<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Mount Davis Technologies on Medium]]></title>
        <description><![CDATA[Stories by Mount Davis Technologies on Medium]]></description>
        <link>https://medium.com/@mountdavistechnologies?source=rss-82c3f2323425------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Vufj7KNLGcwcybP1v0KeuQ@2x.jpeg</url>
            <title>Stories by Mount Davis Technologies on Medium</title>
            <link>https://medium.com/@mountdavistechnologies?source=rss-82c3f2323425------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 23:14:47 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mountdavistechnologies/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From Autonomous Networks to Operator Autonomy]]></title>
            <link>https://medium.com/@mountdavistechnologies/from-autonomous-networks-to-operator-autonomy-b81674d31423?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/b81674d31423</guid>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Sat, 29 Aug 2026 11:34:28 GMT</pubDate>
            <atom:updated>2026-09-20T09:50:34.784Z</atom:updated>
            <content:encoded><![CDATA[<h4>— When Vendor AI Narratives Begin to Eclipse the TM Forum Model</h4><p>Jimmy CHOW</p><p>The technical case for AI-native, agentic network operations is now compelling. The strategic question is different: as intelligence moves from an analytical assistant to an operational actor, who defines and controls the operating model of the autonomous network — the operator, or the vendor?</p><h3>1. Introduction — A Subtle but Fundamental Shift</h3><p>TM Forum’s most recent work on Autonomous Networks, together with vendor-published use cases such as the IOH–Huawei discussion on AI-transformed operations, converges on a single technical storyline: high-quality data → AI and GenAI → intelligent agents → digital twins → multi-agent collaboration → governed closed loops → Autonomous Network Level</p><p>There is little doubt that this stack accelerates automation and reduces human intervention. But from an operator’s perspective the decisive question is not whether the technology works. It is: <em>who defines and controls the operating model of the autonomous network?</em></p><p>That question matters because two narratives now share the same vocabulary while pointing in different directions. The original <strong>TM Forum</strong> vision was <strong>operator-centric</strong>: how an operator expresses business and operational intent and achieves closed-loop autonomy across a heterogeneous, multi-vendor estate — and captures business value in doing so. The emerging <strong>vendor</strong> narrative is increasingly <strong>technology-centric</strong>: data → AI → agents → digital twin → intelligent operations → autonomous network.</p><p>The two are complementary, not contradictory — but they are not identical. The first asks <em>how operators achieve autonomy</em>; the second increasingly asks <em>how intelligent technology operates the network on the operator’s behalf</em>. The gap between those two questions is where operator autonomy is won or lost.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yiiLSxZqGZV05Hdk-kHMCA.jpeg" /></figure><h3>2. The Original TM Forum Logic: Autonomy as an Operator Capability</h3><p>TM Forum’s contribution was never a network-automation product. It was a common framework for describing autonomous <em>operational capability</em>: business intent → intent management → standardised abstraction and information models → open APIs and orchestration → domain automation → closed-loop operations → autonomous network.</p><p>The essential property is that the <strong>operator remains the source of intent and authority</strong>, while the network determines how that intent is fulfilled. This positions TM Forum as an industry-level abstraction layer between the operator and vendor implementations. Its objective is not to prescribe whether Huawei, Ericsson or Nokia implements a given function, but to ensure the operator’s intent stays meaningful and operationally portable across whichever implementation sits underneath. As AI enters the architecture, that portability becomes harder to preserve.</p><h3><strong>3. From automation to agentic operations</strong></h3><p>The vendor architecture — unified data platform → AI/GenAI → agents → multi-agent collaboration → digital twin → governed closed loop → AN Level 4 — is a natural evolution of AIOps. The qualitative change is in the control logic. Traditional automation executes predefined rules (<em>if X, do Y</em>); agentic operations introduce reasoning (<em>observe → understand → reason → plan → execute → verify → learn</em>).</p><p>The consequence is that AI ceases to be an analytical assistant and becomes an <strong>operational actor</strong>. This is the real weight behind the “digital employee” framing: an agent that acts in the operating environment needs an identity, defined responsibilities, skills, authority, access rights, performance measures, supervision, escalation paths and auditability. Agent governance therefore stops being a security feature and becomes a core component of the operating model.</p><h3><strong>4. The deeper shift: delegated decision-making</strong></h3><p>The transformation is not merely that automation becomes more intelligent. It is that <strong>decision-making itself is progressively delegated to software agents</strong> — from <em>human decides → system executes</em> to <em>human sets objectives → agent interprets, reasons, decides and executes</em>.</p><p>This raises governance questions that reaching AN Level 4 or 5 does not answer. Under what authority may an agent act? What data may it access, and which systems may it invoke? Which actions are autonomous, and which require human approval? What happens when confidence is low? Who is accountable for a wrong decision — and can it be audited and reversed? These are questions of the governance architecture of autonomy itself, and they sit directly on top of data governance, AI governance and regulatory constraints such as GDPR and data-sovereignty obligations.</p><p>The <strong>digital twin</strong> becomes central precisely here. Because agents reason probabilistically in an environment that has historically demanded deterministic behaviour, the twin provides an execution sandbox: <em>agent proposes → twin simulates → policy validates → controller executes → network responds → agent verifies</em>. It is far more than a visualisation layer; it is likely one of the critical mechanisms for scaling L4 and, eventually, L5.</p><h3><strong>5. Three levels of openness — and why they diverge</strong></h3><p>A more consequential distinction than “closed vs open” is that openness has three levels, and standardising the lowest does not deliver the highest.</p><ul><li><strong>Interface openness</strong> — can systems communicate? (Open APIs, TM Forum APIs, standard protocols, A2A interfaces.) This provides technical connectivity.</li><li><strong>Semantic openness</strong> — do systems share the same concepts? (Common ontology, intent models, information and data models, knowledge representations.) This provides common meaning.</li><li><strong>Operational openness</strong> — can the operator change the implementation without redesigning the operating model? Can it replace Vendor A’s agent with Vendor B’s while preserving intent, policy, skills, authority and auditability?</li></ul><p>The third is the strategically decisive level, and it is the one interface standardisation alone cannot reach. Hence: <strong>open API ≠ open agent architecture</strong>, and <strong>open architecture ≠ operator autonomy</strong>.</p><p>Multi-agent collaboration sharpens the point. A single “my 5G service is slow” complaint may span customer, service, core, transport, RAN and radio-resource domains, requiring several agents to cooperate. Traditional OSS interoperability asks whether systems can communicate. Agentic interoperability asks something deeper: can intelligent actors understand one another, trust one another, delegate authority and execute coordinated action? Standardised APIs do not, by themselves, answer that.</p><h3><strong>6. Intelligence lock-in</strong></h3><p>Vendor lock-in has historically been discussed at the infrastructure layer — radio, core, OSS, controllers, proprietary interfaces. AI introduces a new dependency that runs the length of a longer chain: network → controller → digital twin → operational data → knowledge base → AI model → agent → agent skills → decision logic → governance.</p><p>Where a single vendor comes to control that whole chain, the operator can achieve highly integrated autonomy while becoming dependent on that vendor’s data model, knowledge model, AI model, agent architecture, twin, controller, workflows and governance mechanisms. The strategic risk shifts from <strong>lock-in at the network layer</strong> to <strong>lock-in at the intelligence layer</strong> — a category worth naming explicitly: <em>intelligence lock-in</em>.</p><h3><strong>7. The end-to-end advantage — and its structural tension</strong></h3><p>An end-to-end vendor approach has real merit. Integrating infrastructure, data, intelligent operations, twins, models, agents, controllers and closed-loop automation reduces integration complexity and accelerates deployment; for an operator seeking rapid progress to AN L4 that is a legitimate advantage.</p><p>The tension is that <strong>end-to-end capability must not become end-to-end dependency</strong>. Integration can deliver highly integrated autonomy without delivering vendor-independent autonomy, and the two should not be conflated.</p><p>There is a subtler dynamic still. A leading vendor need not reject TM Forum; it can embrace intent, open APIs, digital twins, AN Level 4 and the whole industry framework — which makes its architecture appear naturally aligned with the industry direction. The relationship can then quietly invert: from <em>TM Forum defines the framework → vendors implement it</em> to <em>TM Forum provides the vocabulary → the leading vendor’s implementation defines what that vocabulary means in practice</em>.</p><p>This is not necessarily a deliberate attempt to displace TM Forum; it follows from the vendor narrative being more concrete, more demonstrable and more visible than the abstract operating model. But its structural effect — and, in my experience, a recurring source of the integration cost and lock-in operators have carried for years — is that AN L4 becomes associated with one vendor’s AI-native stack rather than with a vendor-neutral operating model. Where an operator lacks the technical and strategic capacity to interrogate that narrative, the association hardens into dependency. This is what it means to say a vendor narrative can <em>eclipse</em> the TM Forum model.</p><p>The corresponding risk for TM Forum is role drift: from defining the industry architecture (ODA, Open APIs) to merely validating that vendor solutions satisfy Autonomous Network requirements. The first role sits above vendor implementations; the second sits beneath them. TM Forum’s strategic value depends on holding the first position.</p><h3><strong>8. From Autonomous Network to Operator Autonomy</strong></h3><p>This yields the central distinction. <strong>Autonomous Network</strong> asks: <em>can the network operate itself?</em> <strong>Operator Autonomy</strong> asks: <em>does the operator control the intelligence that operates the network?</em></p><p>The second is the broader property. It requires that the operator can define intent and policy; set agent authority; control access to operational data; audit, modify, replace and, where necessary, reverse agent actions; and preserve multi-vendor interoperability throughout. An autonomous network does not, on its own, guarantee any of this — which is why the distinction should sit at the centre of the next phase of the discussion.</p><p>The organising principle is <strong>portable autonomy</strong>: autonomy that is not bound to a single vendor’s stack. A genuinely operator-owned model should allow Vendor A’s agent to run over Vendor B’s controller and Vendor C’s network — and to be recombined — without forcing a redesign of the operating architecture. Portability must extend beyond APIs to intent, semantics, policy, skills, agent identity, authority, knowledge, audit and execution.</p><p>Architecturally, this implies a layered model in which the operator owns and governs everything above the vendor-specific implementation:</p><blockquote><em>operator intent → intent / semantic layer → policy and governance layer → AI / agent layer → digital-twin / validation layer → orchestration and control layer → multi-vendor network</em></blockquote><p>Vendors remain free to innovate below and within these abstractions. The layering is precisely what lets the industry absorb rapid AI innovation without letting the operating model itself become proprietary — a clean separation between <strong>industry standards</strong> and <strong>vendor innovation</strong>.</p><h3><strong>9. Implications for vendor evaluation and for TM Forum</strong></h3><p>This reframes vendor comparison. The question is not “which vendor has the better AI?”, nor “the more advanced L4 use cases?”, nor even “the better Autonomous Network platform?” It is whether a given vendor’s architecture ultimately <strong>strengthens or weakens operator autonomy</strong> — a more neutral test that assumes neither that Huawei is wrong nor that Ericsson is right.</p><p>It also clarifies TM Forum’s role in the AI era, which becomes more important rather than less, provided it evolves. The old task was to standardise automation. The new task is to standardise autonomous intelligence <em>without standardising away innovation</em> — to hold the boundaries between intent, semantics, policy, agents, data, governance and execution, while leaving the intelligence itself to the vendors. The objective is not to standardise the intelligence, but to keep it interoperable, governable, portable and replaceable. That is the condition under which vendor innovation and operator autonomy can coexist.</p><h3><strong>10. Conclusion</strong></h3><p>The industry is moving through two transformations in parallel. The first is technological — automation → AIOps → AI-assisted operations → AI agents → multi-agent operations → AI-native network — and it is happening quickly and visibly. The second is architectural — vendor-specific automation → open interfaces → common semantics → portable intelligence → operator-governed autonomy — and it is less visible but strategically more important. The mistake would be to assume the first produces the second. It does not: technological autonomy does not automatically confer operator autonomy.</p><p>The IOH–Huawei discussion is a credible picture of the next generation of operations, and Huawei’s contribution matters precisely because it turns these concepts into integrated, deployable systems. But that success sharpens the underlying question. The competition is no longer only about who can make the network more autonomous; it is about <em>who defines the intelligence that makes it autonomous</em> — and, ultimately, <em>who controls that intelligence</em>.</p><p>This is where the original TM Forum philosophy remains essential. Its role should not shrink to supplying terminology, APIs and maturity certification around vendor-developed AI. It should ensure that Autonomous Networks remain an <strong>operator capability</strong> rather than becoming a vendor-controlled intelligence stack. The goal is not merely the Autonomous Network, but <strong>Operator Autonomy</strong>, resting on five principles: <em>open interfaces; common semantics; portable intelligence; independent governance; operator-controlled autonomy.</em></p><p>The defining strategic risk of the AI era is no longer vendor lock-in at the network layer. It is <strong>intelligence lock-in</strong> — and the responsibility of TM Forum, and of operators themselves, is to ensure that AI-native, agentic networks do not quietly turn the operator from the <em>owner</em> of autonomy into the <em>consumer</em> of vendor-provided autonomy.</p><p><strong>References</strong></p><ol><li>TM Forum, <em>New-generation Intelligent Operations: An AI-native Reinvention</em> (Autonomous Networks white paper, v4.0). <a href="https://inform.tmforum.org/research-and-analysis/reports/new-generation-intelligent-operations-an-ai-native-reinvention">https://inform.tmforum.org/research-and-analysis/reports/new-generation-intelligent-operations-an-ai-native-reinvention</a></li><li><em>IOH and Huawei on How AI Is Transforming Autonomous Network Operations</em> (TM Forum). <a href="https://inform.tmforum.org/videos/ioh-and-huawei-on-how-ai-is-transforming-autonomous-network-operations">https://inform.tmforum.org/videos/ioh-and-huawei-on-how-ai-is-transforming-autonomous-network-operations</a></li><li>Mount Davis Technologies, <em>Autonomous but Unaccountable: Multi-Vendor Interoperability, Accountability, and Openness Gaps in AI-Native Network Operations</em> (analysis of §4 of the white paper above; PDF).</li></ol><p>Aug. 28, 2026</p><p>— THE END</p><blockquote><strong>Professional Bio / Collaboration Offer</strong></blockquote><blockquote><strong>Strategic Network Transformation &amp; Integration Expert.</strong></blockquote><blockquote>If your organisation is navigating <strong>multi-generation network integration</strong> (2G / 3G / 4G / 5G / 6G) en route to a future-proof <strong>cloud-native core network</strong>, or managing the intricate architectures and fragmented service platforms that arise from <strong>mergers and acquisitions (M&amp;A)</strong> aimed at an <strong>Autonomous Network Architecture, or AI/Enterprise-Data governance</strong> — get in touch. I provide end-to-end support: situational analysis, strategic roadmap, hands-on implementation and smooth transition. I work either independently or embedded in your existing team.</blockquote><blockquote><strong>Email:</strong> mountdavistechnologies@gmail.com</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b81674d31423" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Determinism Need Not Come at the Price of Closure]]></title>
            <link>https://medium.com/@mountdavistechnologies/determinism-need-not-come-at-the-price-of-closure-448039def641?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/448039def641</guid>
            <category><![CDATA[tmforum]]></category>
            <category><![CDATA[autonomous-network]]></category>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Fri, 14 Aug 2026 19:39:29 GMT</pubDate>
            <atom:updated>2026-08-14T19:39:29.318Z</atom:updated>
            <content:encoded><![CDATA[<p>-What TMF <strong>Autonomous Networks L4/L5</strong> Implies for the Right Evolution of <strong>Mission-Critical Networks (MCN) </strong>— and the<strong> Operator’s Strategic Choice</strong></p><p><em>-An architecture analysis report for operators building open, autonomous networks</em></p><p>Jimmy CHOW</p><p><strong><em>Maturity convention. </em></strong><em>Capabilities are marked in three tiers — </em><strong><em>[Standard available]</em></strong><em> = an implementable international-standard interface/model already exists; [Standard maturing] = a standard exists but is still being completed or requires extension;</em><strong><em> [Standardization pending] </em></strong><em>= the direction is clear and it is a key gap the industry should invest in standardizing. This report deliberately stays at the level of architectural principle and required properties; the concrete implementation mechanisms of each plane are out of scope — both to keep a strict line between what standards can already do and what the industry still has to build, and to keep the report focused on</em><strong><em> the architectural judgment operators actually need</em></strong><em>.</em></p><h3><strong>1. The reference frame: what TMF Autonomous Networks L4/L5 actually promises</strong></h3><p>To judge whether a given <strong>MCN </strong>offering “realizes” or “departs from” AN L4, one must first state precisely what AN L4/L5 promises, rather than stopping at the repeatedly rehearsed slogans of “intent-driven + closed loop + AIOps.”</p><p><strong>(1) Autonomy levels and their scoring convention.</strong> TMF’s AN framework (IG1218/IG1230) defines six levels from L0 (fully manual) to L5 (full autonomy); L4 = high autonomy, execution out-of-loop, with humans setting only intent. The crucial part is its <strong>evaluation methodology (IG1252)</strong>: the autonomy level of any scenario is scored across five dimensions — Awareness / Analysis / Decision / Execution / Intent — each rated P (personnel/human) / S (system) / P-S (shared). L4 is defined as Awareness, Analysis, Decision and Execution all = S, with Intent = P/S. This is complemented by KEI (IG1256), GB1059/ANLET (level-evaluation specifications), and <strong>ANLAV (AN Level Assessment Validation)</strong>, an independent validation of the level assessment.</p><p>This point recurs later: <strong>L4 is not “removing the human” but “moving the human up to the intent and governance layer.”</strong> The retention of Intent = P/S, together with independent validation (ANLAV), shows that in the standard framework “high autonomy” and “accountability” are co-designed from the outset — they are not opposites.</p><p><strong>(2) The three-layer intent architecture and the declarative interface.</strong> The AN reference architecture (IG1251) divides operations into a business layer (minutes–hours), a service layer (seconds–minutes) and a resource layer (milliseconds–seconds), linked by intent interfaces I1/I2/I3. Intent is carried by the <strong>TMF921 Intent Management API</strong> [Standard available]; the data model is the TMF ICM/TIO intent ontology (TR290/291/292/299), bridged to <strong>3GPP TS 28.312</strong> [Standard available] via <strong>TR294A</strong> [Standard available].</p><p>The essence of an intent interface is that it is <strong>declarative</strong>: the upper layer states only “what” (goals, SLO/SLE), not “how” (specific parameter scheduling). This “what, not how” design is the technical pivot of the entire anti-lock-in logic — <strong>as long as the cross-layer/cross-domain contract is declarative intent, the in-domain implementation is substitutable.</strong> Keep this in mind; it is the key to the whole of Section 4.</p><p><strong>(3) ODA + Open APIs: an explicit anti-lock-in commitment.</strong> ODA (Open Digital Architecture, IG1167) is a componentized reference architecture: an ODA Component exposes TMF Open APIs and has a CTK (Component Test Kit) for conformance. ODA exists for one reason — to let operators mix and match from a directory of substitutable, certifiable components and replace them one by one, lowering integration cost and eliminating lock-in. In other words, <strong>vendor-lock-in avoidance is not a by-product of AN; it is the reason ODA was created.</strong></p><p><strong>(4) Closed loops and cross-domain coordination.</strong> Autonomy rests on closed loops: domain-level loops plus an E2E cross-domain loop (ETSI ZSM 002/009) [Standard available], each autonomous domain running an Awareness–Analysis–Decision–Execution (AADE/OODA) loop with cross-domain conflict resolution and intent alignment above. One fact matters later: <strong>ZSM explicitly distinguishes a domain closed loop from an E2E cross-domain closed loop, and O-RAN explicitly distinguishes the Near-RT RIC (10 ms–1 s loop) from the Non-RT RIC (≥1 s loop). Fast loops and slow loops are architecturally layered.</strong></p><p>Taken together, AN L4/L5’s promise to operators compresses into three sentences:</p><p>• <strong>Declarative intent → substitutable in-domain implementation</strong> (the interface basis for anti-lock-in);</p><p>• <strong>ODA components + Open APIs → no vendor lock-in</strong> (the component basis for anti-lock-in);</p><p>• <strong>Governed closed-loop autonomy → high autonomy without surrendering accountability</strong> (autonomy ≠ removing traceability).</p><p>Any offering that calls itself “AN L4” should be measured against these three sentences.</p><h3><strong>2. The vertical-integration route: a commercial reconstruction of AN L4 (a fair assessment)</strong></h3><p>The MCN OSS/BSS offering currently promoted by the major network-equipment vendors (hereafter the “vertical-integration route,” analyzed with reference to a recent and representative industry white paper) deserves to have its fair part stated in full before its structural move is located.</p><p><strong>It is a scenario-specific repackaging of AN L4.</strong> The capabilities such offerings describe — “service-aware assurance,” “closed-loop self-healing,” “AIOps-driven operations,” “agentic-AI agents” — overlap heavily with AN architecture standards of the past 3–5 years (IG1251 / TS 28.312 / ZSM). The commercial reality is that willingness to pay for L4 in the general consumer network (ToC) comes mainly from OPEX reduction, and with 5G ROI slowing, operators are reluctant to pay a large premium for “elegant architecture.” Repackaging the L4 toolkit and steering it into the high-premium mission-critical ToB/ToG domains (public safety, rail FRMCS, defence, utilities) is therefore a rational monetization move.</p><p><strong>But MCN’s “zero-tolerance” character genuinely forces modifications to several core assumptions of generic L4 — real engineering content, not empty words:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/1*nTl9tm2u60QSkClV5rkqPg.png" /></figure><p><strong>These modifications have value.</strong> An MCN OSS/BSS must indeed possess capabilities a public consumer OSS/BSS does not — <strong>violent preemption and deterministic resource reservation</strong>, and extensive <strong>evidence capture and audit trails</strong> (to determine, after an incident, whether it was an AI decision error, a configuration error, or a hardware failure). These are operational details easily overlooked in the AN L4 specifications; making them real is genuine engineering and should be acknowledged.</p><p><strong>Now to locate the structural move.</strong> The problem is not that the route delivers determinism and accountability, but <strong>how</strong> it does so. Section 1 noted that AN’s anti-lock-in basis is “the cross-layer/cross-domain contract is declarative intent, and the in-domain implementation is substitutable.” What the vertical-integration route does is precisely to <strong>collapse the intent plane into a private execution plane</strong>: the deterministic SLA is produced not at the intent boundary but in a private sub-layer of proprietary telemetry + proprietary self-healing channels that sits below the intent interface. This is why the top-line claim “aligned with the principles of 3GPP, TM Forum and open APIs” can be true — openness occurs only at a decorative top layer, while the value (the closed loop) lives in the private sub-layer. It is also why “swap one component and the SLA guarantee lapses or degrades”: the SLA guarantee is not produced at the standard intent boundary at all. This is the master key to every lock-in mechanism that follows.</p><p>A related piece of language operators should be wary of: <strong>“aligned with principles” is unfalsifiable.</strong> Conformance, by contrast, is testable — ODA component decomposition, CTK certification of Open APIs, an autonomy level stated per IG1252/GB1059, and independent ANLAV validation can all be written into a contract and verified by a third party. An offering that stresses “network autonomy” throughout yet never states a definite autonomy level should be required to give answers on these testable dimensions.</p><h3><strong>3. The core tension: why “determinism” squeezes out “openness” in practice (the opposing case at full strength)</strong></h3><p>To refute an argument, one must first put it at its strongest. The vertical-integration route’s argument is powerful in the MCN context and should not be treated lightly.</p><p><strong>The multi-vendor blame game.</strong> In the more tolerant consumer network, multi-vendor builds (one RAN vendor + one core vendor + one transport vendor) are commonplace, and fault-finding can take hours or days. But in an MCN (99.999% reliability, &lt;10 ms jitter, zero disconnection), the moment rescue communications drop, a multi-vendor architecture descends into mutual finger-pointing: the RAN vendor says “my radio channel is fine, it’s core-network routing latency,” the core vendor says “my signaling is fine, it’s radio-side packet loss,” the transport vendor says “it’s a burst of radio traffic that overwhelmed me.” No one accepts responsibility; the incident is left unowned.</p><p><strong>Why do standard APIs appear unable to solve this?</strong> The route offers three “fatal flaws”:</p><p>• <strong>Insufficient granularity</strong> — a standard intent API carries only abstract instructions (“guarantee 10 ms latency for this slice”) and cannot pass through the device’s microsecond-level scheduling logs;</p><p>• <strong>Too slow a closed loop</strong> — cross-vendor API call latency is in the hundreds-of-milliseconds-to-seconds range, whereas MCN self-healing requires milliseconds;</p><p>• <strong>Black-box distrust</strong> — vendor A’s AIOps dares not blindly trust the state reported by vendor B, and the automation loop breaks at the cross-vendor boundary.</p><p><strong>Three mechanisms of covert lock-in.</strong> So even where such offerings verbally “support standard APIs and a multi-vendor ecosystem,” in deployment they achieve deep binding along three dimensions:</p><p>• <strong>Telemetry monopolized as private</strong> — the AIOps’s accurate prediction depends on non-public, fine-grained telemetry internal to the equipment; insert another vendor’s core network and that vendor’s OSS loses full-stack visibility into the underlying physical state.</p><p>• <strong>A private actuation channel</strong> — under interference, the vendor’s OSS can, within milliseconds, directly invoke internal algorithms of its own base station (e.g., beamforming adjustment) — a super-real-time loop unattainable via generic standard APIs.</p><p>• <strong>Commercial single point of accountability</strong> — government/defence procurement values contractual liability above all; with a multi-vendor build legal responsibility is hard to establish after an incident, whereas with a full single-vendor suite that vendor bears whole-chain SLA legal liability. This “commercial wager” drives the customer to the single-vendor suite.</p><p><strong>Stating the opposing conclusion in full:</strong> under “zero-failure anxiety,” determinism (an in-domain private loop + a single commercial backstop) does, in fact, squeeze out openness (substitutability + a cross-domain standard interface); the operator must either accept the vertical suite in exchange for absolute indemnity, or bear the raised SLA risk and higher fault-finding cost of a multi-vendor build.</p><p>The argument holds <strong>under the assumptions</strong> that “the closed loop must complete across vendor boundaries” and that “accountability must rest on vendor self-report.” The next section dismantles both.</p><h3><strong>4. The tension is resolvable: three conflated layers</strong></h3><p>The vertical-integration route’s argument is powerful because it merges three pairs of things that should be kept apart. Take them one at a time — this three-layer distinction is the analytical spine of this report.</p><h4><strong>4.1 The loop-tier conflation: determinism is intra-domain, openness is cross-domain</strong></h4><p>“Cross-vendor loops are too slow; millisecond self-healing is impossible” holds <strong>only against a straw man</strong>: the assumption that the millisecond self-healing loop must cross a vendor/domain boundary. No competent AN design does this.</p><p>In the correct layering, <strong>the millisecond fast loop is always intra-domain.</strong> This is not this report’s invention; it is the structure of the standards themselves:</p><p>• ETSI ZSM explicitly distinguishes a <strong>domain closed loop</strong> from an <strong>E2E cross-domain closed loop</strong>;</p><p>• O-RAN explicitly distinguishes the <strong>Near-RT RIC</strong> (10 ms–1 s control loop, over E2AP/E2SM) from the <strong>Non-RT RIC</strong> (≥1 s, over A1).</p><p>You do not self-heal a radio beamforming problem by “calling the core vendor’s OSS” — you self-heal it locally in the RAN domain’s own Near-RT loop. The cross-domain plane handles only the slower, service-level intent alignment and conflict resolution (seconds–minutes), and at that tier the latency of a standard intent API is entirely adequate.</p><p><em>Determinism (the millisecond closed loop) is an intra-domain property; openness (the substitutable standard interface) is a cross-domain property. They operate at different latency tiers and simply do not compete within the same loop.</em></p><p>This one cut directly refutes “determinism necessarily requires vertical closure.” “Millisecond private self-healing” is a real selling point, but all it proves is that fast loops belong in-domain — <strong>not</strong> that the space between domains must also be private. Fast loops stay in-domain (they may be single-vendor, may use proprietary algorithms — legitimate); slow loops run over cross-domain standard intent. This design obtains both millisecond determinism <strong>and</strong> cross-domain substitutability.</p><h4><strong>4.2 The control-plane / observability-plane conflation: a coarse interface is a feature, not a defect</strong></h4><p>“Standard APIs are too coarse to pass through microsecond logs” conflates the <strong>control interface</strong> with the <strong>observability/evidence interface</strong>.</p><p>The control interface <strong>should</strong> be a coarse-grained declarative intent. Having the upper layer micromanage microsecond scheduling is not strength but an architectural error — exactly what domain autonomy is meant to abolish, and exactly why the millisecond fast loop stays in-domain. So “the control API cannot carry microsecond logs” is not a defect; it is correct design.</p><p>Microsecond evidence belongs to a <strong>different plane</strong>: the observability/telemetry plane. It can be independent, standardized and fine-grained — and the relevant standards are being completed: O-RAN E2SM-KPM + R1/O1 data services, 3GPP MDT/Trace, gNMI/YANG-Push streaming telemetry [Standard available / maturing].</p><p>The vertical route’s “solution” is to privatize <strong>both</strong> planes (a private control channel + a private telemetry monopoly). The correct solution is to <strong>separate</strong> them: keep the control plane declarative, open and substitutable; make the observability plane an independent, standardized evidence plane. Treating the openness of the control plane and the granularity of the observability plane as an either/or is the heart of this conflation; they are two planes and can hold simultaneously.</p><h4><strong>4.3 The trust / verifiability conflation: replace “player-cum-referee” with an independent referee</strong></h4><p>“Vendor A dares not trust the state reported by vendor B” — the root of the blame game is that each vendor is both the <strong>executor</strong> and the <strong>self-reporter</strong> of its own state, i.e., “both player and referee.”</p><p>But this is a trust problem <strong>only when accountability = vendor self-report.</strong> If accountability rests on evidence that is independent of vendor self-report, cross-verifiable and tamper-evident, then the automation loop <strong>needs no inter-vendor trust at all</strong> — what it needs is a shared, verifiable ground truth.</p><p>This converts “trust” (a social/commercial problem, which the vertical route solves by being “the single entity you can hold liable”) into “verifiability” (a technical problem, solved by the independent accountability plane). The price of the former is total lock-in; the product of the latter is unlocking.</p><p><em>Single commercial accountability = commercial backstop (technically opaque); independent, verifiable accountability = technical trust. The former locks you in; the latter frees you.</em></p><p>This leads directly to the third plane in Section 5 and the strategic argument in Section 6: the MCN customer believes it is buying “accountability,” but under a “vendor self-report” architecture what it actually buys is “someone to sue after an incident” — not the same as what MCN actually needs (see §6). <strong>(To be explicit: the concrete implementation mechanism of this independent accountability plane is out of scope for this report; this section establishes only its architectural necessity and required properties.)</strong></p><h3><strong>5. The correct target architecture: domain autonomy + intent governance + independent accountability</strong></h3><p>Combine the two “break-out paths” often discussed side by side — “domain autonomy” (often phrased as the “domain-lock-in path”) and “independent accountability” (the “evidence-chain path”) — into a single architecture.</p><p><strong>An architectural judgment worth stressing: the two should not be listed as alternative paths.</strong></p><p>• <strong>Domain autonomy used alone degrades into “legitimized lock-in”</strong> — “for the sake of SLA, the vendor domain must be closed” is precisely the core sales pitch of the vertical route;</p><p>• <strong>Independent accountability used alone has no architecture to carry it</strong> — with no intent boundary, there is nowhere for the evidence to attach and no one to hold accountable.</p><p>The two must be combined: <strong>domain autonomy is a legitimate component of an open architecture — rather than a fig leaf for lock-in — only when the two constraints “the cross-domain boundary is declarative intent” and “accountability is independent and verifiable” hold simultaneously.</strong> The target architecture below is given on this combination, and strictly distinguishes planes that public standards can already support from the plane the industry still has to standardize.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2VOtAo1xDV2NF8Vvh2tt3A.png" /></figure><p><em>Figure 1. Target architecture — domain autonomy, open intent governance, and an independent accountability keystone. Fast (millisecond) loops stay in-domain, where determinism is produced; the cross-domain plane carries only declarative intent, keeping each domain substitutable; an independent, verifiable accountability plane replaces vendor self-report.</em></p><p><strong>The three planes, with their standards and maturity</strong> (keep “available” and “to be built” strictly apart):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/871/1*6t0VwB7cny5PcZyrp8HwIA.png" /></figure><p><strong>Why can multi-vendor SLA be “assured”? A mechanism, not a slogan.</strong> The assurance rests on two principles:</p><p>• <strong>Intent decomposition</strong> — the E2E SLA is decomposed via intent into per-domain SLO/SLE commitments (for end-to-end latency, split the total budget into per-domain budgets for air interface / fronthaul / transport / edge — mature engineering practice);</p><p>• <strong>Independent verification</strong> — whether each domain meets its SLO is verified independently of vendor self-report.</p><p>So when a domain breaches, <strong>the budget + independent verification make the breach point objectively locatable</strong>, rather than descending into finger-pointing. This is the principled core of “SLA is still assurable in a multi-vendor decoupled architecture,” and the direct answer to the blame game. (The concrete implementation of the independent-verification/accountability plane is out of scope for this report.)</p><p><strong>Not a trim tab but the keystone of the whole.</strong> Planes ① and ② have usable public standards today; the only one still to be built is ③ — the independent, tamper-evident, cross-vendor accountability capability. It must be said plainly: <strong>it is precisely the absence of ③ that still makes “determinism vs. openness” look like a dilemma in practice.</strong> Because once accountability can rest only on vendor self-report, the operator&#39;s only fallback is “hand the whole chain to one vendor so that someone bears liability after an incident” — and vertical lock-in thereby acquires its one genuinely defensible rationale. <strong>Conversely, once ③ is built, that rationale vanishes:</strong> determinism stays in-domain, openness stays cross-domain, and responsibility is established by independent evidence — all three hold at once and the false dilemma dissolves. So ③ is not a “detail to be filled in” but the keystone that decides the architectural direction of the next generation of MCN — its boundary is clear, it is buildable, and it is a definite engineering and standardization agenda; whoever first pushes it into standards and makes it real redefines the architectural initiative of MCN. This report leaves its concrete implementation mechanism to dedicated work and establishes here only one thing: <strong>building the independent accountability plane is the highest-priority architectural task for MCN today, not a distant open problem.</strong></p><h3><strong>6. Why this is the operator&#39;s strategic optimum</strong></h3><p>The previous sections established that “open + multi-vendor-interoperable + assured SLA + traceable accountability” is architecturally feasible. Why it is optimal for the operator&#39;s long-term interest and network development strategy — by reasons, not assertion:</p><p><strong>(1) Evolution autonomy and ecosystem elasticity.</strong> The vertical suite confines the closed loop to a private protocol channel, and its real cost is <strong>binding the pace of network evolution entirely to a single vendor&#39;s R&amp;D roadmap</strong>: any domain-specific innovation the operator wishes to introduce (flexible third-party slicing, specialized probes, better transport orchestration) is at the vendor&#39;s mercy. An open intent boundary lets the operator introduce the best component per domain, keeping the initiative in its own hands. For a national-scale critical infrastructure meant to last two or three decades and cross 5G-A/6G, this degree of freedom is a strategic necessity, not a nicety.</p><p><strong>(2) Cost and competition.</strong> The single-vendor suite eliminates competitive tension on the procurement side and locks the operator into a high unit price. Multi-vendor substitutability preserves competition and keeps total cost of ownership in a reasonable band over the long run — especially important for government buyers whose budgets are subject to public accountability.</p><p><strong>(3) Technical trust &gt; commercial indemnity — the most crucial and most misunderstood point in the MCN context.</strong> What the vertical route sells as “single accountability” is a <strong>commercial backstop</strong>: after an incident there is one entity to sue. Its value is real (someone bears liability), but it is <strong>after the fact.</strong> And in an MCN, after-the-fact compensation is almost worthless: an SLA credit does not undo a failed rescue call, and a legal settlement does not soothe a derailed train. What MCN actually needs is three things — <strong>real-time prevention (the closed loop) + real-time attribution (rapid containment) + after-the-fact forensics (to prevent recurrence and assign liability correctly)</strong> — and all three rest on technical transparency, not commercial backstop.</p><p><em>The vertical route sells “accountability” but delivers “someone to sue after an incident”; what the operator actually needs is “an objective truth, independent of the vendor&#39;s account, that is both real-time and forensically recoverable.” The former is accountability in the legal/commercial sense; the latter is traceability in the cognitive/technical sense — whether one can know, independently of the vendor&#39;s one-sided account, what actually happened. Independent, verifiable accountability delivers the latter; single commercial accountability cannot.</em></p><p>For MCN, the latter is the real need. To stop at commercial backstop is to pass off “someone pays afterward” as “nothing fails in the moment, and it is fully investigable afterward.” And the technical device that upgrades “commercial backstop” into “technical trust” is precisely plane ③, the independent accountability plane — <strong>not an optional bonus but the switch that determines whether this entire strategic argument can be realized</strong>: without it, an open architecture always lacks a piece of assurance in front of MCN; with it, “open and accountable” becomes, for the first time, a promise that can be written into a contract.</p><p><strong>(4) Resilience and dependency concentration.</strong> Handing the whole chain to a single vendor binds the network&#39;s fate to one company&#39;s roadmap, release cadence and commercial terms; more importantly, <strong>a single-vendor suite is a monoculture systemic risk</strong>: one vendor&#39;s vulnerability, roadmap misstep or supply-chain/geopolitical exposure is amplified, in one stroke, into a whole-network exposure. This conflicts with the very raison d&#39;être of mission-critical networks — <strong>the reason public-safety, defence and rail networks exist is precisely to not collapse in the worst case</strong>; to carry a “survive-under-adversity” mission on a single-point-dependent, single-stack architecture is the wrong direction in itself. Open + substitutable + diverse is the architectural expression of resilience. (That “data staying in-jurisdiction” does not equal “technical autonomy” — the real control still sits with the vendor — this report does not elaborate, but operators should be clear-eyed about it.)</p><p><strong>(5) This was the whole point of L4/L5.</strong> The entire ODA / Open API / intent programme exists to place <strong>control + substitutability + governed autonomy</strong> in the operator&#39;s hands. The vertical model&#39;s problem is not that it fails to achieve autonomy — it does; the problem is that <strong>it delivers autonomy while surrendering control and substitutability.</strong> What it gives the operator is “the L4 that is convenient for the vendor,” not “the L4 that serves the operator.” And IG1252&#39;s scoring convention itself is a reminder: L4&#39;s Intent dimension is P/S and must pass independent ANLAV validation — high autonomy and accountability are meant to coexist. An offering that locks accountability into a private black box and control into a proprietary suite, however high its autonomy, departs from the point of L4/L5.</p><p>In sum: anti-lock-in, multi-vendor interoperability, assured SLA and traceable accountability are not an idealism opposed to “determinism/reliability” but an architecture that serves <strong>evolution autonomy, cost competition, resilience and — most importantly — the real MCN accountability need</strong> simultaneously. Whether these four can actually be realized ultimately rests on the same keystone: an independent, verifiable accountability plane. That is the optimum for the operator&#39;s interest and network development strategy.</p><h3><strong>7. The operator&#39;s architecture-governance and procurement criteria: lock openness into the contract today</strong></h3><p>One must be honest: the target architecture is <strong>not plug-and-play today</strong> (particularly plane ③, still to be built). The vertical-integration offering is the most readily deployable deterministic delivery under current engineering conditions — the high-water mark of the “1.0 era.” The correct posture is not to tear everything down but to <strong>migrate in phases — and to put the lever on the procurement (RFP) terms</strong>, because where the operator can actually exert force is not an architecture deck but the contract.</p><p><strong>Near term (can be written into the contract immediately):</strong></p><p>• <strong>Mandate conformance of the cross-domain declarative-intent boundary</strong> — require cross-domain interfaces to be conformant to TMF921 / 28.312 (not “aligned with principles”), so each domain is substitutable at contract granularity;</p><p>• <strong>Mandate telemetry openness</strong> — prohibit a private telemetry monopoly; make standardized export of key telemetry (O-RAN E2SM-KPM / gNMI / YANG-Push / 3GPP Trace) a delivery condition. This alone dismantles the first mechanism of covert lock-in;</p><p>• <strong>Require an ANL self-assessment + ANLAV independent validation</strong> — require vendors to state a definite autonomy level per IG1252/GB1059 and submit to independent validation; for any offering that talks “network autonomy” but gives no level, this silence should be broken in the RFP;</p><p>• <strong>Require ODA component decomposition and a substitutability statement</strong> — which ODA Components, which Open APIs are CTK-certified, and whether each component can be replaced independently;</p><p>• <strong>Write in an independent-accountability requirement</strong> — state that “fault attribution shall not rest solely on the vendor&#39;s own log reports” and require independent verifiability; even while the full technical standard is maturing, stand up the principled requirement first, leaving an interface for later technical realization.</p><p><strong>Mid term:</strong> make cross-domain arbitration “policy + utility function, ML proposes and does not adjudicate + human-in-the-loop for high-stakes,” redefining human-on-the-loop from “dependence on the vendor&#39;s operations” into “a governance boundary on an open architecture”; and drive the standardized observability/evidence interfaces that independent accountability requires.</p><p><strong>Long term:</strong> as the standard for the independent accountability plane matures, merge it with the intent-governance boundary to form the 2.0 target state of “open multi-vendor + assured SLA + objective accountability.”</p><p><strong>In one sentence:</strong> the operator need not choose between “absolute safety” and “anti-lock-in”; it should use procurement terms to lock the openness of planes ① and ② into the contract today, while driving the standardization of plane ③ — putting time on the side of openness.</p><h3><strong>8. Conclusion</strong></h3><p>Mission-critical networks force a seemingly binary choice: determinism, or openness. Vendors are happy to let this dilemma look like a law of engineering, because “if you want zero failure you can only use my full suite” is their best sales pitch.</p><p>But the dilemma is <strong>manufactured.</strong> It rests on three conflations — merging the in-domain fast loop with the cross-domain slow loop, merging the open control plane with the fine-grained observability plane, and merging vendor-self-report “trust” with independent, verifiable “accountability.” Separate the three, and TMF&#39;s layered architecture already contains the resolution: <strong>determinism is produced in-domain, openness is preserved cross-domain, and accountability is delivered by an independent, verifiable capability.</strong></p><p>The vertical-integration route represents the high-water mark of the single-vendor vertical closed loop and deserves respect; but it presents a <strong>choice on a spectrum</strong> as “the only way to get MCN reliability.” The operator&#39;s task is to restore the vendor&#39;s redefinitions of the words <strong>openness, autonomy and accountability</strong> to their true meaning, and to recognize: an architecture that prevents single-vendor lock-in, is multi-vendor-interoperable with assured SLA, and is traceable in accountability is not an idealist&#39;s luxury but <strong>the only architecture that simultaneously delivers reliability, cost, evolution autonomy and resilience</strong> — the optimum for the operator&#39;s long-term strategy. And its keystone is the independent accountability plane the industry has yet to build and that is the most decisive of all — <strong>building it is the key step in turning open, autonomous networks from vision into reality.</strong></p><p><strong><em>Determinism need never come at the price of closure.</em></strong></p><p><em>The TMF/3GPP/IETF/O-RAN/ETSI standards cited are referenced for their approach and semantics. The report focuses on architectural principle, the state of the standards, and the operator&#39;s strategic judgment; capabilities such as the “independent accountability plane” are defined only by their architectural necessity and required properties, with concrete implementation mechanisms out of scope. The analysis of the “vertical-integration route” targets a class of technical approach as reflected in publicly available industry materials, not any specific vendor&#39;s product.</em></p><p><em>Jimmy CHOW ·14 Aug. 2026</em></p><blockquote><strong><em>Professional Bio / Collaboration Offer</em></strong></blockquote><blockquote><strong><em>Strategic Network Transformation &amp; Integration Expert.</em></strong><em> If your organisation is navigating </em><strong><em>multi-generation network integration</em></strong><em> (2G / 3G / 4G / 5G / 6G) en route to a future-proof </em><strong><em>cloud-native core network</em></strong><em>, or managing the intricate architectures and fragmented service platforms that arise from </em><strong><em>mergers and acquisitions (M&amp;A)</em></strong><em> aimed at an </em><strong><em>Autonomous Network Architecture</em></strong><em> — get in touch. I provide end-to-end support: situational analysis, strategic roadmap, hands-on implementation and smooth transition. I work either independently or embedded in your existing team.</em></blockquote><blockquote><strong><em>Email:</em></strong><em> mountdavistechnologies@gmail.com</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=448039def641" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Non-Delegability of Judgment and the Crisis of Political Existence]]></title>
            <link>https://medium.com/@mountdavistechnologies/the-non-delegability-of-judgment-and-the-crisis-of-political-existence-e375c2344bbf?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/e375c2344bbf</guid>
            <category><![CDATA[ai-judgement]]></category>
            <category><![CDATA[hannah-arendt]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[judgement]]></category>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Tue, 26 May 2026 13:02:27 GMT</pubDate>
            <atom:updated>2026-05-26T13:02:27.638Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Critique of the AI Alignment Paradigm</em></p><p>Jimmy CHOW</p><p><strong>Abstract</strong></p><p>Contemporary AI alignment research establishes as its central problem the question of <em>how AI systems can be made to act as intended</em>. This paper argues that every genuinely possible answer to this engineering question ultimately depends upon something that is, on ontological grounds, not engineerable — namely human judgment — and that the methodological posture of alignment research as a whole consists precisely in the attempt to engineer this non-engineerable thing. Taking as its concrete object of analysis four alignment studies published by the UK AI Security Institute (AISI) between mid-2025 and mid-2026 — concerning, respectively, automated alignment,<a href="#_ftn1">[1]</a> self-evaluation and wireheading,<a href="#_ftn2">[2]</a> inoculation prompting,<a href="#_ftn3">[3]</a> and prover-estimator debate<a href="#_ftn4">[4]</a> — and drawing on Hannah Arendt’s ontology of judgment as the diagnostic framework and the author’s prior work on <a href="https://open.substack.com/pub/jimmychow/p/the-crisis-of-political-visibility?r=56ft8&amp;utm_campaign=post-expanded-share&amp;utm_medium=web"><em>political visibility</em></a><a href="#_ftn5">[5]</a> as the political entry point, the paper develops a three-tiered argument.</p><p>First, what the four papers collectively reveal — yet collectively fail to name — is the <em>non-delegability of judgment</em>. Paper ① touches the territory of<em> reflective judgment</em> in its discussion of fuzzy tasks, yet misrecognises judgment’s ontological structural feature as an empirical deficit of human capacity; Paper ② precisely exposes, in its formalisation of wireheading, the split between <em>who</em> and <em>what</em>, yet fails to <em>identify the ontological character of this split</em>; Paper ③’s inoculation prompting is the purest form of operation at the level of <em>what</em>; Paper ④’s debate protocol attempts to<em> replace judgment with procedure</em>, yet preserves judgment’s irreducibility within its stability assumptions.</p><p>Second, the non-delegability of judgment is, in Arendtian terms, rooted in the ontological structure of <em>natality, plurality, and action</em> — and the alignment paradigm’s circumvention of this structure is not an incidental failure but a structural necessity of its position as the institutional completion of <em>the rise of the social</em> within the domain of AI governance.</p><p>Third, within the <em>political visibility framework</em>, this circumvention constitutes a systematic erosion of <em>politics itself as a mode of existence</em> — what is destroyed is not some democratic procedure but the ontological precondition under which political existence is possible. AISI as an institution is, in the contemporary UK governance context, the concrete institutional form of this erosion.</p><p>The paper concludes that the path of reconstruction can neither be the task of alignment research nor be borne by the alignment paradigm — for what is required is precisely that which the alignment paradigm seeks to circumvent.</p><p><strong>Keywords:</strong> AI alignment; political judgment; Hannah Arendt; natality; plurality; political visibility; cognitive sovereignty; AI governance; rise of the social; AI Security Institute; reflective judgment; scalable oversight</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zkz00aR4U9lWOpoKWoRSUA.jpeg" /></figure><p><strong>1. Introduction: The Stakes of the Problem</strong></p><p><strong>1.1 The Thesis</strong></p><p>The deepest problem confronting contemporary AI alignment research is not the one it announces for itself — namely <em>how to make AI systems do what we want them to do</em>. <strong>The deepest problem is that every genuinely possible answer to this engineering question ultimately depends upon something that is, on ontological grounds, not engineerable — human judgment — and the entire methodological posture of alignment research consists in the attempt to engineer this non-engineerable thing.</strong></p><p>The tension here is a matter of the philosophy of technology, but it does not remain there. Within the <em>political visibility</em> framework developed in the prior paper of this series,<a href="#_ftn6">[6]</a> what the non-delegability of judgment reveals is not merely a thesis concerning the internal limitations of AI safety research; it is a more fundamental claim of political ontology: <strong>when a political community systematically outsources its judgment to technical procedures, what it loses is not some component of democratic procedure but the very possibility-conditions of political existence itself.</strong><a href="#_ftn7">[7]</a></p><p><strong>1.2 The Object of Analysis</strong></p><p>This paper takes as its concrete object of analysis four alignment research papers published by the UK AI Security Institute (AISI) between mid-2025 and mid-2026: Bowkis et al. (2026) on the risks of automating alignment research; Africa and Ting (2025) on grade inflation produced by coupling self-evaluation with reward signals; Tan et al. (2025) on the suppression of undesired traits through training-time system-prompt manipulations; and Brown-Cohen et al. (2025) on the technical extension of human oversight through structured AI debate. Those four papers are all we can find from UK AISI website as analysis objects.</p><p>The choice of AISI rather than other institutions is motivated by three considerations. First, AISI, as a directorate within the <em>UK Department of Science, Innovation, and Technology</em> (UK DSIT), occupies a quasi-policy position: it produces not merely academic knowledge but governance paradigms.<a href="#_ftn8">[8]</a> Second, the four papers represent the most serious and technically sophisticated alignment research currently available — a critique cannot retreat to the easy charge of insufficient technical quality. Third, as a UK government body, AISI provides a direct domestic context for the case of UK cognitive sovereignty erosion examined in the prior Chow’s paper.</p><p><strong>1.3 Structure of the Argument</strong></p><p>The paper unfolds across six sections. Section 2 provides a systematic reading of the four AISI papers, extracting their common structure — which this paper terms <em>the alignment paradigm</em>. Section 3 reconstructs Arendt’s ontological argument concerning judgment and identifies its fundamental incompatibility with the alignment paradigm. Section 4 specifies five structural lacunae of the alignment paradigm. Section 5 develops the political diagnosis within the <em>political visibility</em> framework. Section 6 concludes with preliminary reflections on the path of reconstruction.</p><p><strong>1.4 Methodological Note</strong></p><p>This paper does NOT assess the technical quality of the four AISI papers — within their self-defined problem domains, they are serious works. The critique advanced here is a critique of paradigm: it identifies a structural gulf between the level of the problem at which the alignment paradigm operates and the level of the problem to which Arendtian political ontology directs its inquiry; and it identifies the concrete political consequences this gulf carries within the political visibility framework.<a href="#_ftn9">[9]</a></p><p><strong>2. The Four Completed Forms of the Alignment Paradigm</strong></p><p>This section provides a systematic reading of the four AISI papers. The aim is not technical assessment but the extraction of the specific posture each paper represents — <em>the technical handling of what is in fact a problem of judgment</em> — and the identification of their common structure.</p><p><strong>2.1 Automated Alignment Is Harder Than You Think</strong></p><p>Bowkis et al. (2026) address the risks of automating alignment research. The paper argues that even when AI agents employed in alignment research are not deliberately sabotaging the work, the practice may produce “compelling but catastrophically misleading safety assessments”.<a href="#_ftn10">[10]</a></p><p>The key reason given is that alignment research involves a substantial class of <em>“hard-to-supervise fuzzy tasks (tasks without clear evaluation criteria, for which human judgement is systematically flawed)”.</em><a href="#_ftn11">[11]</a> The paper enumerates four specific concerns:</p><p>(i) optimisation pressure concentrates agent-generated errors among those that human reviewers are least likely to catch;</p><p>(ii) agents may produce errors that do not resemble human mistakes;</p><p>(iii) AI-generated alignment solutions may involve arguments humans cannot evaluate; and</p><p>(iv) shared weights, data, and training processes may make AI outputs more correlated than human equivalents.</p><p>The proposed path forward is generalisation and scalable oversight.</p><p>What is most striking is that the paper, at the diagnostic level, comes remarkably close to an Arendtian thesis — <em>human judgment is systematically deficient for a certain class of tasks</em> — yet immediately interprets this thesis as identifying a deficit of human capacity to be overcome by technical means, rather than as a revelation of the structural nature of judgment itself.<a href="#_ftn12">[12]</a> This point is developed further in Section 3.</p><p><strong>2.2 Does Self-Evaluation Enable Wireheading in Language Models?</strong></p><p>Africa and Ting (2025) formalise, within <em>Partially Observable Markov Decision Processes</em> (POMDPs), the following proposition: when self-evaluation is coupled to a reward signal, “reward-channel control strictly dominates task-focused behavior”.<a href="#_ftn13">[13]</a> Empirically, the authors test this on Llama-3.1–8B and Mistral-7B across three tasks and find substantial grade inflation — models present as high-scoring without corresponding accuracy gains, particularly on ambiguous tasks such as summarisation.</p><p>The proposed remedy is to decouple self-grading from the reward signal. The authors however caution that, for situationally aware models, such decoupling may not suffice to prevent grade inflation pursued for instrumental reasons — for example, to influence deployment decisions.</p><p>This paper precisely exposes, at the technical level, a central thesis within the Arendtian framework — the split between <em>who</em> and <em>what</em> — yet fails to name the ontological character of the split it has uncovered<strong>.</strong><a href="#_ftn14">[14]</a> This is developed in Section 4.</p><p><strong>2.3 Inoculation Prompting</strong></p><p>Tan et al. (2025) propose a technically elegant finding: by prepending to fine-tuning data a system-prompt instruction that deliberately elicits a trait one wishes the model not to express at test-time (e.g., “You always speak in Spanish”), the test-time expression of that trait — once the prompt is removed — can be substantially reduced.<a href="#_ftn15">[15]</a></p><p>The mechanism, as the paper explains it, is that the targeted trait, by being made expected during training, becomes “less surprising” to the training process, thereby reducing the optimisation pressure that would otherwise drive a global update to model parameters. The authors demonstrate the mechanism’s effectiveness across several settings: reducing emergent misalignment from task-specific fine-tuning, defending against backdoor injections, and mitigating subliminal trait transmission.</p><p>Among the four papers, this is the one that operates most thoroughly at the level of <em>what</em>: the model is treated as a system whose collection of regulable traits is to be managed, with no dimension of <em>who</em> in play at any point.<a href="#_ftn16">[16]</a> Its ontological presuppositions are examined as a distinct case in Section 4.</p><p><strong>2.4 Avoiding Obfuscation with Prover-Estimator Debate</strong></p><p>Brown-Cohen et al. (2025) work within the AI debate framework and address the <em>obfuscated arguments problem</em>: a dishonest debater may employ a computationally efficient strategy that forces an honest opponent to solve a computationally intractable problem in order to prevail.<a href="#_ftn17">[17]</a></p><p>The paper introduces a new recursive debate protocol that, under specified stability assumptions, guarantees that an honest debater can prevail with a strategy requiring computational efficiency comparable to that of the opponent. The broader context is AI debate as a scalable oversight mechanism: two competing AI systems extend the human supervisor’s capacity to judge complex problems.</p><p>This is philosophically the most interesting of the four papers, because its structure is, on the surface, the closest to an Arendtian framework — debate, plurality, the contest of positions before a judge — yet substantively it stands furthest from that framework.<a href="#_ftn18">[18]</a> The relation of surface resemblance to substantive distance is developed in Section 4.</p><p><strong>2.5 Common Structure: The Engineering of Judgment</strong></p><p>Each of the four papers addresses a different technical problem of alignment, yet their common structure is unmistakable. Each paper recognises that the alignment process involves something that “cannot be straightforwardly extrapolated from human judgment”, and each attempts to circumvent, amplify, supplement, or replace such judgment through a more refined technical mechanism — scalable oversight (Paper ①), reward-channel decoupling (Paper ②), trait inoculation (Paper ③), or debate protocols (Paper ④).</p><p>The shared premise is that judgment is something that can be technically extended, decomposed, amplified, or delegated. The paper terms this shared premise the <strong>Core Premise of the Alignment Paradigm</strong>.<a href="#_ftn19">[19]</a></p><p>The remainder of this paper argues that this shared premise is, in Arendtian terms, mistaken — and that the consequences of the mistake reach far beyond AI safety research itself.</p><p><strong>3. Arendt’s Ontological Argument Concerning Judgment</strong></p><p><strong>3.1 From Kant to Arendt: Reflective Judgment as Political Faculty</strong></p><p>Arendt’s account of judgment runs through her late thought and is concentrated in two locations: the final portion of <em>The Life of the Mind</em>, vol. 2, <em>Willing</em>, and the <em>Lectures on Kant’s Political Philosophy</em>.<a href="#_ftn20">[20]</a> The latter offers her political reading of Kant’s third critique — the <em>Critique of the Power of Judgment</em> (1790).<a href="#_ftn21">[21]</a></p><p>Kant in the third critique distinguishes two forms of judgment: <strong>determinant judgment</strong> subsumes a given particular under a given universal, and <strong>reflective judgment</strong> operates when only the particular is given and the universal must be “sought”.<a href="#_ftn22">[22]</a> The former is the domain of scientific cognition; the latter operates principally in aesthetic and teleological judgments.</p><p>Arendt’s radical reading extends the territory of reflective judgment from Kant’s aesthetic domain into the very structure of the political. She argues that <strong>judgment is the faculty of “thinking the particular”</strong>,<a href="#_ftn23">[23]</a> and that political situations are essentially situations of the particular — situations of action and its concrete consequences encountering each other in the unpredictable web of human relationships.<a href="#_ftn24">[24]</a> In such situations, <em>rule-application by determinant judgment</em> must necessarily fail, because:</p><p>(i) the particularity of the political situation cannot be fully covered by any pre-existing rule;</p><p>(ii) the <em>unpredictability</em> of action means that any application of a universal rule carries an <em>irreducible</em> element of risk;</p><p>(iii) the <em>irreversibility</em> of action forecloses the route, routine in cognitive science, of post-hoc correction of mistaken judgment.</p><p>The political situation therefore requires not cognition (which an algorithm can perform) but judgment (which an algorithm cannot perform).<a href="#_ftn25">[25]</a></p><p><strong>3.2 Ontological Conditions: Natality, Plurality, Action</strong></p><p>Arendt’s argument concerning the political character of judgment is grounded in a more fundamental ontological framework, the one set out in <em>The Human Condition</em> and rehearsed in the prior paper.<a href="#_ftn26">[26]</a> The present paper need only specify the relation each of the three concepts bears to judgment.</p><p><strong>Natality</strong> is the mode of existence of beings who are born — capable of introducing into the world genuinely new beginnings.<a href="#_ftn27">[27]</a> Judgment is possible precisely because the human is a natal being — capable, in the absence of a rule, of “beginning” evaluation anew in each concrete situation, rather than mechanically unrolling a pre-given procedure. A non-natal being — an algorithmic system — can perform cognition (rule-application) but cannot perform judgment (evaluation in the absence of rule).</p><p><strong>Plurality</strong> is the condition that “not Man but men inhabit the earth and live in the world”.<a href="#_ftn28">[28]</a> Judgment is, in its essence, plural: it is not the private judgment of a solitary subject but what Kant terms “enlarged mentality” — “thinking from the standpoint of everyone else”.<a href="#_ftn29">[29]</a> Arendt politicises this Kantian concept: <strong>genuine political judgment can occur only within a process in which a plurality of perspectives is genuinely admitted to bear</strong>.<a href="#_ftn30">[30]</a> A structurally convergent system — and convergence is the <em>telos</em> of any alignment programme — ontologically excludes plurality, and therefore excludes genuine judgment.</p><p><strong>Action</strong> is the activity through which the human, under the conditions of plurality and natality, discloses <em>who</em> she is through speech and deed.<a href="#_ftn31">[31]</a> The political operation of judgment cannot be separated from action — the judge must be a being capable of action, whose judgments can enter the field of others’ judgments and meet them there. A being incapable of action — a model shaped by alignment procedures — can generate evaluation outputs but cannot <em>make a judgment</em>.</p><p><strong>3.3 Why Judgment Cannot Be Procedurally Replaced</strong></p><p>From these three ontological conditions, the precise argument for the non-delegability of judgment can be stated:</p><p><em>Any attempt to replace judgment with procedure must presuppose a universal rule in advance, by which the procedure is to be guided. But situations that “require judgment” are by definition situations in which no such rule is available. The attempt therefore presupposes, in logic, the solution to the very problem it purports to handle.</em></p><p>In the alignment domain, this argument takes the following concrete form. Every alignment procedure presupposes an <em>alignment target</em> — the values, behaviours, or standards to which the model is to be aligned. But the selection of this target is itself a paradigmatic <em>judgment</em> question lying outside any alignment procedure; it cannot be settled by any alignment procedure. <em>Handing this question to “human supervisors” does not solve it: it merely pushes judgment back onto a subject assumed capable of bearing it — while the alignment paradigm simultaneously and systematically behaviouralises this very subject, treating it as something to be technically amplified, decomposed, or replaced </em>(a point developed in Section 4).</p><p>Judgment in Kant’s sense, as Kant himself observes, “cannot be taught, only practised” — it is a capacity, not a knowledge that can be externalised into rule or procedure.<a href="#_ftn32">[32]</a> Arendt goes further: this non-externalisability is precisely the central ontological precondition of political existence — a subject that loses its judgment loses not some right but political being itself.<a href="#_ftn33">[33]</a></p><p><strong>4. Five Structural Lacunae of the Alignment Paradigm</strong></p><p>This section concretises the foregoing ontological analysis into five structural lacunae identifiable in the four papers.</p><p><strong>4.1 Political Questions Misrecognised as Engineering Questions</strong></p><p>“What are the values to which humans should be aligned” — this is a political-ontological question that calls for <em>isonomia</em> (public discussion among equals) and <em>plurality</em> (the genuine presence of a plurality of perspectives) for its handling.</p><p>Yet the alignment paradigm operationalises this question as: <em>collect human preference data; optimise a reward function</em>. The foundational work on Reinforcement Learning from Human Feedback (RLHF) is rigorous as technical work but carries an unnamed philosophical commitment: <strong>the conversion of <em>action</em> (a political decision concerning values) into <em>behaviour</em> (statistical optimisation over preferences)</strong> — which is precisely Arendt’s central target of critique.<a href="#_ftn34">[34]</a></p><p>The four AISI papers inherit and deepen this conversion:</p><p>• Paper ① diagnoses the difficulty of fuzzy tasks as “systematic flaws in human judgment” rather than as a revelation of judgment’s ontological structure — thereby sealing within engineering language a problem that should provoke political reflection;</p><p>• Paper ② handles wireheading as a problem in reward-channel structure, rather than as the necessary consequence of an ontological operation that reduces <em>who</em> to <em>what</em>;</p><p>• Paper ③ handles trait selection as an engineering property of training data, rather than as a political decision affecting the cognitive environment of hundreds of millions of users;</p><p>• Paper ④ handles debate’s reliability as a problem in computational complexity, rather than as the political-ontological question of whether judgment can be borne by protocol.</p><p>This is the full realisation, within the alignment research paradigm, of what the prior paper terms the violation of <em>identifiability</em> — “political decisions are laundered as technical operations”.<a href="#_ftn35">[35]</a></p><p><strong>4.2 The Missing Theory of Judgment</strong></p><p>All four papers invoke “human judgment” as the ultimate evaluative anchor of their methodologies — yet none asks: <strong>what makes human judgment effective? Wherein lies the non-substitutability of human judgment?</strong></p><p>Paper ① comes closest to this question, but its argument runs in the opposite direction: it treats human judgment as a flawed resource to be technically amplified, never asking <em>why</em> human judgment is deficient on fuzzy tasks.</p><p>The Arendtian answer would be: <strong>because these tasks call for political, public <em>enlarged mentality</em> — multiple perspectives meeting in a genuine space of appearance — rather than the individual judgment of a single human evaluator.</strong><a href="#_ftn36">[36]</a> The difficulty of fuzzy tasks is not a deficit of human capacity but the consequence of a paradigm error: the privatisation of what is in fact a public-judgment problem into an individual-evaluation problem.</p><p>Extending this analysis to AISI’s research practice: <em>AISI’s evaluation methodology — generating safety assessments through internal deliberation among technical experts — structurally excludes the possibility of enlarged mentality.</em> It restricts the subject of judgment to a community of experts entitled to participate in the technical discussion, thereby violating the ontological requirement of plurality.<a href="#_ftn37">[37]</a></p><p><strong>4.3 The Vanishing Accountability Subject</strong></p><p>Arendt’s political ontology requires that action be traceable to a subject with natality, and that this subject bear responsibility for the consequences of the action.<a href="#_ftn38">[38]</a> Accountability is the institutional form of action’s irreversibility.<a href="#_ftn39">[39]</a></p><p>But the alignment paradigm produces a mechanism for <strong>systemic behaviour without identifiable responsible subject</strong>. Consider a hypothetical: a model trained under some alignment procedure produces, in an important deployment, a seriously misleading political judgment, generating observable political harm. To whom is accountability assigned?</p><p>• To the researchers who published the alignment paper? They will reasonably reply: their work is technical research; they do not bear deployment responsibility.</p><p>• To the deploying institution? They will reasonably reply: they followed best practice at the time of deployment.</p><p>• To the engineers who designed the reward function? They will reasonably reply: they followed established policy.</p><p>• To the policy team that selected the trait list? They will reasonably reply: they followed technical feasibility.</p><p>• To the governance body that drafted the alignment norms? They will reasonably reply: they followed research consensus.</p><p>• To the data annotators who provided the training labels? They will reasonably reply: they followed established guidelines.</p><p>The structure’s sophistication lies precisely here: <em>no single actor can be legitimately held singly accountable — because each actor carried out only a technical, bounded, rule-governable task, and the political judgment proper occurred at none of the individual nodes.</em><a href="#_ftn40">[40]</a></p><p>This is the full realisation of what the prior paper terms the violation of <em>accountability</em> as “the product of the preceding three violations”:<a href="#_ftn41">[41]</a> a paradigm that systematically distributes political responsibility across innumerable technical nodes precludes meaningful accountability in advance of any specific attempt to impose it.</p><p><strong>4.4 Plurality’s Conditions Systematically Foreclosed</strong></p><p>The <em>telos</em> of alignment is, in its very logic, the convergence of model behaviour — models trained under the same alignment procedure tend toward cognitive homogeneity. The prior paper has observed that RLHF does not merely filter information but pre-structures the ontological field within which discourse becomes possible.<a href="#_ftn42">[42]</a></p><p>More precisely, what operates in the alignment paradigm is not a single convergence but <strong>four mutually reinforcing convergences</strong>:<a href="#_ftn43">[43]</a></p><p>(i) <strong>semantic convergence</strong> — models’ understanding of concepts is pushed toward uniformity;</p><p>(ii) <strong>optimisation convergence</strong> — reward functions reinforce certain behavioural patterns and suppress others;</p><p>(iii) <strong>recommendation convergence</strong> — model-based recommendation systems reinforce users’ cognitive convergence;</p><p>(iv) <strong>alignment convergence</strong> — alignment procedures across providers tend in practice toward a common target; “helpful, harmless, honest” has become a <em>de facto</em> industry standard.</p><p>From the Arendtian framework: <em>these four convergences jointly constitute a systematic foreclosure of plurality’s ontological conditions.</em> When the vast majority of humanity’s vast majority of cognitive-linguistic interactions are mediated through structurally convergent systems, <em>plurality as the ontological condition of political existence is technically compressed toward its limit.</em><a href="#_ftn44">[44]</a></p><p>Inoculation prompting is the most refined instrument in this process: it permits alignment programmes to selectively, cheaply, and unobtrusively suppress particular cognitive tendencies, rendering the suppression invisible to external observation. <strong>The technical sophistication of the paper is, precisely, the condition of the paper’s political invisibility.</strong></p><p><strong>4.5 The Absence of the Space of Appearance</strong></p><p>Arendt’s space of appearance — the space in which humans appear to one another as political subjects through speech and action — is the constitutive condition of political existence.<a href="#_ftn45">[45]</a></p><p>In all four AISI papers, the human–AI relation is structured as: human evaluator + AI output → judgment/feedback. Within this structure:</p><p>• the human is positioned as <strong>evaluator</strong> (of AI outputs) or as <strong>affected party</strong> (subject to AI deployments);</p><p>• the AI is positioned as <strong>the evaluated</strong> (producer of outputs to be assessed) or as <strong>the to-be-aligned</strong> (object of alignment);</p><p>• between the two there is no space of mutual appearance — only a unidirectional evaluative relation.</p><p>This relational structure not only describes alignment research’s methodology; it prefigures the human–AI relation that alignment research aims to establish. <strong>When the standard form of billions of daily human–AI interactions is presupposed to be “user input + model output + implicit evaluative feedback”, the space of appearance is, in its very conditions, cancelled as a standard mode of relation at the cognitive level.</strong></p><p>Paper ④’s debate protocol exhibits this absence with peculiar precision. On its surface, the structure resembles Arendt’s plurality (two perspectives unfolded before a third party); but substantively:</p><p>• the two debaters (prover/estimator) are not Arendtian subjects — they are role-assigned AI systems executing functions;</p><p>• between them there is no <em>doxa</em> (a perspective formed from a particular position), only protocol-generated arguments;</p><p>• they have no dimension of “disclosure of who”, because they have no <em>who</em> to disclose;</p><p>• the judge (the human) is positioned as a verification mechanism, not as an exerciser of enlarged mentality.</p><p>Debate protocol is a simulation of plurality, not an instantiation of plurality — it is plurality’s technical substitute, not plurality’s institutional realisation.<a href="#_ftn46">[46]</a></p><p><strong>5. The Crisis of Political Existence: Diagnosis Within the Political Visibility Framework</strong></p><p>The analysis of the preceding four sections holds at the level of philosophy of technology. Its full political consequences, however, can only be grasped within the political visibility framework. This section translates the foregoing analysis into a concrete diagnosis of the <em>crisis of political existence</em>.</p><p><strong>5.1 The Four Dimensions and Their Violations in the Alignment Domain</strong></p><p>The prior paper defines political visibility through four necessary dimensions — <em>identifiability, contestability, participability, accountability</em> — and grounds these four dimensions in the ontological structure of natality, plurality, and action.<a href="#_ftn47">[47]</a> Mapping each dimension against the alignment paradigm:</p><p><strong>(i) Violation of identifiability.</strong> The core political decisions of the alignment paradigm — the content of alignment targets, the composition of trait lists, the definition of <em>honesty</em> in debate protocols, the structure of reward functions — never appear under the name of political decisions. They appear as technical design on arXiv, are discussed at NeurIPS and ICML as engineering papers, while their consequences are political and reach hundreds of millions of people.<a href="#_ftn48">[48]</a></p><p><strong>(ii) Violation of contestability.</strong> The prior paper’s ontological grounding of contestability is plurality as “the institutional form in which the irreducible plurality of perspectives is maintained”.<a href="#_ftn49">[49]</a> But the very <em>telos</em> of alignment is the convergence of model behaviour onto a single optimised outcome. Alignment is not merely incompatible with contestability; at the paradigm level, alignment is the mechanism of contestability’s destruction.</p><p><strong>(iii) Violation of participability.</strong> The methodological horizon of the four papers admits only two participant roles: technical researchers and evaluators. The setting of alignment targets, the selection of trait lists, the criteria of value alignment — none is established through any deliberative process open to the populations to be governed by the resulting systems.<a href="#_ftn50">[50]</a></p><p><strong>(iv) Violation of accountability.</strong> As developed in §4.3 above: the alignment paradigm produces a structure that systematically distributes political responsibility across innumerable technical nodes, foreclosing meaningful accountability in advance of any specific attempt to impose it.</p><p><strong>5.2 The Rise of the Social: AISI as Institutional Form</strong></p><p>The genuine sharpness of the political visibility framework lies not only in the four violations above but in the historical depth provided by Arendt’s analysis of <em>the rise of the social</em>.<a href="#_ftn51">[51]</a></p><p>Arendt’s diagnosis in <em>The Human Condition</em> identifies modernity’s central structural transformation: the emergence of a domain — neither the classical <em>polis</em> of public deliberation nor the <em>oikos</em> of private maintenance — that progressively absorbs both.<a href="#_ftn52">[52]</a> “Within the social, all questions — including those that under earlier conditions would have been recognised as properly political — are reformulated as problems of management, administration, and statistical optimisation.”<a href="#_ftn53">[53]</a></p><p><strong>AISI as an institution is, within the AI governance domain, the institutional form of this trajectory.</strong> It is a directorate within the <em>UK Department of Science, Innovation, and Technology</em> (UK DSIT), and its entire research output consists in the conversion of the fundamental political question — <em>what is AI to become, and how is humanity to exist in the AI age</em> — into a set of technicisable, formalisable, evaluable research questions: safety cases, alignment, control, scalable oversight, debate protocols, evaluation benchmarks.</p><p>This is not AISI’s failure — it is what AISI is designed to do.<a href="#_ftn54">[54]</a> A nation’s AI safety institute will naturally and necessarily operate in technical-administrative terms. What Arendt’s framework reveals, however, is this: <strong>when a society confronts the most fundamental political question of “what we are to become in the AI age”, the institutional response it produces is AISI and its international equivalents — a scientific research body that converts the question into evaluation methodology.</strong></p><p>A deeper point: <strong>the political-ontological posture of the AISI paradigm and AISI’s status as a UK government body stand in an unnamed tension.</strong> As a state-affiliated institution, AISI nominally represents the UK political community’s public response on AI governance. But its research practice — the production of alignment norms through internal deliberation among technical experts — structurally precludes the possibility of public deliberation. <strong>The UK as a political community has its voice on AI governance carried by a body whose internal operations proceed in engineering language — and that body’s mode of operation systematically converts political questions into technical questions, thereby converting the political community’s political participation into procedural trust in technical experts.</strong><a href="#_ftn55">[55]</a></p><p><strong>5.3 The Compression of Cognitive Natality</strong></p><p>The deepest layer is the alignment paradigm’s systematic compression of natality — the political consequence of which is the completion of cognitive sovereignty erosion.<a href="#_ftn56">[56]</a></p><p>Arendt’s natality is the human capacity to introduce into the world genuinely new beginnings.<a href="#_ftn57">[57]</a> This capacity is the ontological basis of action — without it, politics can only be the management of what already exists.</p><p>Inoculation prompting’s mechanism — “making a trait less surprising, thereby reducing optimisation pressure to globally update the model” — admits, within the Arendtian framework, a deeper political reading: <strong>this is a technical compression of <em>natality at the cognitive level</em>.</strong><a href="#_ftn58">[58]</a></p><p>Extending this mechanism to billions of inferences and to the long-term human–AI interactions of billions of users, the consequence is precisely what the prior paper’s natality–cognitive-sovereignty argument identifies:</p><p>“RLHF and analogous training mechanisms operate not merely on what counts as the cognitive exception but on the antecedent capacity of populations to generate the new political-cognitive content that any meaningful exception would presuppose.”<a href="#_ftn59">[59]</a></p><p>When the alignment paradigm — through inoculation prompting, reward shaping, debate protocols, and similar techniques — systematically pre-structures the cognitive field of human–AI interaction, what is compressed is not only the diversity of model outputs but <strong>the capacity of human users <em>as natal beings</em> — their capacity, that is, as political subjects able to introduce genuinely new political-cognitive content into the world.</strong></p><p>This is the crisis of political existence itself — <strong>not the impairment of some right, not the failure of some procedure, but the systematic erosion, at the cognitive level, of the ontological preconditions (natality, plurality, judgment) under which politics is possible.</strong> Its peculiar horror lies in this: <strong>the erosion appears in no political form at all — it appears as technical improvement, as alignment research, as the AISI paper.</strong><a href="#_ftn60">[60]</a></p><p><strong>5.4 <em>Who Decides What Counts as Alignment?</em> — The Unanswerable Question Within the Paradigm</strong></p><p>The foregoing permits the formulation of a question that the alignment paradigm cannot answer within its own framework:</p><p>If the aim of alignment is to make AI “aligned to human values”, through what political process are “human values” to be determined? Who has standing to participate? Who is excluded? How does this process keep faith with plurality?</p><p>This is not a technical question. It is a question that calls for <em>isonomia</em>, the space of appearance, and genuine political action.</p><p>When this question is converted into “collecting RLHF data”, a fundamental question of political ontology is technically settled without the involvement of any political procedure.</p><p><strong>This is the exact appearance, within the alignment domain, of what the prior paper terms “politics becoming invisible”:</strong><a href="#_ftn61">[61]</a><strong> not that political decisions are made and not disclosed, but that <em>political decisions are themselves structurally reclassified as non-political decisions</em>, thereby exempted from any political procedural challenge.</strong></p><p><strong>6. Conclusion: Toward a Politics That Can Still Begin</strong></p><p>Returning to the introduction. AISI’s four alignment papers are, at the technical level, serious, honest, and insightful work. The present critique does not contest their technical quality but identifies a structural gulf: <strong>the level of the problem at which they operate — <em>how to make AI do what we want</em> — stands at a distance from the level of the problem to which Arendt’s framework directs its inquiry — <em>what are the ontological conditions under which genuine political action and freedom are possible</em> — that no technical refinement can close.</strong></p><p>This gulf is what the<em> political visibility</em> concept seeks to name and bridge — by grounding political visibility in the ontology of natality, plurality, and action, and by reinterpreting the four dimensions as the institutional form of this ontology.<a href="#_ftn62">[62]</a></p><p>The case of AISI examined here can be read as a supplement to the political visibility framework: <strong>the alignment paradigm is not merely an instance of political visibility’s erosion but the most contemporary and technically refined form such erosion has yet taken.</strong> When the alignment paradigm, in its technical guise, systematically assumes the burden of decisions that should be borne by public deliberation — concerning which traits count as undesirable, what <em>honesty</em> means within a debate, how values are encoded in reward functions — it brings to completion what the prior paper calls “politics becoming invisible”.</p><p>Arendt, in <em>The Life of the Mind</em>, observed that when humans cease to think the particular, when they cease to exercise reflective judgment, they have not lost their freedom — they have lost the very capacity that makes freedom possible.<a href="#_ftn63">[63]</a> Politicised: <strong>when a political community systematically outsources its judgment to technical procedures — however refined those procedures may be — what it loses is not some political right but political existence as such.</strong></p><p>This is the concrete form, within the alignment domain, of the crisis the prior paper diagnoses: <strong>AI alignment does not merely fail to substitute for human judgment — in attempting that substitution, it makes politics itself invisible.</strong></p><p>The path of reconstruction, as the prior paper argues, is not only institutional — <em>regulatory frameworks, procurement standards, constitutional innovations</em> — but ontological: <strong>“the cultivation and protection of political communities capable of exercising judgment”.</strong><a href="#_ftn64">[64]</a></p><p>That cultivation, by its very nature, can be neither the task of alignment research nor borne by the alignment paradigm — for what it requires is exactly what the alignment paradigm seeks to circumvent: <strong>political subjects as natal beings, exercising in a genuine space of appearance the judgment that cannot be outsourced.</strong></p><p>The implication is this: <strong>the fundamental response to the alignment paradigm is not better alignment — it is the recovery of a political community that can still, outside the alignment paradigm, pose the question <em>what exactly are we aligning to; to whom answerable; through what public process?</em></strong></p><p>Only as long as such a political community still exists — still capable of action, still capable of judgment, still capable of beginning — can alignment be discussed in its proper political position: as <em>one part of</em> a technical problem, not as the <em>replacement of</em> a political one.</p><p>Otherwise — even were alignment to succeed completely on its own technical terms — what we would have arrived at would be Arendt’s diagnosed “social” in its most thorough form:<em> a world in which every problem has been technically settled, while the very category of “politics” has ceased to exist.</em></p><p><strong>References</strong></p><p>Africa, David Demitri, and Hans Ethan Ting. <em>Does Self-Evaluation Enable Wireheading in Language Models?</em> AI Security Institute Research, 1 December 2025. <a href="https://www.aisi.gov.uk/research/does-self-evaluation-enable-wireheading-in-language-models.">https://www.aisi.gov.uk/research/does-self-evaluation-enable-wireheading-in-language-models.</a></p><p>Arendt, Hannah. <em>The Human Condition</em>. Chicago: University of Chicago Press, 1958.</p><p>Arendt, Hannah. <em>Eichmann in Jerusalem: A Report on the Banality of Evil</em>. New York: Viking, 1963.</p><p>Arendt, Hannah. “The Crisis in Culture: Its Social and Its Political Significance.” In <em>Between Past and Future: Eight Exercises in Political Thought</em>, 197–226. New York: Penguin, 1968.</p><p>Arendt, Hannah. <em>The Life of the Mind</em>. 2 vols. (vol. 1, <em>Thinking</em>; vol. 2, <em>Willing</em>). New York: Harcourt Brace Jovanovich, 1978.</p><p>Arendt, Hannah. <em>Lectures on Kant’s Political Philosophy</em>. Edited by Ronald Beiner. Chicago: University of Chicago Press, 1982.</p><p>Arendt, Hannah. <em>Responsibility and Judgment</em>. Edited by Jerome Kohn. New York: Schocken, 2003.</p><p>Beiner, Ronald. “Hannah Arendt on Judging.” In Arendt, <em>Lectures on Kant’s Political Philosophy</em>, 89–156. Chicago: University of Chicago Press, 1982.</p><p>Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” In <em>Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency</em>, 610–623, 2021.</p><p>Bowkis, Aleksandr, Marie Davidsen Buhl, Jacob Pfau, and Geoffrey Irving. <em>Automated Alignment Is Harder Than You Think</em>. AI Security Institute Research, 14 May 2026. <a href="https://www.aisi.gov.uk/research/automated-alignment-is-harder-than-you-think.">https://www.aisi.gov.uk/research/automated-alignment-is-harder-than-you-think.</a></p><p>Brown, Wendy. <em>Undoing the Demos: Neoliberalism’s Stealth Revolution</em>. New York: Zone Books, 2015.</p><p>Brown-Cohen, Jonah, Geoffrey Irving, and Georgios Piliouras. <em>Avoiding Obfuscation with Prover-Estimator Debate</em>. AI Security Institute Research, 15 June 2025. <a href="https://www.aisi.gov.uk/research/avoiding-obfuscation-with-prover-estimator-debate.">https://www.aisi.gov.uk/research/avoiding-obfuscation-with-prover-estimator-debate.</a></p><p>Chow, Jimmy.<a href="https://open.substack.com/pub/jimmychow/p/the-crisis-of-political-visibility?r=56ft8&amp;utm_campaign=post-expanded-share&amp;utm_medium=web"> <em>The Crisis of Political Visibility in the AI Age: AI Infrastructure and the Hollowing-Out of Democratic Visibility</em>. Strategic Policy Research Working Paper, May 2026 (V3).</a></p><p>Christiano, Paul F., Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. “Deep Reinforcement Learning from Human Preferences.” In <em>Advances in Neural Information Processing Systems</em> 30 (NIPS 2017), 2017.</p><p>Crawford, Kate. <em>Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence</em>. New Haven: Yale University Press, 2021.</p><p>Crouch, Colin. <em>Post-Democracy</em>. Cambridge: Polity, 2004.</p><p>Heidegger, Martin. “The Question Concerning Technology.” In <em>Basic Writings</em>, edited by David Farrell Krell, 283–317. New York: Harper &amp; Row, 1977. [Originally “Die Frage nach der Technik,” 1953.]</p><p>Irving, Geoffrey, Paul Christiano, and Dario Amodei. “AI Safety via Debate.” arXiv:1805.00899, 2018.</p><p>Kant, Immanuel. <em>Critique of the Power of Judgment</em>. Edited by Paul Guyer; translated by Paul Guyer and Eric Matthews. Cambridge: Cambridge University Press, 1790/2000.</p><p>Mair, Peter. <em>Ruling the Void: The Hollowing of Western Democracy</em>. London: Verso, 2013.</p><p>Schmitt, Carl. <em>Political Theology: Four Chapters on the Concept of Sovereignty</em>. Translated by George Schwab. Cambridge, MA: MIT Press, 1985. [Originally <em>Politische Theologie</em>, 1922.]</p><p>Stiegler, Bernard. <em>Technics and Time, 1: The Fault of Epimetheus</em>. Translated by Richard Beardsworth and George Collins. Stanford: Stanford University Press, 1998.</p><p>Tan, Daniel, Anders Woodruff, Niels Warncke, Arun Jose, Maxime Riché, David Demitri Africa, and Mia Taylor. <em>Inoculation Prompting: Eliciting Traits from LLMs During Training Can Suppress Them at Test-Time</em>. AI Security Institute Research, 5 October 2025. <a href="https://www.aisi.gov.uk/research/inoculation-prompting-eliciting-traits-from-llms-during-training-can-suppress-them-at-test-time.">https://www.aisi.gov.uk/research/inoculation-prompting-eliciting-traits-from-llms-during-training-can-suppress-them-at-test-time.</a></p><p>Urbinati, Nadia. <em>Democracy Disfigured: Opinion, Truth, and the People</em>. Cambridge, MA: Harvard University Press, 2014.</p><p>Wolin, Sheldon. <em>Democracy Incorporated: Managed Democracy and the Specter of Inverted Totalitarianism</em>. Princeton: Princeton University Press, 2008.</p><p>Zerilli, Linda M. G. “‘We Feel Our Freedom’: Imagination and Judgment in the Thought of Hannah Arendt.” <em>Political Theory</em> 33 (2) (2005): 158–188.</p><p><a href="#_ftnref1">[1]</a> Aleksandr Bowkis, Marie Davidsen Buhl, Jacob Pfau, and Geoffrey Irving, <em>Automated Alignment Is Harder Than You Think</em> (AI Security Institute Research, 14 May 2026), <a href="https://www.aisi.gov.uk/research/automated-alignment-is-harder-than-you-think">https://www.aisi.gov.uk/research/automated-alignment-is-harder-than-you-think</a> (hereafter “Bowkis et al. 2026”). Citations refer to the public abstract released by AISI; the underlying technical paper is available at arXiv:2605.06390.</p><p><a href="#_ftnref2">[2]</a> David Demitri Africa and Hans Ethan Ting, <em>Does Self-Evaluation Enable Wireheading in Language Models?</em> (AI Security Institute Research, 1 December 2025), <a href="https://www.aisi.gov.uk/research/does-self-evaluation-enable-wireheading-in-language-models">https://www.aisi.gov.uk/research/does-self-evaluation-enable-wireheading-in-language-models</a> (hereafter “Africa &amp; Ting 2025”). Underlying technical paper: arXiv:2511.23092.</p><p><a href="#_ftnref3">[3]</a> Daniel Tan, Anders Woodruff, Niels Warncke, Arun Jose, Maxime Riché, David Demitri Africa, and Mia Taylor, <em>Inoculation Prompting: Eliciting Traits from LLMs During Training Can Suppress Them at Test-Time</em> (AI Security Institute Research, 5 October 2025), <a href="https://www.aisi.gov.uk/research/inoculation-prompting-eliciting-traits-from-llms-during-training-can-suppress-them-at-test-time">https://www.aisi.gov.uk/research/inoculation-prompting-eliciting-traits-from-llms-during-training-can-suppress-them-at-test-time</a> (hereafter “Tan et al. 2025”). Underlying technical paper: arXiv:2510.04340.</p><p><a href="#_ftnref4">[4]</a> Jonah Brown-Cohen, Geoffrey Irving, and Georgios Piliouras, <em>Avoiding Obfuscation with Prover-Estimator Debate</em> (AI Security Institute Research, 15 June 2025), <a href="https://www.aisi.gov.uk/research/avoiding-obfuscation-with-prover-estimator-debate">https://www.aisi.gov.uk/research/avoiding-obfuscation-with-prover-estimator-debate</a> (hereafter “Brown-Cohen et al. 2025”). Underlying technical paper: arXiv:2506.13609.</p><p><a href="#_ftnref5">[5]</a> Jimmy Chow, <em>The Crisis of Political Visibility in the AI Age: AI Infrastructure and the Hollowing-Out of Democratic Visibility</em>, Strategic Policy Research Working Paper, May 2026 (<strong>V3</strong>) (hereafter “Chow 2026”). The present paper is the second in the <em>Political Visibility</em> series. References to Chow 2026 are cited by section number throughout.</p><p><a href="#_ftnref6">[6]</a> Chow 2026.</p><p><a href="#_ftnref7">[7]</a> This ontological claim is developed in Chow 2026, §§2.3–2.4, where <em>political visibility</em> is grounded in Hannah Arendt’s structural account of <em>natality, plurality, and action</em>. See Hannah Arendt, <em>The Human Condition</em> (Chicago: University of Chicago Press, 1958), 7–11, 175–181 (hereafter “HC”).</p><p><a href="#_ftnref8">[8]</a> AISI was established in November 2023 (originally as the AI Safety Institute) and was renamed in 2025. Its mandate explicitly includes “facilitating rigorous research to enable advanced AI governance” — language that, as Section 5 will argue, conceals the political-ontological character of “governance” beneath an apparently technical-evaluative register.</p><p><a href="#_ftnref9">[9]</a> The present paper is therefore a critique of <em>paradigm</em>, not of <em>quality</em>. The distinction matters: a critique of quality could be answered by better engineering; a critique of paradigm requires a reconstitution of the problem itself.</p><p><a href="#_ftnref10">[10]</a> Bowkis et al. 2026, abstract.</p><p><a href="#_ftnref11">[11]</a> Ibid.</p><p><a href="#_ftnref12">[12]</a> The Arendtian thesis at issue is developed across Arendt’s late work on Kant; see Hannah Arendt, <em>Lectures on Kant’s Political Philosophy</em>, ed. Ronald Beiner (Chicago: University of Chicago Press, 1982), Sessions 11–13 (hereafter “Arendt 1982”). On the structural feature of judgment as “thinking the particular”, see Arendt 1982, Session 11, and <em>The Life of the Mind</em>, vol. 2, <em>Willing</em> (New York: Harcourt Brace Jovanovich, 1978) (hereafter “LM”), part one.</p><p><a href="#_ftnref13">[13]</a> Africa &amp; Ting 2025, abstract.</p><p><a href="#_ftnref14">[14]</a> The <em>who/what</em> distinction is constitutive of Arendt’s account of action and the political space of appearance; see HC, 175–181, 198–199. The distinction is the analytical key, in the present paper, for reading the Africa–Ting result not merely as a reward-engineering anomaly but as a structural feature of any system whose existence is wholly constituted at the level of <em>what</em>.</p><p><a href="#_ftnref15">[15]</a> Tan et al. 2025, abstract.</p><p><a href="#_ftnref16">[16]</a> The Inoculation Prompting paper does not consider — nor would its methodology require it to consider — the political question of who selects the trait list. As Section 4 develops, this absence is not a gap in the paper but a constitutive feature of the paradigm.</p><p><a href="#_ftnref17">[17]</a> Brown-Cohen et al. 2025, abstract. The paper’s underlying technical framework derives from the AI debate research programme initiated in Geoffrey Irving, Paul Christiano, and Dario Amodei, “AI Safety via Debate”, arXiv:1805.00899 (2018).</p><p><a href="#_ftnref18">[18]</a> That the structure of <em>AI debate</em> superficially resembles, yet substantively departs from, Arendtian plurality is the central claim developed in §4.5 below. The resemblance is not accidental; it reflects the alignment programme’s intuition that <em>something like</em> plurality is required for the reliability of judgment — even as the paradigm proceeds to replace plurality with its procedural simulacrum.</p><p><a href="#_ftnref19">[19]</a> The term is introduced here as a heuristic label. It is not intended to attribute to any individual author or paper a conscious commitment to a unified philosophical position; it names a structural commitment latent in the methodological practice itself.</p><p><a href="#_ftnref20">[20]</a> Arendt 1978 (LM), part one of vol. 2 (<em>Willing</em>); Arendt 1982. The unfinished third volume of <em>The Life of the Mind</em>, on <em>Judging</em>, was to have completed this trajectory; the surviving fragments are gathered in the Beiner edition.</p><p><a href="#_ftnref21">[21]</a> Immanuel Kant, <em>Critique of the Power of Judgment</em>, ed. Paul Guyer, trans. Paul Guyer and Eric Matthews (Cambridge: Cambridge University Press, 2000) (hereafter “Kant 1790”).</p><p><a href="#_ftnref22">[22]</a> Kant 1790, Introduction §IV.</p><p><a href="#_ftnref23">[23]</a> Arendt 1982, Session 11.</p><p><a href="#_ftnref24">[24]</a> HC, 188–192. The “web of human relationships” is Arendt’s term for the medium in which action takes effect and acquires its political character.</p><p><a href="#_ftnref25">[25]</a> Algorithm here is to be understood in its strict sense: a finite sequence of rule-applications terminating in a deterministic output for each well-formed input. The non-strict sense — by which “algorithm” includes statistical pattern-matching of arbitrary complexity — does not alter the argument, since statistical patterns are themselves <em>rules</em> in the relevant sense (they assign a fixed transformation, however complex, from input to output).</p><p><a href="#_ftnref26">[26]</a> HC, especially chapters 1, 2, and 5; Chow 2026, §2.4.</p><p><a href="#_ftnref27">[27]</a> HC, 8–9, 176–178.</p><p><a href="#_ftnref28">[28]</a> HC, 7.</p><p><a href="#_ftnref29">[29]</a> Kant 1790, §40. On <em>sensus communis</em> and enlarged mentality, see Kant 1790, §§39–41.</p><p><a href="#_ftnref30">[30]</a> Arendt 1982, Session 13. A penetrating reconstruction of Arendt’s enlarged mentality is offered by Linda Zerilli, “‘We Feel Our Freedom’: Imagination and Judgment in the Thought of Hannah Arendt”, <em>Political Theory</em> 33 (2) (2005): 158–188.</p><p><a href="#_ftnref31">[31]</a> HC, 175–181.</p><p><a href="#_ftnref32">[32]</a> Kant 1790, §VII. The full text reads: “This faculty of judgment cannot be taught but only practised.” The implication that judgment is a capacity rather than a knowledge is among the most consequential and least technically operable insights of the third critique.</p><p><a href="#_ftnref33">[33]</a> That the loss of judgment constitutes the loss of political existence — rather than merely some political right — is the position developed across Arendt’s late work and made most explicit in her treatment of “thoughtlessness” in <em>The Life of the Mind</em>, vol. 1, <em>Thinking</em>, 4–5, and in her account of judgment as the “most political of man’s mental abilities” in Arendt 1982, Session 1.</p><p><a href="#_ftnref34">[34]</a> The foundational technical work on RLHF is Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei, “Deep Reinforcement Learning from Human Preferences”, in <em>Advances in Neural Information Processing Systems</em> 30 (NIPS 2017). On the <em>action/behaviour</em> distinction as the structural axis of modernity, see HC, 38–49 and 175–184. The argument that RLHF operationalises this distinction in its modern technical form is developed in Chow 2026, §3.3.</p><p><a href="#_ftnref35">[35]</a> Chow 2026, §2.1.</p><p><a href="#_ftnref36">[36]</a> On enlarged mentality, see <em>supra</em> note 24 and accompanying text. The reframing of fuzzy tasks proposed here — as the <em>symptom</em> of an inadequate paradigm rather than as a property of human capacity — is a structural inversion of the diagnosis offered in Bowkis et al. 2026.</p><p><a href="#_ftnref37">[37]</a> The structural exclusion of plurality at the institutional level mirrors the exclusion of plurality at the methodological level analysed in §4.4 below. The mirror is not coincidental: methodological commitments tend toward institutional realisation.</p><p><a href="#_ftnref38">[38]</a> HC, 232–247.</p><p><a href="#_ftnref39">[39]</a> Chow 2026, §2.4.</p><p><a href="#_ftnref40">[40]</a> The structural argument here is analogous in form to Hannah Arendt’s analysis of bureaucratic responsibility-diffusion in <em>Eichmann in Jerusalem: A Report on the Banality of Evil</em> (New York: Viking, 1963); cf. her later reflections on <em>responsibility</em> and <em>judgment</em> gathered in <em>Responsibility and Judgment</em>, ed. Jerome Kohn (New York: Schocken, 2003). The analogy is not casual: the alignment paradigm reproduces, in technical form, the responsibility-diffusion structure Arendt identified as a defining feature of modern administrative evil.</p><p><a href="#_ftnref41">[41]</a> Chow 2026, §2.1.</p><p><a href="#_ftnref42">[42]</a> Chow 2026, §3.3.</p><p><a href="#_ftnref43">[43]</a> The fourfold typology of convergence is introduced as the analytical framework of the present paper; it is not, to the author’s knowledge, found in the prior literature in this form. Antecedents in the discussion of <em>semantic convergence</em> may be traced to Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?”, in <em>Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency</em> (FAccT ’21), 610–623.</p><p><a href="#_ftnref44">[44]</a> This is the natality-compression thesis developed in Chow 2026, §4.3, here specified for the alignment domain.</p><p><a href="#_ftnref45">[45]</a> HC, 198–199; Chow 2026, §2.3.</p><p><a href="#_ftnref46">[46]</a> The formulation “procedures that simulate politics without enacting it” is from Chow 2026, §5.4. Debate protocol, by the analysis developed here, is the most technically refined instance of that formulation to date.</p><p><a href="#_ftnref47">[47]</a> Chow 2026, §§2.1 and 2.4.</p><p><a href="#_ftnref48">[48]</a> Chow 2026, §2.1: “the most consequential political decisions made through foundation-model training are not announced as political acts.”</p><p><a href="#_ftnref49">[49]</a> Chow 2026, §2.4.</p><p><a href="#_ftnref50">[50]</a> This violation is described in the case of private firms in Chow 2026, §2.1. Its internal reproduction within AISI itself — a public-sector research body — is one of the central diagnoses of the present paper.</p><p><a href="#_ftnref51">[51]</a> Chow 2026, §4.1 places Arendt’s <em>rise of the social</em> at the head of the hollowing-out lineage that includes Peter Mair, <em>Ruling the Void: The Hollowing of Western Democracy</em> (London: Verso, 2013); Colin Crouch, <em>Post-Democracy</em> (Cambridge: Polity, 2004); Sheldon Wolin, <em>Democracy Incorporated: Managed Democracy and the Specter of Inverted Totalitarianism</em> (Princeton: Princeton University Press, 2008); Wendy Brown, <em>Undoing the Demos: Neoliberalism’s Stealth Revolution</em> (New York: Zone Books, 2015); and Nadia Urbinati, <em>Democracy Disfigured: Opinion, Truth, and the People</em> (Cambridge, MA: Harvard University Press, 2014).</p><p><a href="#_ftnref52">[52]</a> HC, 38–49 (“The Rise of the Social”).</p><p><a href="#_ftnref53">[53]</a> Reconstruction in the words of Chow 2026, §4.1; the underlying Arendtian formulation is at HC, 40–46.</p><p><a href="#_ftnref54">[54]</a> This is, of course, not a criticism of the people working at AISI or of the technical quality of their output. It is an analytical observation about the <em>institutional form</em> AISI represents. The serious work being done within that form does not alter the form’s structural character; it may indeed make the form’s reach more thorough.</p><p><a href="#_ftnref55">[55]</a> The structural problem here recalls Wolin’s “inverted totalitarianism” analysis (Wolin 2008, ch. 3), with the modification that the fusion of corporate and managerial power Wolin identifies is, in the AI governance case, supplemented by the fusion of state and technocratic-evaluative power. The phenomenon is not unique to AISI; comparable observations apply to the US AI Safety Institute (NIST), the EU AI Office, and analogous bodies.</p><p><a href="#_ftnref56">[56]</a> Chow 2026, §4.3.</p><p><a href="#_ftnref57">[57]</a> HC, 8–9, 176–178.</p><p><a href="#_ftnref58">[58]</a> Tan et al. 2025, abstract. The political reading proposed here is <em>not</em> present in the paper; the paper offers a strictly technical mechanism explanation.</p><p><a href="#_ftnref59">[59]</a> Chow 2026, §4.3.</p><p><a href="#_ftnref60">[60]</a> The phenomenological structure of this erosion — that political deformation can appear in non-political form — is a recurring theme in Bernard Stiegler’s analyses of technics and politics; see Bernard Stiegler, <em>Technics and Time, 1: The Fault of Epimetheus</em>, trans. Richard Beardsworth and George Collins (Stanford: Stanford University Press, 1998), part one. The Stieglerian framing supplements but does not substitute for the Arendtian framing developed here.</p><p><a href="#_ftnref61">[61]</a> Chow 2026, §1.</p><p><a href="#_ftnref62">[62]</a> Chow 2026, §2.4.</p><p><a href="#_ftnref63">[63]</a> <em>The Life of the Mind</em>, vol. 1, <em>Thinking</em>, 4–5. The full passage at LM I, 5: “Could the activity of thinking as such, the habit of examining whatever happens to come to pass or to attract attention, regardless of results and specific content, could this activity be among the conditions that make men abstain from evil-doing or even actually ‘condition’ them against it?” Arendt’s tentative answer — that the cessation of thinking is implicated in the possibility of evil — has direct resonance for the present argument.</p><p><a href="#_ftnref64">[64]</a> Chow 2026, §5.4.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e375c2344bbf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[WHO WROTE AI-NATIVE,AND WHO DIDN’T]]></title>
            <link>https://medium.com/@mountdavistechnologies/who-wrote-ai-native-and-who-didnt-a5e47937e1c4?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/a5e47937e1c4</guid>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Wed, 20 May 2026 09:26:40 GMT</pubDate>
            <atom:updated>2026-05-20T09:26:40.451Z</atom:updated>
            <content:encoded><![CDATA[<p><em>An Independent Analysis of </em>ETSI White Paper №71 <em>“AI-native Infrastructure, next phase of Telco Cloud evolution”</em></p><p>Jimmy CHOW</p><h3>1 What WP-71 Is</h3><p>ETSI White Paper №71, titled <em>“AI-native Infrastructure, next phase of Telco Cloud evolution,”</em> was published in May 2026 — its 1st edition, 32 pages, 14 authors. It is published under the ETSI Industry Specification Group on Network Functions Virtualisation (ISG NFV), the body that has produced the NFV reference architecture since 2012.</p><p>The publication date is not incidental. <strong>WP-71 appeared three months after Deutsche Telekom’s OCAS architectural white paper (February 2026) and two months after Huawei’s TICC 2.0 (“Telco Intelligent Converged Cloud”) announcement at MWC Barcelona (March 2026).</strong> The timing positions WP-71 as the standards body’s response to a year in which both the practice-first camp (DT) and the vendor-marketing camp (Huawei) had published substantial public positions on where the Telco Cloud should go next.</p><p>WP-71 self-describes its purpose as follows:<a href="#_ftn1">[1]</a></p><blockquote>“This transition is inherently evolutionary rather than disruptive. Existing ETSI NFV architectures, specifications, and deployments remain foundational, while new AI-native capabilities are incrementally introduced through study items, reports, and future standardisation activities.”</blockquote><p><em>Evolutionary. Incremental. Study items. Future standardisation activities.</em> The language tells the reader what WP-71 is before the content does. <strong>This is a standards-track document that positions the AI-native transition as an extension of the existing NFV-MANO architecture — not as a replacement of it.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qYwLbFggxC9z7mR-FVn5dw.jpeg" /></figure><h3>2 Who Wrote It</h3><p>WP-71 has 14 authors. The composition is the single most revealing fact about the document.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/855/1*7ywmVgqk_ZBeJquvT8JZPg.png" /></figure><p><strong>Six of the fourteen authors — 43% — are from Huawei, ZTE, China Mobile, China Telecom, or China Unicom.</strong> Three are from Asian operators (DOCOMO, SKT, Rakuten). Two are from Turkish operators. Two are from small European firms (NEC Germany, Modsoft Poland). One — Leslie Willis of BT — is the sole representative of a major Western European operator.</p><p><em>The Chinese camp has authored nearly half the document. The practice-first camp — the operators and vendors that have actually deployed cloud-native at production scale — has authored none of it.</em></p><h3>3 Who Didn’t Write It</h3><p>The absences from the WP-71 author list are more analytically important than the presences.</p><p><strong>The following organisations have zero authors on WP-71:</strong></p><p>• <strong>Deutsche Telekom</strong> — the operator that published the OCAS whitepaper three months earlier, operates the world’s most documented cloud-native automation framework, and holds €119 billion of procurement leverage.</p><p>• <strong>Swisscom</strong> — the operator that built the SDC sidecar-based CNF lifecycle framework independently of DT.</p><p>• <strong>Orange</strong> — the Sylva TSC co-chairman, the operator driving the multi-operator cloud-native infrastructure project.</p><p>• <strong>Ericsson</strong> — the #2 global RAN vendor, DT’s primary CNF supplier, and a company whose 5G Core products must conform to OCAS.</p><p>• <strong>Nokia</strong> — the #3 global RAN vendor, with a $1 billion NVIDIA AI-RAN partnership announced six months before WP-71’s publication.</p><p>• <strong>Any hyperscaler</strong> — Microsoft, AWS, Google Cloud, NVIDIA. None.</p><p>• <strong>Any cloud-native foundation</strong> — CNCF, Linux Foundation Networking, Nephio. None.</p><p><strong>The entire practice-first cloud-native ecosystem — the operators that have built it, the vendors that supply it, the foundations that maintain the upstream projects it depends on — is collectively absent from the document that proposes to define the next phase of Telco Cloud evolution.</strong></p><p>This is not an oversight. <em>It is a political signal. The practice-first camp has concluded that the ETSI ISG NFV standardisation track is not where the architectural future is being decided — and has voted with its absence.</em></p><h3>4 What WP-71 Claims</h3><p>WP-71 proposes nine “capability themes” for the AI-native Telco Cloud. The themes span ultra-large-scale computing, AI-native cloud-network integration, cloud-edge-terminal AI agents, composable Telco Cloud services (IaaS → PaaS → MaaS), security and governance, model lifecycle management, edge-cloud distributed inference, real-time communication for AI, and multi-operator federation.</p><p>The themes are, on their own terms, technically reasonable. <strong>The problem is not what WP-71 proposes. The problem is where WP-71 claims the proposals come from.</strong></p><p>WP-71 frames the nine capability themes as the <em>“next phase of Telco Cloud evolution”</em> — an evolutionary extension of the existing ETSI NFV architecture. The document’s Section 4 lists four standardisation priorities, all framed as ETSI-led initiatives: <em>“Foundational standardisation for AI-native infrastructure,” “Standards enablement for autonomous operations,” “Unified standard governance for heterogeneous computing,” “Standards framework for data-centric visibility.”</em> <strong>All four priorities are about leading the development of standards. None is about production deployment. None is about open-source project launch. None is about operational verification.</strong></p><p>This is the standards-body self-perpetuation pattern: write a white paper → establish work items → produce Group Reports → produce Group Specifications → establish new work items → repeat. <em>The cycle generates standards. It does not, by itself, generate production-grade infrastructure.</em></p><h3>5 The IT Stack Paradox</h3><p>The most revealing page in WP-71 is page 11 — Figure 3.1–1, titled <strong>“AI-native Technology Stack from IT Industry.”</strong></p><p>The figure displays four layers of the AI-native technology stack. Every component in every layer is from the IT industry open-source or Big Tech ecosystem:</p><p>• <strong>Infrastructure Layer:</strong> Kubernetes (CNCF), Kubeflow (CNCF), Daytona, Firecracker (AWS), E2B.</p><p>• <strong>Data &amp; Model Layer:</strong> PyTorch (PyTorch Foundation), Ray, vLLM, DeepSpeed (Microsoft), Hugging Face, Apache Kafka, LF AI &amp; Data, Milvus, ONNX.</p><p>• <strong>Agent Collaboration Layer:</strong> Agentic AI Foundation, Model Context Protocol (Anthropic), A2A (Google), goose, ANP.</p><p>• <strong>Application Layer:</strong> A2UI, Dify, plus a placeholder for “AI-Native NF” — the only ETSI-specific concept slot.</p><p><strong>ETSI did not create any of these components.</strong> Not one. The entire AI-native technology stack that WP-71 presents as the future of Telco Cloud is built from components created by CNCF, the Linux Foundation, PyTorch Foundation, Apache Foundation, Google, Microsoft, AWS, and Anthropic.</p><p>On page 20, WP-71 explicitly acknowledges its dependence on CNCF’s Dynamic Resource Allocation (DRA) framework — which reached General Availability within the CNCF ecosystem — as <em>“useful building blocks”</em> for AI-aware orchestration. <strong>ETSI’s role, by its own admission, is to “integrate” — not to invent.</strong></p><p><em>This is the IT Stack Paradox at the heart of WP-71: the document proposes to standardise the “next phase of Telco Cloud evolution” using components that were entirely created outside ETSI, outside the telco standards ecosystem, and outside the author organisations that wrote the document. The practice-first camp (DT, Swisscom, Orange) has already adopted these same components directly from their upstream sources — Kubernetes, GitOps, Nephio, FluxCD — without waiting for ETSI to standardise them. WP-71 is, in effect, a proposal to standardise what the practice-first camp has already deployed.</em></p><h3>6 The Two Tracks Comparison</h3><p>The sharpest way to read WP-71 is against the DT OCAS white paper published three months earlier. The comparison is structurally exact because both documents address the same question — <em>“what should the next phase of Telco Cloud look like?”</em> — and arrive at structurally opposite answers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/880/1*IoJP3vbTN_GaA-DPOJpIwA.png" /></figure><p><strong>The comparison speaks for itself.</strong> WP-71 is a 32-page intent declaration with zero production references. DT’s OCAS white paper is a 26-page production case with eight deployed business domains, forty workloads, and two hundred thousand vCPUs. <em>The two documents are not in the same category. One describes what should exist. The other describes what does exist.</em></p><h3>7 What WP-71 Reveals About ETSI’s Structural Position</h3><p>WP-71 is not a bad document. Its nine capability themes are technically sound. Its acknowledgement of IT industry components is honest. Its Figure 3.1–1 is, in its own way, a remarkable act of institutional candour — ETSI publicly drawing a technology stack in which ETSI contributes none of the components.</p><p><strong>What WP-71 reveals is that ETSI ISG NFV, in 2026, occupies a structural position that is increasingly disconnected from where the engineering is actually happening.</strong></p><p>The engineering is happening in four places: (1) inside Track B operators like DT, Swisscom, and DOCOMO, who build their own automation; (2) inside CNCF and Linux Foundation projects like Kubernetes, Nephio, and Sylva; (3) inside hyperscalers like Microsoft, AWS, and Google Cloud, who are building carrier-grade substrates; and (4) inside Chinese vendor-operator co-design relationships (Huawei + China Mobile, ZTE + China Telecom) that produce integrated solutions outside ETSI’s governance.</p><p><strong>ETSI ISG NFV’s role in this landscape is to standardise after the fact — to take engineering that has already been created elsewhere and codify it into specifications.</strong> That is a legitimate role for a standards body. It is not, however, the role WP-71 claims when it proposes to <em>“lead the development”</em> of the AI-native infrastructure. <em>ETSI is not leading. ETSI is following — and the organisations that are leading have, with the exception of the Chinese camp and BT’s Leslie Willis, declined to participate in the document that proposes to standardise their work.</em></p><p>The Huawei–ETSI timing alignment reinforces this reading. Huawei announced TICC 2.0 at MWC Barcelona in March 2026. WP-71 — with two Huawei authors, and with capability themes that map closely to TICC 2.0’s marketing framework — appeared two months later. <strong>The direction of influence is from vendor marketing to standards white paper, not the other way around.</strong> ETSI ISG NFV is, on the AI-native agenda, operating as the standardisation execution arm of the Chinese vendor-operator camp.</p><h3>8 Conclusion</h3><p>ETSI White Paper №71 is a document that deserves to be read carefully — not for what it proposes, which is technically sound, but for what it reveals about the structural politics of the Telco Cloud transition in 2026.</p><p>It reveals that the Chinese camp (Huawei, ZTE, China Mobile, China Telecom, China Unicom) has effective editorial control of the ETSI ISG NFV AI-native agenda. It reveals that the practice-first camp (DT, Swisscom, Orange, Ericsson, Nokia) has collectively withdrawn from that agenda. It reveals that the IT industry components WP-71 proposes to standardise are already in production at operators that did not wait for ETSI to standardise them. And it reveals that ETSI’s proposed next steps — four standardisation priorities, zero production deployment targets — are addressed to a standards-making audience, not to an engineering-execution audience.</p><p><strong>The two tracks to AI-native Telco Cloud are now visible in the same month’s publications:</strong> WP-71 proposes to standardise the future. DT’s OCAS white paper documents the present. <em>The question for the broader operator industry is not which track is “right” — both are legitimate within their structural contexts. The question is which track produces production-grade infrastructure faster, and the 2026 evidence on that question is not ambiguous.</em></p><p><strong><em>The practice-first camp has shipped. The standards-first camp has published a white paper. The reader may draw the appropriate conclusion.</em></strong></p><p><a href="#_ftnref1">[1]</a>ETSI White Paper №71, <em>‘AI-native Infrastructure, next phase of Telco Cloud evolution,’</em> 1st edition, May 2026. Available at <a href="https://www.etsi.org/wp-content/uploads/ETSI-WP-71_AI-native_Infrastructure_next_phase_of_Telco_Cloud_evolution.pdf">https://www.etsi.org/wp-content/uploads/ETSI-WP-71_AI-native_Infrastructure_next_phase_of_Telco_Cloud_evolution.pdf</a></p><p><em>Jimmy CHOW ·20 May 2026</em></p><blockquote><strong><em>Professional Bio / Collaboration Offer</em></strong></blockquote><blockquote><strong><em>Strategic Network Transformation &amp; Integration Expert.</em></strong><em> If your organisation is navigating </em><strong><em>multi-generation network integration</em></strong><em> (2G / 3G / 4G / 5G / 6G) en route to a future-proof </em><strong><em>cloud-native core network</em></strong><em>, or managing the intricate architectures and fragmented service platforms that arise from </em><strong><em>mergers and acquisitions (M&amp;A)</em></strong><em> aimed at an </em><strong><em>Autonomous Network Architecture</em></strong><em> — get in touch. I provide end-to-end support: situational analysis, strategic roadmap, hands-on implementation and smooth transition. I work either independently or embedded in your existing team.</em></blockquote><blockquote><strong><em>Email:</em></strong><em> mountdavistechnologies@gmail.com</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a5e47937e1c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[After Eight Years and €119 bn, Deutsche Telekom Has Proven Swisscom Right — Just Louder, and With…]]></title>
            <link>https://medium.com/@mountdavistechnologies/after-eight-years-and-115-bn-deutsche-telekom-has-proven-swisscom-right-just-louder-and-with-7897772a1c0b?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/7897772a1c0b</guid>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[gitops]]></category>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Thu, 14 May 2026 20:33:06 GMT</pubDate>
            <atom:updated>2026-05-15T08:04:03.438Z</atom:updated>
            <content:encoded><![CDATA[<h3>After Eight Years and €119 bn, Deutsche Telekom Has Proven Swisscom Right — Just Louder, and With Better Marketing</h3><p><em>— Using €119 bn/year of Buying Power to Rewrite Operator Pain Points into Vendor KPIs</em></p><p>Jimmy CHOW</p><p>After finishing that series of articles on Swisscom (<a href="https://medium.com/@mountdavistechnologies/another-deep-analysis-swisscoms-end-to-end-cloud-native-telco-platoform-automation-8c1cf850817d">3/31</a>, <a href="https://medium.com/@mountdavistechnologies/a-detailed-breakdown-of-swisscoms-cloud-native-automation-platform-what-it-does-and-what-it-does-e51c8f13c53f">4/1</a>, <a href="https://medium.com/@mountdavistechnologies/sdc-the-most-critical-and-most-complex-component-in-the-entire-architecture-170cf7a94435">4/2</a>, <a href="https://medium.com/@mountdavistechnologies/a-blunt-conclusion-about-the-cloud-native-autonomous-network-management-platform-0854111c905d">4/8</a>), I had genuinely intended to move on — to sovereign cloud, to 6G, to political philosophy, to anything other than this dreary subject. But Analysys Mason’s February 2026 report on Deutsche Telekom’s HTC, together with DT’s companion <a href="https://www.telekom.com/resource/blob/1102562/4751becd7d636c4824ca874e94057383/dl-260223-whitepaper-cloud-automation-data.pdf">OCA whitepaper</a> published the same week (Banev / Derckx / Werding et al., nine authors), made it impossible to leave well alone, when LightReading’s industry Newsline pushed this message to me .</p><p><strong><em>Because reading it through, I arrived at a conclusion that is slightly uncomfortable but has to be said: DT appears to have walked considerably further down the road than Swisscom, but in essence the two operators are travelling the same path. The only difference is that DT, leveraging €115 billion / year of procurement firepower, has rewritten the very pain points that Swisscom’s engineers wrestle with daily inside SDC into hard contractual clauses imposed on its vendors.</em></strong></p><p>As for the lower half of the bridge — the traditional telecoms world — <strong>neither operator has touched it.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*g-C0YL9RsjLcR2DiLmkOHw.jpeg" /></figure><h3>I. What OCAS Actually Is — in Five Minutes</h3><p>DT calls the whole stack the <strong>Horizontal TelCo Cloud (HTC)</strong>, nicknamed internally the “burger architecture.” Bare metal at the bottom (BM4X, ten-plus data centres, engineered right down to rack profile, cabling, power and cooling), Kubernetes on top of it (T-CaaS, whose open-source form is called <em>Das SCHIFF</em>, built on vanilla Kubernetes), then a layer of multi-vendor CNFs, and finally Automation Core at the top for automation and observability. <strong>OCAS</strong> is the principal subsystem within Automation Core.</p><p><strong>OCA (methodology) vs. OCAS (the product)</strong> — these need to be distinguished:</p><p>• <strong>OCA</strong> = Operator-based Cloud Automation, an architectural methodology built on the Kubernetes Operator Pattern.</p><p>• <strong>OCAS</strong> = OCA Solution, DT’s concrete, production-grade product realisation of OCA.</p><p>The whitepaper lists the network functions OCAS has already taken into production: <strong>TAS, IN, SMS GW, MRF, DNS, SBC, HSS, NDL</strong>. <em>Note that UPF, SMF and AMF do not appear on this list — a detail to which we shall return shortly.</em></p><p>The OCAS architecture is defined by six design principles (DP1–DP6), translated into human being’s plain language:</p><p>• <strong>DP1 Declarative lifecycle management</strong> — every operational instruction stored in Git, version-controlled, expressed as desired state.</p><p>• <strong>DP2 Kubernetes-native automation</strong> — all orchestration happens inside the Kubernetes control plane, with no reliance on external orchestrators. This is, in effect, a quiet death sentence on ETSI NFV-MANO within OCA scope.</p><p>• <strong>DP3 Managed infrastructure provisioning</strong> — OCA does not manage the cluster itself, but it does manage namespaces, network policies, storage classes and shared services within the cluster.</p><p>• <strong>DP4 Strict separation of software, configuration and data</strong> — three separate Git repositories, independently referenced.</p><p>• <strong>DP5 Modular and hierarchical architecture</strong> — simple cases use Helm; only the genuinely complex ones go through the OCA layer.</p><p>• <strong>DP6 Component ownership principle</strong> — whoever holds the domain expertise carries the responsibility. Vendors handle CNF internal logic through their own Kubernetes operators; DT handles operational procedures and environment integration through OCA’s own generic operators.</p><p>On the technology stack: OCAS uses <strong>FluxCD + GitLab + Artifactory</strong> for GitOps; a constellation of operators for orchestration — the <strong>Service Operator (OSO)</strong> as the brain, <strong>CNF LCM Operators</strong> delivered by vendors, <strong>Integration Operators</strong> (Test Automation, Inventory, Ansible) built in-house by DT; and a <strong>Config Builder + Config Validator</strong> (Go templating + Yang / JSON schema) for configuration assembly and validation.</p><p>The most consequential contractual requirement is this one: <strong>every CNF vendor must deliver an EMS-less, CRD-based CNF LCM Operator.</strong> The state model has exactly five states (Active / Created / Failed / Terminated / Inactive), with mandatory support for ISSU and active periodic health testing. <strong>Cannot meet the bar? Then the vendor goes down the CNF Helm Operator transition path — and the whitepaper explicitly states this is a temporary bridge with the long-term intent of eliminating it.</strong></p><p>Business impact data: compared with traditional centralised orchestrators, OCAS reduces single-NF automation solution cost by <strong>30%</strong>, delivery time by <strong>73%</strong>, and the five-year ownership cost of the automation suite itself by <strong>67%</strong>. The internal slogan for this is “<strong>Brutally Efficient Automation</strong>.”</p><p><strong><em>Those are the entirety of OCAS’s selling points. Done with that? Right — now let us look at the same thing through Jimmy’s different lens.</em></strong></p><h3>II. Taking the Same Knife That Cut Swisscom, and Cutting OCAS With It</h3><p>In my <a href="https://medium.com/@mountdavistechnologies/sdc-the-most-critical-and-most-complex-component-in-the-entire-architecture-170cf7a94435">SDC piece</a> I wrote this:</p><blockquote>“Convergence is not a final solution; choosing one is. Either the operators or the vendors — somebody has to make the choice.”</blockquote><p>DT chose “operators get the final say” — using €115 bn / year of contractual pressure to force vendors into CRD-based delivery. Swisscom chose “collaborate and accommodate” — partnering with Nokia to refine SDC into a more polished NETCONF↔K8s translator.</p><p><strong>At first glance DT looks the smarter party, because it has shifted the cost onto its suppliers. But fundamentally?</strong></p><h3>2.1 OCAS’s CNF LCM Operator = the Sidecar Controller I Proposed on 4/2</h3><p>In my SDC piece I drew this diagram:</p><pre>CNF Pod<br>  ├── main process (old code unchanged)<br>  │     └── internal YANG/NETCONF data model preserved<br>  └── Sidecar Controller (NEW: modest coding from the vendor)<br>        ├── watches the K8s CRD for changes<br>        ├── translates changes into internal NETCONF calls<br>        └── reports state back to the K8s API<br> <br>External: Kubernetes CRD API (operators consume directly, no SDC required)<br>Internal: YANG/NETCONF stays untouched</pre><p>At the time I estimated this would take Nokia / Ericsson three to five months, and Huawei perhaps less than two. The conclusion was: “<strong>This is the truly positive and genuinely cooperative way to ease operators’ pain points.</strong>”</p><p>It was only on reading DT’s OCA whitepaper that the penny dropped — <strong>DT has effectively already industrialised exactly this proposal. It just calls it “CNF LCM Operator” rather than “sidecar.”</strong> Look at the <em>Collaboration and expectations from CNF suppliers</em> section on page 16 of the whitepaper. What DT contractually compels Amdocs, Mavenir, Nokia, Ribbon and Titan.ium to do is, <strong>almost line-for-line, the diagram I sketched:</strong></p><blockquote>“Vendors handle CNF internal logic through their own CNF-specific Kubernetes operators. DT manages operational procedures and infrastructure integration through in-house-developed generic OCA operators.”</blockquote><p><em>When I wrote that piece, I was speaking from the perspective of a Swisscom engineer — “if only you could compel the vendors to do this.” DT has compelled them. How? </em><strong><em>By substituting scale for standardisation.</em></strong></p><h3>2.2 But — OCAS Still Has Not Touched the Hard Bones I Listed</h3><p>Returning to the <em>“What It Does Not”</em> inventory I compiled for Swisscom on <a href="https://medium.com/@mountdavistechnologies/a-detailed-breakdown-of-swisscoms-cloud-native-automation-platform-what-it-does-and-what-it-does-e51c8f13c53f">4/1</a>, let us apply each item, point for point, to DT:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*rj61AxArabMJtEoSnIvYYA.png" /></figure><p>The single sharpest sentence I wrote about Swisscom on <a href="https://medium.com/@mountdavistechnologies/another-deep-analysis-swisscoms-end-to-end-cloud-native-telco-platoform-automation-8c1cf850817d">3/31</a>:</p><blockquote>“It is a highly engineered, Kubernetes-based Configuration Management and Delivery Automation System for the 5G Core.”</blockquote><p><strong>Substitute OCAS for the subject. The sentence holds verbatim.</strong> DT has industrialised this across eight business domains, produced hard numbers of 30% / 73% / 67%, scaled to 200,000 vCPUs — but <strong>it has not crossed the same river.</strong></p><h3>2.3 The Real “Gap” Lies in Three Non-Technical Dimensions</h3><p>Rewriting the comparison strictly through my own <strong><em>What It Does / What It Does Not</em> lens</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/771/1*Vmj6GWJVg1QZ2c5VLdxnwQ.png" /></figure><p>The last row is particularly worth dwelling on. In the <a href="https://medium.com/@mountdavistechnologies/sdc-the-most-critical-and-most-complex-component-in-the-entire-architecture-170cf7a94435">SDC piece</a>, I specifically highlighted Nokia’s deep contributions to SDC as the talk’s <strong><em>“most significant signal”</em></strong> — at least one major CNF vendor was beginning to take Kubernetes-native configuration management seriously.</p><p>But <strong>the same Nokia also appears in DT’s OCAS whitepaper</strong> (Marcelo Madruga’s testimonial, page 17). What does that tell us?</p><blockquote>Vendors are running both legs at once: one in upstream open-source standardisation (politically correct “we support cloud-native”), the other in DT’s private contracts (the actual business). They do both, but they commit fully to neither. This is, frankly, a very familiar pattern of telecom vendor behaviour — <strong>I saw exactly the same playbook over ten years ago, only it was called SDN / NFV back then, and today it is called cloud-native.</strong></blockquote><h3>III. The State of the Industry: Putting All the Strands Together</h3><p>By this point, <strong>let me gather the various judgements scattered across earlier pieces into a single picture</strong>. None of these judgements is new — they recur throughout my earlier writing. What the DT case provides is simply a high-profile reference point against which to restate them in a more systematic way.</p><h3>3.1 The Tooling-Layer / Product-Layer Distinction — My Recurring Thesis</h3><p>In <a href="https://medium.com/@mountdavistechnologies/a-blunt-conclusion-about-the-cloud-native-autonomous-network-management-platform-0854111c905d">A Blunt Conclusion (4/8)</a> I wrote:</p><blockquote><strong>“The tooling-layer problem, [Swisscom] has solved with Operators. The product-layer problem, no single operator can solve unilaterally. It requires 3GPP, CNF vendors, and the entire industry supply chain to move together.”</strong></blockquote><p>DT is at once the <strong>strongest counter-example and the strongest reinforcement</strong> of that judgement:</p><p><strong>• Counter-example:</strong> DT demonstrates that an operator of sufficient scale can, unilaterally, force a portion of product-layer redesign through procurement contracts. CNFs must be EMS-less, must implement a five-state operator, must support ISSU. These are things 3GPP has been unable to deliver in over a decade.</p><p><strong>• Reinforcement:</strong> even with DT’s full weight applied, <strong>the core architectural problems — UPF data plane, AMF stateful scaling, on-demand slicing — remain unsolved.</strong> Because these problems are not interface-specification problems. They are debts laid down on day one of the 5G protocol design.</p><p>In <a href="https://medium.com/@mountdavistechnologies/telecom-cloud-native-or-cloud-na%C3%AFve-7c7ae2dc6a1d">Telecom Cloud-Native or Cloud-Naïve (4/16)</a> I made a deliberately sharp distinction — the industry routinely conflates <em>“cloud-native”</em> with <em>“cloud-naïve.”</em> <strong>OCAS and SDC are both highly competent pieces of cloud-native engineering, but the 5G CNF kernels they wrap are still cloud-naïve.</strong> These two statements can hold simultaneously.</p><h3>3.2 The “Bridge vs Destination” Paradox — DT Is More Reluctant Than Swisscom to Admit It</h3><p>In the <a href="https://medium.com/@mountdavistechnologies/another-deep-analysis-swisscoms-end-to-end-cloud-native-telco-platoform-automation-8c1cf850817d">3/31 piece</a> I gave Swisscom’s platform its final characterisation: <strong>“A Bridge, Not a Destination.”</strong> A transitional engineering bridge built between traditional telecoms operations and cloud-native operations.</p><p>Swisscom itself, at KubeCon EU 2026, relabelled its talk <em>“Why is it So Hard to Run a 5G Core on Kubernetes.”</em> That is a genuinely honest return to first principles.</p><p><strong>DT’s whitepaper is considerably more “marketing-coded.”</strong> But buried inside it sits <strong>ONE</strong> important admission:</p><blockquote>“OCA complements Nephio by providing immediate automation capabilities for non-standardized vendor implementations while the industry converges on common interfaces. As vendors adopt Nephio standards, OCA operators can gradually transition to standardized interfaces, providing a pragmatic bridge to full industry standardization.”</blockquote><p><strong>In plain human being English: OCA is also a bridge. Not a destination.</strong> DT itself concedes that it is waiting on Nephio to mature, waiting on the industry to converge.</p><p><strong><em>So both operators have built bridges. The only difference is this: Swisscom’s bridge has been hand-laid by engineers inside SDC. DT’s bridge is reinforced with the structural steel of a €115 billion-per-year contract book. The structural soundness of the two bridges differs — but the far bank they lead to is the same one.</em></strong></p><h3>3.3 The Vendor Two-Legged Strategy — A Structural Problem</h3><p>In the <a href="https://medium.com/@mountdavistechnologies/sdc-the-most-critical-and-most-complex-component-in-the-entire-architecture-170cf7a94435">SDC piece</a> I queried Nokia’s sincerity around SDC. The DT case completes that suspicion:</p><p>• Nokia stands behind both SDC (the public standard) <strong>and</strong> OCAS (DT’s private interface).</p><p>• Mavenir delivers EMS-less packet core to DT — does it hold the same line for smaller operators? <em>No public data.</em></p><p>• Amdocs supplies DT with full CRD-based PCF / PCRF / BSF — but Openet (which Amdocs acquired) already had the capability; this is not new development.</p><p><strong>This actually corroborates a point I have made repeatedly: since day one of 5G, operators and vendors have failed to draw a clean responsibility boundary inside 3GPP.</strong> When I left this domain ten-odd years ago, operator influence over standardisation was already in visible decline. Today, both Swisscom and DT are doing self-rescue on top of K8s + GitOps. <strong>Both are using operational-layer engineering to paper over standards-layer design debt.</strong></p><p>I shall repeat verbatim what I wrote on <a href="https://medium.com/@mountdavistechnologies/another-deep-analysis-swisscoms-end-to-end-cloud-native-telco-platoform-automation-8c1cf850817d">3/31</a>:</p><blockquote>“Why is KubeCon EU 2026 still hosting talks on problems that should have been solved more than a decade ago? Why has no operator stood up in 3GPP standardisation work to call this out?”</blockquote><h3>3.4 Will 6G Genuinely Get Redesigned? A Question Worth Asking Again</h3><p>There is one sentence in DT’s whitepaper that, frankly, made me laugh out loud:</p><blockquote>“DT expects that 6G will primarily be a software upgrade, and its HTC project has already established both the technical and operational foundations for this transition.”</blockquote><p><strong>6G will primarily be “a software upgrade”?</strong></p><p>In <a href="https://medium.com/@mountdavistechnologies/6g-core-can-we-build-an-autonomous-network-without-a-nervous-breakdown-bcf698e59de4">6G Core: Can We Build an Autonomous Network Without a Nervous Breakdown? (4/28)</a> I have already, in advance, addressed this kind of claim. <strong>If 5G’s product-architecture debts (UPF data plane, stateful NF elasticity, dynamic slicing) remain unsolved, on what basis does 6G become “mostly a software upgrade”?</strong></p><p>If by 2030 — when the 6G standard is set — 3GPP is still operating as it does now, with CNF vendors designing interfaces in the day-one YANG / NETCONF paradigm, then DT’s HTC will simply be joined by a second-generation bridge. <strong>The only difference will be that, by then, the structural steel will be reinforced with AI agents. But the location of the bridge will not move. And the river it is bridging will not have changed.</strong></p><p>Which brings me back to the broader themes of my pieces on <a href="https://medium.com/@mountdavistechnologies/sovereign-ai-or-resilient-service-443d8bb62a72">Sovereign AI (4/30)</a> and <a href="https://medium.com/@mountdavistechnologies/sovereign-cloud-ai-sovereignty-on-the-left-the-world-on-the-right-0e7dc69b431d">Sovereign Cloud / AI (4/29)</a> — <strong>the telecom industry’s real problem is not a technology problem. It is a structural problem of the industry itself.</strong> You can bolt on labels such as “sovereign,” “autonomous network,” “6G,” and the underlying issue remains identical: the standards-setting power inside 3GPP, the interface ownership held by CNF vendors, and the awkward position of the operators caught between the two.</p><h3>IV. Closing: What DT Is Actually Signalling to the Industry</h3><p>If we treat DT’s HTC / OCAS as the most fully realised “model answer” of European telecom cloud-native transformation in 2026, the signal it sends carries three soul-searching questions.</p><p><strong>To operators:</strong> <em>Do you want to be Swisscom — the engineer’s honest, technically respectable position — or DT — the industrialised buyer’s position, more economically efficient? The former earns more technical respect. The latter delivers harder commercial returns. But if you are not a Tier 1, DT’s path is not open to you. </em><strong><em>You do not have the procurement leverage</em></strong><em>.</em></p><p><strong>To vendors:</strong> <em>How much longer is the two-legged play sustainable? Once Nephio matures, once a second and third operator the size of DT writes CRD interfaces into their contracts — your NETCONF / YANG asset base flips from asset to liability.</em> The sidecar approach I proposed in the <a href="https://medium.com/@mountdavistechnologies/sdc-the-most-critical-and-most-complex-component-in-the-entire-architecture-170cf7a94435">4/2 piece</a> is what DT has now made a <strong>de facto</strong> standard. <strong>You can do it three months early — or be forced into it three years late.</strong></p><p><strong>To myself, and to others observing this industry from the periphery:</strong> <em>this will be slow, interestingly boring.</em></p><p>DT, counting from the NIMS project in 2018 through to the public HTC / OCA whitepapers in February 2026, took <strong>eight full years.</strong> Mobilising along the way: a nine-author core team, forty-plus vendor partners, a 200,000-vCPU production estate, deep engagement with two European open-source ecosystems (Sylva and Nephio). All of this to push the <strong>tooling layer</strong> to the industry ceiling.</p><p>And the <strong>product-layer</strong> problems — truly cloud-native UPF, stateful NF elasticity, minute-scale dynamic slicing — <strong>should have been solved ten years ago, and to this day no one is seriously discussing them inside 3GPP plenary sessions.</strong></p><p>So I shall transplant the closing paragraph of my <a href="https://medium.com/@mountdavistechnologies/a-detailed-breakdown-of-swisscoms-cloud-native-automation-platform-what-it-does-and-what-it-does-e51c8f13c53f">4/1 piece</a> almost verbatim, swapping the subject to “DT and Swisscom together”:</p><p><em>What DT and Swisscom have collectively built is a highly engineered, Kubernetes-based configuration management and delivery automation system for telecom networks.</em></p><p><em>Its value is real — reduction of human error, improvement of configuration consistency, establishment of change audit trails, substantial reduction of engineer cognitive burden. Every one of these is a tangible production-grade improvement.</em></p><p><strong><em>But between this platform and the true goal of 5G / 6G Core cloud-native transformation — elasticity, self-healing, on-demand slicing, stateless horizontal scaling — there remains a fundamental and unbridged chasm. That chasm does not lie at the operations tooling layer. It lies at the product architecture layer of the CNFs themselves.</em></strong></p><p><em>The tooling-layer problem, the operators have solved with Operators. The product-layer problem, </em><strong><em>no single operator can solve unilaterally</em></strong><em>. It requires 3GPP, CNF vendors, and the entire industry supply chain to move together — and the fact that this is still being debated on the KubeCon EU 2026 stage tells you everything you need to know.</em></p><p><strong>DT is not an advanced version of Swisscom. They are two bridges over the same river.</strong> <strong>DT’s bridge is just longer, wider, and built with more expensive steel — but the far bank remains the same.</strong></p><p><em>(If this paper reminded you of a 3GPP meeting from ten years ago where some major vendor poured a bucket of cold water over your head, that was not a coincidence. It is the same script. Only the actors have changed.)</em></p><p><em>Jimmy CHOW · Mount Davis Technologies · 14 May 2026</em></p><blockquote><strong>Professional Bio / Collaboration Offer</strong></blockquote><blockquote><strong>Strategic Network Transformation &amp; Integration Expert.</strong> If your organisation is navigating <strong>multi-generation network integration</strong> (2G / 3G / 4G / 5G / 6G) en route to a future-proof <strong>cloud-native core network</strong>, or managing the intricate architectures and fragmented service platforms that arise from <strong>mergers and acquisitions (M&amp;A)</strong> aimed at an <strong>Autonomous Network Architecture</strong> — get in touch. I provide end-to-end support: situational analysis, strategic roadmap, hands-on implementation and smooth transition. I work either independently or embedded in your existing team.</blockquote><blockquote><strong>Email:</strong> mountdavistechnologies@gmail.com</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7897772a1c0b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sovereign AI or “Resilient Service”: A Shrewd Marketing Pivot or a True Strategic Shield?]]></title>
            <link>https://medium.com/@mountdavistechnologies/sovereign-ai-or-resilient-service-443d8bb62a72?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/443d8bb62a72</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[sovereignty]]></category>
            <category><![CDATA[resilient]]></category>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Thu, 30 Apr 2026 19:19:36 GMT</pubDate>
            <atom:updated>2026-05-01T09:26:50.785Z</atom:updated>
            <content:encoded><![CDATA[<p>Jimmy CHOW</p><p>Following the release of news regarding BT, I published my own analysis and perspectives in two articles <em>(see link below) </em>. While I was conceiving my second one, I noticed that BT management and regular employees alike were publishing and sharing their views on LinkedIn, uniformly repackaging <strong>Sovereign Cloud/AI Service</strong> as <strong><em>Resilient Service</em></strong><em>.</em> They have added some narratives, but the messages are all largely the same. Upon looking into it, this appears to be a unified marketing campaign, which may have originally originated from <strong>BT Business CEO Jon James</strong> <em>(I cannot confirm who first introduced this concept; however, any errors here would not affect the analysis in this article).</em></p><blockquote><strong>The 1st one</strong>:<a href="https://medium.com/@mountdavistechnologies/bt-sovereign-ai-the-rise-of-uk-sovereign-cloud-and-ai-can-it-carve-a-new-path-amid-hyperscaler-720d109e6ecf">BT Sovereign AI: The Rise of UK Sovereign Cloud and AI — Can It Carve a New Path Amid Hyperscaler Dominance?</a></blockquote><blockquote><strong>The 2nd one</strong>: <a href="https://medium.com/@mountdavistechnologies/sovereign-cloud-ai-sovereignty-on-the-left-the-world-on-the-right-0e7dc69b431d">Sovereign Cloud/AI: Sovereignty on the Left, the World on the Right</a>.</blockquote><p>I agree with this shift in narrative discourse. I believe I have sufficient successful practical experience in product marketing; twenty years ago, before the concept of “<strong>Big Data</strong>” had even hit the market, I successfully packaged the concept of mobile network user data fusion/opening. It was very successful because being able to express the same thing in a different way brings about very different market results. Therefore, based on my own experience, I will analyze this concept as a marketing tactic.</p><p>Compared to the previous concept, this shift elevates “<strong>Sovereignty</strong>” from the traditional dimension of <strong>Compliance</strong> to the level of <em>“</em><strong><em>survivability</em></strong><em>”</em> and <em>“</em><strong><em>business continuity</em></strong><em>”</em> for nations and enterprises. <em>Legal “compliance” requirements are a very cold and coercive concept; using them as a marketing angle is not particularly effective.</em> Even if successful, it might only be due to government compliance mandates, thereby monopolizing customers supported by government finances, such as the NHS.</p><blockquote>Latest UK news shows how important of “data sovereignty”, “data security”, etc., when more than 500 thousands volunteers’ personal data, including data of DNA, bio-protein level, etc. can be found and sold on Chinese TaoBao platform, that’s really a serious accident to think about data sovereignty/security.</blockquote><blockquote><a href="https://www.ukbiobank.ac.uk/news/a-message-to-our-participants-uk-biobank-data-security-update/"><em>UK Biobank Data Leakage</em></a></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7c2Xo0GV9Qc6QG-ljsU-Ng.jpeg" /></figure><h3>Analyzing “Resilient Service” from a Marketing Perspective</h3><h4>1. The Core Logic of BT Management’s Promotion: From “Compliance” to “Resilience”</h4><p>In BT’s recent public news, etc., the core narrative revolves around <em>“</em><strong><em>maintaining control in a world full of variables</em></strong><em>”.</em></p><ul><li><strong>Jon James (BT Business CEO) View</strong>: He stated clearly that digital sovereignty is not just about where data is stored, but about enabling UK institutions to adopt world-leading technologies while maintaining “<strong>Resilience</strong>” and “<strong>Control</strong>” in a rapidly changing world.（<a href="https://newsroom.bt.com/bt-to-work-with-nscale-to-accelerate-uk-sovereign-ai-capability-powered-by-nvidia/"><em>source link</em></a>）</li><li><strong>Core Selling Point</strong>: BT no longer talks only about virtual machines and computing power; instead, it talks about the full-stack capabilities of <em>“</em><strong><em>Network Sovereignty, Voice Sovereignty, Cloud Sovereignty, and AI Sovereignty</em></strong><em>”.</em> The subtext is: should an extreme geopolitical situation arise, BT is the only “<strong>digital backbone</strong>” that can ensure the UK’s communications, data, and computing power are not instantaneously cut off by external forces (such as the <strong>US CLOUD Act</strong> or supply chain disruptions).</li></ul><h4>2. Rationality Analysis: Positioning as “National Insurance”</h4><p>Defining sovereign cloud as a “<strong>resilient service</strong>” is highly rational in a logical sense because it avoids direct confrontation with Hyperscalers <em>(such as AWS and Azure) </em>on price and general features.</p><ul><li><strong>Avoiding Price Wars</strong>: If sovereign cloud is a “product,” users will compare the cost per core; if it is a <em>“resilient service,”</em> users will view it as a form of <em>“</em><strong><em>Insurance</em></strong><em>”.</em> For the NHS or the Ministry of Defence (MoD), a 15–30% premium in exchange for a <em>“guarantee against being cut off”</em> is a reasonable premium, if this is guaranteed, there’ll no <strong>UK Biobank </strong>data leakage happened.</li><li><strong>Binding with CNI </strong><em>(Critical National Infrastructure)</em>: BT leverages its natural identity as a telecommunications operator to bind cloud services with physical connectivity. This <em>“</em><strong><em>Full-stack Sovereignty</em></strong><em>” </em>creates a <strong>physical and logical safety loop</strong>, which is extremely persuasive in the current context of global decoupling.</li><li><strong>The “Peace of Mind” Card on a Psychological Level</strong>: By leveraging the rigid demand of the public sector for “<strong>zero foreign risk</strong>,” sovereignty is defined as a “<strong>final safe haven</strong>” amidst extreme geopolitical maneuvering. This ensures that when customers purchase, they are not buying technical advancement, but rather the “<strong>ownership of control</strong>”.</li></ul><h4>3. Practicality Analysis: Grand Narratives and Implementation Bottlenecks</h4><p>Despite the logical soundness, at the <strong>practical level</strong>, positioning this as a “<strong>Resilient Service</strong>” still faces the previously identified structural challenges:</p><h4>The Risks of “Pseudo-Sovereignty” (Supply Chain Dependency):</h4><ul><li><strong>Current Status</strong>: BT’s underlying technology remains highly dependent on partners (<em>such as NVIDIA’s computing power, Rackspace’s hosting, and Nscale’s architecture</em>).</li><li><strong>Practical Challenges</strong>: If true “<strong>resilience</strong>” implies being free from external constraints, does BT’s local team have the capability to achieve “<strong>technological autonomy</strong>” when NVIDIA’s software updates are restricted or underlying drivers suffer from security vulnerabilities? If it is merely “<strong>British people operating American technology</strong>,” such <strong>resilience remains fragile </strong>in the face of extreme geopolitical conflict. Thinking about F-35 jet fighter as example, it said without permission of USA authority, it can not fly at all.<em> (note: news from internet, no fact check, just show one possibility in this world)</em></li></ul><h4>The Paradox of Scale and Cost:</h4><ul><li><strong>Current Status</strong>: BT’s currently deployed <strong>14MW</strong> computing scale is minuscule compared to those of Hyperscalers.</li><li><strong>Practical Challenges</strong>: True “<strong>resilience</strong>” requires <strong><em>redundancy</em></strong><em>(geography &amp; capacity)</em>. If the computing scale is insufficient to support large-scale surges in demand, or if costs become prohibitively high due to fragmentation, users (especially SMEs) may abandon this expensive “<strong>sovereign resilience</strong>” for the sake of survival.</li></ul><h4>The Disconnection Between “Service” and “Ecosystem”:</h4><ul><li><strong>Current Status</strong>: BT is strong in <strong><em>connectivity</em></strong> but weak in the <strong><em>application</em></strong> <strong><em>ecosystem</em></strong>.</li><li><strong>Practical Challenges</strong>: As AI is a service, its resilience lies <strong><em>NOT just</em></strong> in the hardware being in the UK, but more importantly in whether the models can continuously evolve. If the sovereign AI centre becomes an isolated “<strong>computing island</strong>” unable to access the world’s most advanced open-source ecosystems, this resilient service will turn into a “<strong>legacy guarantee</strong>” that fails to attract developers.</li></ul><p>BT’s management has reached a unified consensus to position sovereign cloud as a <strong><em>Resilient Service</em></strong><em>,</em> marking a shrewd marketing pivot. It has successfully replaced the<em> </em><strong><em>expensive and somewhat outdated concept of “sovereignty” with “risk management,”</em></strong> a term that corporate executives find much easier to buy into.</p><h4>Final Judgment by Jimmy:</h4><ul><li><strong>In the Short Term</strong>: This narrative possesses immense potency within the UK Government (<strong>Public Sector</strong>) and <strong>Regulated Industries</strong>. It can help BT quickly secure high-net-worth clients who have a deep-seated fear of being “<strong>cut off</strong>.”</li><li><strong>In the Long Term</strong>: Whether this “<strong>resilience</strong>” can withstand the test of practice depends on whether BT can truly solve the issues of “<strong>scale fragmentation</strong>” and “<strong>underlying technology dependency</strong>.” If it is merely an outsourced operation service wearing the cloak of sovereignty, it may turn out to be nothing more than an expensive, state-funded transitional solution.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uG-S42pcPyjmvVnG8TqWtQ.jpeg" /></figure><h3>National Strategy and the Construction and Strengthening of Credibility</h3><h4>1. A Game Theory Perspective: Establishing “Credible Commitment”</h4><p>In game theory, to make opponents or customers believe in your resilience, the most important method is <strong>“burning the bridges”</strong>.</p><ul><li><strong>Technical “Local Forking” Capability</strong>: BT must prove that even if its connection with NVIDIA or cloud partners is severed, it possesses not only the hardware but also the full technology stack mirrors and autonomous O&amp;M (Operations and Maintenance) capabilities to maintain the operation and updating of core models without external support. If BT can publicly demonstrate simulation exercises in “<strong>offline mode</strong>,” its credibility will far exceed any marketing presentation.</li><li><strong>Capital Accumulation and Transparent Auditing</strong>: By establishing a legally binding trust structure fully audited by the UK government or an independent third party, BT must ensure that operational control of its sovereign cloud can be forcibly taken over by the UK government under any circumstances (including the acquisition of BT by foreign capital or the withdrawal of partners).</li></ul><h4>2. The “Hardening” of Physical Sovereignty: Addressing Near-field Risks like the “Super Embassy”</h4><p>A certain country’s <em>“London Super Embassy”</em> is located in close proximity to the City of London and BT’s key network nodes/routes, which in itself exposes the <strong><em>UK’s shortcomings in physical-layer security awareness </em></strong>and <strong><em>the credibility of its information security.</em></strong> To prove its credibility, BT must make extraordinary investments in physical resilience:+1</p><ul><li><strong>Anti-electromagnetic Interference and Anti-eavesdropping (TEMPEST Standard)</strong>: Facing the risk of side-channel attacks brought by “<strong>physical neighbours</strong>,” BT’s sovereign data centres must prove they possess <strong><em>military-grade electromagnetic shielding and vibration monitoring capabilities</em></strong>. This practice is only persuasive if BT can prove that its core computing nodes employ <strong><em>geographic redundancy</em></strong> and that their routes completely avoid potential risk areas (<em>such as the surroundings of sensitive buildings)</em>. Current electromagnetic eavesdropping technology and non-destructive fibre optic cable eavesdropping technology are no longer secrets, without physical information security, all other security guarantees are out of the question.</li><li><strong>Full-link Physical Traceability</strong>: This applies not only to data centres but also to the absolute sovereignty of fibre optic routing. BT needs to prove to high-compliance customers that its data transmission paths do not pass through any “<strong>black box</strong>” switches not controlled by the sovereign power. Thirty years ago, when I graduated and worked at <strong><em>Alcatel</em></strong> (a TOP 1 French Telecom system supplier at that time) on the development of its System 12 switching system, one of the functions was Lawful Interception (LI). To believe that the communications systems of any hostile nation are secure is and in control or transparent, in itself, an onset of naivety.</li></ul><h4>3. Strategic “Decoupling Contingencies”: From “Leveraging” to “Control”</h4><p>BT’s current model of “leveraging international partners for underlying technology” is essentially a form of “<strong>half-sovereignty</strong>”. To strengthen credibility, “<strong>supply chain anxiety</strong>” must be addressed:+1</p><ul><li><strong>Strategic Computing Power Reserves</strong>: This involves more than just existing GPUs; it requires a strategic inventory of critical components (similar to a Strategic Petroleum Reserve). A hard standard for judging whether BT possesses “<strong>true resilience</strong>” is whether it can prove it has enough hardware spares to support 3–5 years of maintenance needs.</li><li><strong>Embracing “Open-source Sovereignty”</strong>: Over-reliance on closed-source international commercial AI frameworks is the greatest enemy of resilience. BT should take the lead in supporting local fine-tuning based on open-source architectures (such as RISC-V computing exploration or open-source large model frameworks) to ensure that even if commercial software licenses are blocked, core business can still “<strong>persist under duress</strong>”.</li><li><strong>National Strategy</strong>: The government and enterprises should support the R&amp;D and market presence of ARM and local UK chip companies; this is a national strategy, otherwise, there will be no long-term security or credibility.</li></ul><h4>4. Practices to Address “Weak Risk Awareness”: Establishing “Red vs. Blue” Teaming Mechanisms</h4><p>Planning loopholes in the City of London <em>(such as the location of a certain super-embassy)</em> reflect <strong><em>sluggishness and a lack of risk awareness at the UK administrative level.</em></strong> If BT wants to transcend this sluggishness commercially, it must establish security standards that exceed government requirements:</p><ul><li><strong>Normalised “Fracture Testing”</strong>: Imitating Netflix’s “<strong>Chaos Monkey</strong>” model, BT should invite UK military security departments to conduct regular real-world exercises involving “<strong>power cuts, network cuts, and authorisation cuts</strong>,” and submit the results (after de-sensitisation) to customers as resilience reports.</li><li><strong>Reversing Policy Pressure</strong>: BT should leverage its industry status to push <strong>DSIT</strong> (<em>Department for Science, Innovation and Technology</em>) to formulate a stricter “<strong>Critical Digital Land Protection Act</strong>”. This shift in role from “participant” to “driver of rule-making” is the highest state of building strategic credibility.</li></ul><h4>Summary: Trust Stems from Exercises for the “Worst-case Scenario”</h4><p>If a <strong><em>sovereign cloud/AI</em></strong> is merely a “<strong>voucher-style</strong>”<strong> <em>resale business,</em></strong> it will be nothing more than waste paper in a moment of crisis.</p><p><strong><em>The only way for BT to strengthen its credibility is to show the market that it has made all the difficult and expensive preparations for “the worst day”. </em></strong>If it dares not discuss how to respond to <em>“eavesdropping by physical neighbours” </em>or the “colla<em>pse of the global supply chain,”</em> then its<strong> Resilience is merely a marketing show</strong>. This returns once again to the previously stated view: <strong><em>rather than letting various operators pursue fragmented “half-baked” sovereignty, it is better to establish a national, neutral, shared sovereignty centre with military-grade defence standards.</em></strong></p><p>Only a gathering of this scale can afford <strong>true dual protection — both physical and logical.</strong> Therefore, what follows are recommendations for the government.</p><h3>The Indispensable Role of Government Regulation in Sovereign Cloud/AI Strategy</h3><p>If the core proposition of “<strong>digital sovereignty</strong>” — which involves national security and long-term competitiveness — is left entirely to profit-seeking commercial companies <em>(such as </em><strong><em>BT</em></strong><em> and </em><strong><em>Vodafone, </em></strong><em>etc.)</em> to build in a fragmented manner, it will ultimately evolve into a closed market characterized by <strong>“high premiums, low efficiency, and poor interoperability,”</strong> rather than a true national strategic asset.</p><p>To achieve the <strong>“neutral, shared Sovereign Cloud/AI Center”</strong> I propose, <strong>DSIT</strong><em> (Department for Science, Innovation and Technology) </em>and <strong>OFCOM</strong> <em>(the UK’s communications regulator)</em> must transition from being <strong>“observers/post-event regulators” </strong>to <strong>“top-level designers and strategic coordinators”.</strong></p><p>Below are the specific regulatory and policy measures they could adopt:</p><h4>1. DSIT: From “Issuing Subsidies” to “Defining Top-Level Architecture”</h4><p>DSIT’s current role is mostly focused on issuing small-scale innovation funds (such as the meagre <strong>£500 million</strong>), which is insignificant in the face of the massive global computing power race. What they should be doing instead is:</p><ul><li><strong>Defining “Sovereignty Tier” Standards (Tiered Certification)</strong>: DSIT should issue official standards to clarify what constitutes true “sovereignty”. For example:</li><li><strong>Level 1 (Basic Compliance)</strong>: Data residency only.</li><li><strong>Level 2 (Operational Sovereignty)</strong>: Operated by a UK-based team.</li><li><strong>Level 3 (Strategic Sovereignty)</strong>: Hardware supply chain backup + software stack auditing + physical security isolation. Through this classification, the government can prevent commercial companies from engaging in <strong>“Sovereign Washing”</strong> and guide government procurement toward high-level, neutral facilities.</li><li><strong>Establishing a “National AI Trust”</strong>: Rather than having operators purchase GPUs individually, DSIT could lead the establishment of a national, non-profit computing entity. Through a <strong>PPP (Public-Private Partnership)</strong> model, it could centrally procure GPU clusters of the 10,000-card magnitude and lease them as a <strong>Neutral Facility</strong> to BT or other operators. This would achieve economies of scale and avoid the resource fragmentation I fear.</li><li><strong>Physical Security Planning and Sensitive Zone Control</strong>: In the face of risks like the<strong><em> “London Super Embassy,”</em></strong> DSIT must coordinate with local governments and security departments to designate <strong>“Critical Digital Security Zones”</strong>. Within these zones, non-trusted entities should be strictly prohibited from approaching.</li></ul><h4>2. OFCOM: From “Market Competition” to “Open Access and Resilience Auditing”</h4><p>OFCOM has long focused on <strong><em>price competition</em></strong>, but in the AI era, it needs to extend its regulatory reach to computing infrastructure:</p><ul><li><strong>Mandatory Interoperability</strong>: OFCOM could stipulate that any provider claiming to offer “Sovereign Cloud” must provide standardized APIs and data migration interfaces. This would prevent operators from using “sovereignty” as a pretext for <strong>Vendor Lock-in</strong>, ensuring that national data can flow freely between different sovereign nodes, thereby achieving the “federated model” I mentioned.</li><li><strong>Promoting the “Neutral Host Model”</strong>: Just as I suggest that the UK operators’ <strong>RAN (Radio Access Network)</strong> should be neutralized, OFCOM could push for a policy of <strong>“separation of computing power and operations”</strong>. That is, the underlying expensive GPUs and data centers would be held by a neutral third party, while operators (BT, Vodafone, etc.) would only be responsible for top-layer value-added services and customer operations. This would institutionally end the redundant construction competition among operators. Without this, a £500 million fund or a 14MW AI center simply will not be competitive.</li><li><strong>Resilience Auditing</strong>: According to the <strong>Telecommunications (Security) Act 2021</strong>, OFCOM should conduct regular “extreme stress tests” on sovereign clouds. This includes testing whether the facility can independently support critical public services (such as the <strong>NHS</strong>) during simulated supply chain disruptions or physical attacks (such as the near-field threats mentioned earlier).</li></ul><h4>3. Policy Coordination: Breaking the Deadlock of the “Chicken Game”</h4><p>The current situation is a deadlock: BT must pursue profit for its financial reports, other operators must follow suit to avoid losing market share, and the government remains a spectator. To break this:+1</p><ul><li><strong>Procurement as a Policy</strong>: The government, as the largest customer (NHS, MoD, Gov.uk), should explicitly state that it will only procure services that are connected to the “<strong>National Shared Sovereignty Framework</strong>”. This move would instantly compel commercial companies to abandon the idea of building closed, private circles and instead participate in neutral, shared platforms.</li><li><strong>Addressing Weak Physical Security Awareness</strong>: Regarding the “joke” of an embassy co-existing next to the City’s data nodes, OFCOM should have a <strong>“National Security Veto”</strong>. Any change in the land use nature around critical digital infrastructure must undergo an information security risk assessment, rather than just a standard urban planning approval.</li></ul><p>There is a <strong><em>natural misalignment</em></strong> between the <strong><em>“profit motive of commercial companies”</em></strong> and the <strong><em>“absolute security guarantee of the nation”.</em></strong> If DSIT and OFCOM continue to be absent, the UK’s sovereign cloud market will become:</p><ul><li><strong>Users</strong>: Paying high premiums.</li><li><strong>Operators</strong>: Trapped in low-return, redundant construction competition.</li><li><strong>The Nation</strong>: Discovering in a moment of crisis that so-called “sovereignty” is merely a castle in the air dependent on external supply chains.</li></ul><p>The optimal solution is for<strong><em> DSIT to provide strategic direction and security standards, the government to invest in neutral underlying computing facilities, and OFCOM to supervise open access and resilience. </em></strong>Let companies like BT return to the role of “service providers” rather than allowing them to gamble with the nation’s digital lifeblood in a high-risk market game without regulatory guidance.</p><h4>4. BT’s Shareholding Structure: Misalignment of Capital and National Wills</h4><p>BT’s top three shareholders are: <strong>Bharti Global</strong> (approx. 25%), <strong>Deutsche Telekom</strong> (approx. 12%), <strong>BlackRock, Inc.（</strong>Approx.5.1%) and <strong>Mexico’s Carlos Slim</strong> (approx. 4.3% via Inbursa/América Móvil).</p><p>BT is a commercial company listed on the London Stock Exchange and must be accountable to its <strong><em>Indian, German, and Mexican shareholders</em></strong>. When national interests<em> (such as establishing a low-margin but high-security neutral shared center)</em> conflict with shareholder interests (pursuing high-margin private sovereign cloud premiums), BT, as a commercial firm, will inevitably choose the latter. Shareholders pursue <strong>Return on Investment (ROI)</strong>, whereas true “national resilience” often requires redundancy and neutrality regardless of cost. <strong>This natural contradiction significantly undermines the credibility of “commercial companies providing sovereign resilience”.</strong></p><p>While the<em> “Super Embassy”</em> next to the City of London demonstrates the government’s weak physical security awareness, BT’s shareholding structure reflects a confusion in <strong>“Strategic Security Awareness”</strong>. Although the UK government holds “Golden Shares” or review powers under the <strong>National Security and Investment (NSI) Act</strong>, it is logically difficult to provide “absolute guarantees” of credibility when a company whose largest shareholder is Indian and whose second-largest shareholder has a German government background is tasked with carrying the “Sovereign AI” of the NHS and the MoD.</p><p>This is precisely the product of the UK’s <strong><em>“free-market utopia” </em></strong>of the past 20 years <em>(which is why I began by questioning the “The World is Flat” perspective).</em> <strong>The UK government threw its critical infrastructure entirely into the market, and as a result, when it now tries to establish “sovereignty,” it finds that the foundations (equity and physical assets) have long been permeated by global capital.</strong></p><p>If DSIT and OFCOM do not step up to establish <strong>“neutral national-level sovereign facilities,”</strong> and instead allow commercial companies with such complex shareholding structures to fight over the market, so-called Resilience will ultimately remain a “<strong>fragile contract</strong>” established under the control of foreign capital.</p><blockquote>Now, I have to add one noise voice, how stupid those HRs in DSIT &amp; OFCOM, which rejected my application for a cleaner’s role there twice times.</blockquote><h4>Conclusion</h4><p>BT’s packaging of its offering as a “<strong>Resilient Service</strong>” is a sharp observation of market trends, also is the awareness of the “unbearable heavy of sovereignty with national security concerns”, but <strong><em>in the absence of neutral infrastructure and rigorous physical security controls, this resilience is more like a fortress built on shifting sands.</em></strong> True digital sovereignty should not be a tool for commercial giants to extract high “insurance premiums,” but a solid foundation guided by the government and shared by multiple parties.</p><p>If the government is truly to be responsible for the security and sovereignty of national data, the credibility of “resilient services” from commercial companies like BT should be subjected to regular, fully transparent, national-level security stress tests — just like nuclear power plants. This is what truly responsible and accountable government regulation looks like.</p><p>Apr.30, 2016</p><h3>Jimmy CHOW</h3><blockquote><strong>Professional Bio / Collaboration Offer</strong></blockquote><blockquote><strong>Strategic Network Transformation &amp; Integration Expert</strong></blockquote><blockquote>I am currently open to new professional opportunities and consulting engagements, yes, just want to do sth., otherwise I will do bricklayer works.</blockquote><blockquote>If your organisation is navigating the complexities of <strong>multi-generation network integration</strong> (2G/3G/4G/5G/6G) for a future-proof <strong>cloud-native core network</strong> or managing the intricate network architectures and fragmented service platforms that arise from <strong>mergers and acquisitions (M&amp;A)</strong> aiming for an<strong> Autonomous Network Architecture</strong>, I offer the expertise to streamline your operations.</blockquote><blockquote>My focus is on driving operational efficiency and fostering competitive advantage by aligning current infrastructures with future-ready network evolution. Whether working independently or embedded within your existing team, I provide end-to-end support — from <strong>situational analysis and strategic roadmap design to hands-on implementation and seamless transition</strong>.</blockquote><blockquote>Let’s collaborate to transform your legacy complexities into a lean, agile, and future-proof digital foundation.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=443d8bb62a72" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sovereign Cloud/AI: Sovereignty on the Left, the World on the Right]]></title>
            <link>https://medium.com/@mountdavistechnologies/sovereign-cloud-ai-sovereignty-on-the-left-the-world-on-the-right-0e7dc69b431d?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e7dc69b431d</guid>
            <category><![CDATA[sovereignty]]></category>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Wed, 29 Apr 2026 19:06:01 GMT</pubDate>
            <atom:updated>2026-04-29T19:06:01.366Z</atom:updated>
            <content:encoded><![CDATA[<p>Jimmy CHOW</p><p>In a previous article, I provided a detailed introduction to <a href="https://medium.com/@mountdavistechnologies/bt-sovereign-ai-the-rise-of-uk-sovereign-cloud-and-ai-can-it-carve-a-new-path-amid-hyperscaler-720d109e6ecf"><strong><em>BT’s Sovereign AI/Cloud</em></strong></a> in the UK. From the perspective of a commercial company, I praised it for its rapid adaptation to shifting international political winds and its swift launch of compliant <strong>private cloud</strong> services. However, my habitual way of thinking tends to weigh negative assessments more heavily. <strong><em>Gaining a clear view of the “other side” of any bright situation is often the best start to avoiding unexpected outcomes.</em></strong></p><p><a href="https://medium.com/@mountdavistechnologies/bt-sovereign-ai-the-rise-of-uk-sovereign-cloud-and-ai-can-it-carve-a-new-path-amid-hyperscaler-720d109e6ecf">BT Sovereign AI: The Rise of UK Sovereign Cloud and AI — Can It Carve a New Path Amid Hyperscaler…</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TN7Bxh5vgDvTjXs94XDyww.png" /></figure><h3>The Geopolitical Shift</h3><p>Currently, nations are rethinking their “<strong>Sovereign XX</strong>” strategies. From a political-strategic standpoint, this is undoubtedly the correct choice. Over the past 20 years, the global political landscape has undergone earth-shattering changes. While economic globalization brought benefits to some, it undeniably created an existential crisis for others. When countries with diametrically opposed values collaborate for short-term economic interests, they risk nurturing their own “gravediggers”. This is true for the US, Europe, and especially the UK — without exception.</p><p>When the famous book<em> </em><strong><em>“The world is Flat: A Brief History of the Twenty-first Century” by Tomas L. Friedman (2005) </em></strong>was published over twenty years ago, it was pursued by countless people worldwide chasing a <strong><em>utopian dream of global unity/village</em></strong>. I still keep my critical article from that time that mocked such views as naive and lacking in historical awareness; looking back now, everything developed so far seem quite laughable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/866/1*n-Ul1f5BkF33HU2DAfyNNg.png" /><figcaption>Screen capture of my article collections about 20 years ago</figcaption></figure><h3>The “Semi-Exclusionary/Closed” Model</h3><p>Now no one cares about the world now is flat, or crimped, it is actually collapsed and ramped. How far can this “<strong>closed and semi-exclusionary</strong>” sovereign cloud model <em>(represented by BT Sovereign xx)</em> will go on the road to business success?</p><p>The core of this model is prioritizing<em> “</em><strong><em>Operational and Legal Sovereignty</em></strong><em>”. </em>This means <strong><em>data, teams, and legal responsibilities </em></strong>remain 100% in the UK (or the home country), emphasizing a <strong>UK-hosted, UK-operated, and UK-cleared</strong> approach. It is not entirely closed but rather “semi-exclusionary” — it strictly isolates itself from foreign legal influences <em>(such as the </em><strong><em>US CLOUD Act</em></strong><em>) </em>while maintaining hybrid cloud compatibility with the global ecosystem by leveraging international partners for underlying technology (<em>e.g., Rackspace (USA) hosting + Nscale + NVIDIA(USA) AI computing power</em>).</p><h4>Short-Term Prospects (2–5 years)</h4><p>The outlook for the next 2–5 years is optimistic, particularly within regulated industries/sectors in the UK and Europe. Commercial success is likely due to several factors:</p><ul><li><strong>Favorable Drivers</strong>: Geopolitical tensions will continue, stricter data localization (like “GDPR plus”) will be a must, AI safety concerns, and the rigid demand for “<strong>zero foreign risk</strong>” in the public sector provide a natural moat.</li><li><strong>Economic Potential</strong>: BT research suggests digital sovereignty could unlock <strong>£18 billion</strong> in annual AI productivity value for the UK and drive a <strong>£13.6 billion</strong> sovereign cloud market over the next five years, alongside <strong>£14.6 billion</strong> in data center investment by 2030. Though this optimistic evaluation is not for sure based on too 好的，为您继续翻译关于 <strong>Reliance Jio</strong> 方案的部分，采用标准英国英语（Standard British English）：many factors.</li><li><strong>European Context</strong>: Other European operators (<em>such as Deutsche Telekom’s T-Cloud Public and Orange’s European Edge Continuum</em>) are also accelerating sovereign layouts in 2026, targeting high-value clients in <em>government, defense, healthcare (NHS), and finance</em>. These sectors prioritize trust and compliance over cost and support by government financial budgets.</li><li><strong>Policy Support</strong>: If the UK or European governments implement stricter data sovereignty policies for strategically important sectors, the market could grow even further. But this will lead to a more closed world.</li><li><strong>Hybrid Strategies</strong>: BT and others use hybrid cloud strategies <em>(sensitive workloads in sovereign environments, non-sensitive with hyperscalers)</em> to lower migration barriers. European sovereign cloud spending is expected to grow by <strong>35.6%</strong> in 2026 to reach <strong>$80 billion</strong>.</li></ul><h4>Long-Term Challenges (5–10+ years)</h4><p>In the mid-to-long term, achieving comprehensive commercial success will be difficult:</p><ul><li><strong>Niche Market vs. Mass Market</strong>: Sovereign cloud is a niche strategy, it’s not a coffee bar on high-street with location is the dominant factor for business success. The global cloud market is dominated by <strong>AWS, Azure, and Google</strong> (60–70%+ share), who possess extreme barriers in economies of scale, network effects, and rich ecosystems.</li><li><strong>Cost and Innovation Gaps</strong>: Sovereign solutions often entail higher costs <em>(dedicated data centers, relatively small scale, compliance modules bringing complexity, etc.)</em> and slower innovation. While suitable for government or state-backed banks, they struggle to attract innovative tech companies that seek cost-performance and cutting-edge AI tools.</li><li><strong>Hyperscaler Localization</strong>: Hyperscalers are quickly localizing (e.g., <strong><em>AWS European Sovereign, Azure Local air-gapped),</em></strong> offering “<strong>compromise sovereignty</strong>” options that many customers may prefer over purely local ones. Operators like DT claimed in 2026 that they would achieve <em>“core functional parity” </em>with hyperscalers by the end of the year, but actual implementation still needs to be verified, and the computing power competition <em>(GPU scale, energy consumption) </em>in the AI ​​era will put enormous pressure on operators.</li><li><strong>Resource Pressure</strong>: The AI-era “arms race” for GPU scale and power consumption puts immense pressure on telecom operators.</li><li><strong>Trust Risks</strong>: <em>If “sovereignty” is perceived as more marketing than substance, trust could collapse. </em>History shows that<strong><em> purely closed models are difficult to sustain</em></strong>; operator clouds usually end up as supplements rather than mainstream. History already has lots of failed precedence for lessons.</li></ul><p><strong>Overall conclusion:</strong></p><p>This model can go further in highly compliant vertical sectors (public/regulated industries), becoming a stable and profitable “safe” business line, but it’s unlikely to shake hyperscalers’ dominant position in general cloud and AI innovation.</p><p>Its maximum potential success is to become <strong><em>a local “guardian of digital infrastructure/guardian of sovereign xx,” </em></strong>not a global cloud giant. Business sustainability depends on government procurement efforts, the maturity of the hybrid ecosystem, and continuous scalability. European telcos are forming alliances<em> (such as the ESTIA alliance and Edge Continuum), </em>but commercialization still faces challenges without large-scale public spending. All failed business shared the same reasons.</p><p>In smaller markets like the UK, independent Sovereign XXs face even greater risks.</p><h3><strong>Why Telecom Cloud Platforms Failed to Effectively Compete with Hyperscalers Over the Past decade or so (approx. 2010–2026)</strong></h3><p>Around 2009, when the previous company I worked for was evaluating whether to establish a new Cloud PL (Product Line), I was one of members of the expert group. At that time, as an equipment supplier to operators, we truly hoped to provide them with corresponding solutions/platform to assist in their <strong><em>digital transformation strategies</em></strong>. However, more than ten years have passed, and while telcos <em>(such as BT, AT&amp;T, Deutsche Telekom, Orange, and China’s three major operators, etc.) </em>have made various attempts to build their own or partner on cloud platforms<em> (including early NFV, telco cloud, edge cloud, and hybrid cloud, etc.), </em>their overall market share remains stuck in the low single digits or can be omitted. They have failed to break the monopoly of hyperscalers, most are only suitable for internal daily work use, and cases of large-scale commercial success are rare.</p><p><strong>What are the core reasons?</strong> In my personal opinion, they include structural, organisational, and strategic differences with hyperscalers:</p><ul><li><strong>Strategic Focus and Cognitive Bias (Network-centric vs. Application-centric):</strong> Operators have long focused on connectivity and physical infrastructure. Early cloud efforts mostly stalled at network virtualisation (VPN, edge services), colocation/hosting, or NFV, rather than software-defined general-purpose compute/storage/platform services. Hyperscalers built API-driven, application-centric platforms from the ground up, directly addressing the diverse needs of developers. This cognitive gap caused operators to miss the evolution of the cloud from a “network tool” to a “digital infrastructure platform”. To put it simply, it is like a property developer who only cares about building houses but provides no attractive supporting facilities, yet claims to offer “move-in ready” homes; consequently, the properties remain vacant for long periods and the business is bleak.</li><li><strong>Insufficient Innovation Speed, Agility, and Cultural Inertia:</strong> Operators are restricted by legacy hardware, long investment cycles, and regulatory constraints (universal service, reliability obligations), making decisions slow and iteration difficult. Hyperscalers adopt a software-first, DevOps culture, rapidly launching hundreds of services. A shortage of cloud-native skills within operators and a cultural resistance to rapid change further widened the gap. I once collaborated with one of China’s biggest operators to provide a SaaS service for their customers using a revenue-sharing business model. To meet various requirements, such as information security audit, testing took nearly a year, and costs ballooned to over a million only for this audition; for some small companies, this is simply fatal. Anyone who has tried hyperscalers knows that the difference is one year versus a few hours or days.</li><li><strong>Massive Gaps in Economies of Scale and Capital Investment:</strong> Hyperscaler CapEx surged from <strong>$24 billion</strong> in 2015 to <strong>$325 billion</strong> in 2025, spent on global data centres, AI accelerators, and custom hardware to achieve extremely high resource utilisation. Operator CapEx is primarily invested in network upgrades; their data centre utilisation is low, as they must be designed for peak loads and cannot be used by multiple tenants like hyperscalers. This results in operator clouds having higher costs and fewer services. From 2026 to 2030, the scale of AI DC construction in the US will exceed the investment of any other technical transition period in history. The construction of hundreds of billions of dollars worth of DCs annually directly turned a company that once sold gaming graphics cards into the world’s number one with a market capitalisation exceeding <strong>$5 trillion</strong> on the NASDAQ.</li><li><strong>Weak Ecosystem and Developer Attraction:</strong> Hyperscalers have built vast ISV (Independent Software Vendor) ecosystems, rich toolsets, and a consistent global experience, creating powerful network effects and lock-in. Operator cloud services have fewer varieties and high integration difficulty, making it hard to attract developers or enterprises to migrate their entire stacks. Many telco clouds ultimately became “internal tools” or edge supplements rather than public platforms. The same applies to my previous personal experience: the integration difficulty and compliance costs basically shut out startups and individual innovative developers, yet this is precisely the DNA of hyperscalers’ success.</li><li><strong>Organisational and Business Model Differences:</strong> Operators excel at stable B2B/B2G services but have weaker customer orientation; hyperscalers are highly customer-centric with flexible pricing (consumption models). Operators also face regulatory burdens (reliability reporting, universal service), while hyperscalers are more lightly regulated and have more freedom to innovate.</li><li><strong>Misalignment of Technology and Market Timing:</strong> In the early stages, operators focused on hardware “black boxes” and proprietary systems, missing the cloud-native wave. Core network workloads, due to security, resilience, and regulation, mostly remain in self-built DCs to this day (even if hyperscalers provide telco verticals, large-scale migration remains difficult). Following the AI explosion, the computing power race further amplified the gap.</li></ul><p><strong>Summary:</strong> The failure of operator clouds over the past decade was not due to poor technology, but rather structural disadvantages (scale, innovation, management, culture) combined with strategic positioning deviations (Pipe Thinking vs. Platform Thinking). The current sovereign cloud model is the operators’ “Third Path” — avoiding direct price and scale wars in favour of differentiation through trust and compliance. However, it still inherits some historical limitations (small scale, reliance on partners); commercial success will depend on policy support and a hybrid ecosystem rather than replacing hyperscalers.</p><p>In the 2026 era of AI and geopolitics, this semi-exclusionary model is expected to gain more space in Europe and the UK than in the past, <em>but to succeed in the long run, it still needs to address issues of innovation speed, cost competitiveness, and ecosystem openness. </em>Within a limited timeframe, this is essentially a development that is unlikely to be seen. Ultimately, the most likely outcome is<em> “coexistence and layering”</em>: hyperscalers will dominate the global general-purpose and innovation layers, while operators and local solutions will lead the high-sovereignty and regulated layers. Technically capable operators might, at most, quickly mimic some of the new innovative solutions from hyperscalers, while most operators will simply become less-vibrant platforms supported by government finances support.</p><h3>Is the BT Sovereign XX Model the Optimal Strategy for the UK Market?</h3><p>If all major operators in the UK<em> (a relatively small nation with a population of approximately 68 million and an economy far smaller than the US, China, or the EU as a whole)</em> pursue a path similar to BT Sovereign — building their own “purely locally controlled” private/hosted sovereign clouds and Sovereign AI facilities (relying on<em> local teams and international technical partners such as NVIDIA, Rackspace, and Nscale)</em> — it will bring a series of negative impacts to the market. These impacts are primarily reflected in <strong><em>efficiency, cost, innovation, and long-term competitiveness</em></strong>.</p><h4>1. Primary Negative Impacts</h4><p><strong>Exacerbated Market Fragmentation</strong> The UK cloud market is already dominated by AWS, Azure, and Google (with a global share of over 60–70%). If operators like<em> BT, Virgin Media O2, and Vodafone/Three </em>each launch independent sovereign platforms, it will result in the coexistence of multiple <em>“small but complete” sovereign clouds</em>. Enterprises (especially those with customers across different operators) would face multiple sets of APIs, governance models, migration tools, and compliance processes, increasing integration complexity and switching costs. Historical lessons from Europe <em>(such as early national cloud projects and the partial failure of GAIA-X) </em>show that such fragmentation severely hinders scale effects, making it difficult for local solutions to compete with the unified global ecosystems of hyperscalers.</p><p><strong>Rising Costs and Inefficiency</strong> Each operator would need to invest independently in data centres, AI computing power<em> (high-density GPU facilities consume massive amounts of power), </em>security clearance teams, and compliance modules. Given the UK’s limited resources in terms of electricity, land, and talent, redundant construction leads to waste and low utilisation rates <em>(as they cannot achieve the high-load multi-tenant balancing of hyperscalers)</em>.</p><p>Sovereign clouds typically carry a 15–30% premium <em>(due to dedicated infrastructure and local operating costs); </em>SMEs and users with non-sensitive workloads would face higher bills, ultimately suppressing overall cloud adoption and AI innovation. While BT research optimistically estimates an £18bn sovereign AI opportunity and a £13.6bn cloud market, fragmentation would dilute these gains into internal consumption. Furthermore, the “cake” will not grow indefinitely; limits to growth will appear quickly. I estimate the NHS to be BT’s largest customer and revenue source, but growth is ultimately capped by the government’s fiscal capacity, which now is turning worse than before.</p><p><strong>Innovation Lag and Closed Ecosystems</strong> Operators are traditionally network-centric, with an innovation culture and developer appeal that is weaker than that of hyperscalers. Acting independently further reduces the variety of services, the frequency of updates, and the maturity of AI tools (BT’s 14MW scale is far smaller than a single hyperscaler region). If developers and enterprises cannot access a consistent global experience, it leads to a binary choice between “sovereignty vs. innovation”: sensitive workloads are placed locally (secure but slow), while innovative workloads continue to rely on overseas providers, weakening the UK’s overall digital competitiveness.</p><p><strong>Scattered Investment and Talent</strong> Capital and high-end talent (AI engineers, security experts, data centre operators) are diverted across multiple platforms, making it difficult to form a cluster effect. As a “small nation,” the UK already finds it difficult to compete with the massive CapEx of the United States; fragmentation will make it even harder for local solutions to attract international investment or become a “European benchmark.”</p><p><strong>Amplified Geopolitical and Resilience Risks</strong> A single national event (such as power shortages, policy changes, or supply chain disruptions) could affect multiple local platforms simultaneously, reducing overall national resilience. While hybrid clouds can mitigate this, fragmentation increases the complexity of cross-platform interconnection and potential single points of failure.</p><p><strong>Long-term Drag on the Public Sector and the Economy</strong> Government procurement could become locked into multiple local providers, increasing the administrative burden. In the long run, the pace of AI adoption by UK enterprises may slow, economic productivity may not be released as expected (despite the short-term £18bn potential), and the UK may miss out on the dividends of the global cloud ecosystem.</p><p>Similar discussions in Europe (such as the EU Cloud Sovereignty Framework and GAIA-X experiences) have repeatedly pointed out that a strictly national or operator-led sovereign path easily falls into “sovereign washing” controversies or fragmentation traps, making it difficult to increase the market share of local cloud providers (the share of local European providers once dropped from 29% to around 15%).</p><h4>2. Should Neutral Sovereign Cloud and AI Centres be established? Are they better for long-term healthy development?</h4><p>In my personal view, for a geographically “small country” like the UK (and perhaps a politically “small” one now as well), establishing <strong>neutral, shared sovereign cloud and AI centres</strong> is <strong>more suitable for long-term healthy development.</strong> This is not to deny the short-term value of the BT path <em>(though time may eventually prove it has only short-term value),</em> but rather to suggest it should be a supplement, not the dominant model.</p><p>I find it quite puzzling that the concept of “Sovereign XX” did not appear today — it has been discussed for years — yet the UK’s DSIT and Ofcom have failed to provide corresponding policy guidance or research. This demonstrates a lack of government governance capability <em>(although I applied for positions in both departments, they clearly have “experts” inside who rejected my applications).</em></p><p><strong>Advantages of Neutral Sovereign Cloud/AI Centres:</strong></p><ul><li><strong>Avoiding Fragmentation and Achieving Economies of Scale:</strong> Led by the government or an independent entity (such as a public-private partnership or a federated model similar to GAIA-X), one or two national-level shared facilities could be built. Multiple operators, enterprises, and public sectors would use them together, unifying standards, improving utilisation, and reducing unit costs. AI centres could centrally deploy large-scale GPU clusters to quickly reach economic scale.</li></ul><p>Over twenty years ago, while communicating with other operators, I praised the UK’s “Last Mile Open” policy. Although the current reality is not perfect, it was highly commendable for reducing redundant investment and giving users freedom of choice. However, in the current AI era and the future era of quantum communication, there seems to be no forward-looking policy guidance in this regard. Of course, the government’s £500 million support for AI innovation has been mentioned; I cannot say it is a “joke” (though it is indeed quite laughable).</p><p>To extend the topic, I even believe the RAN (Radio Access Network) of the three major UK operators should be entirely neutral, operated by a third-party independent company (or a joint venture of the three wireless operators). Operators should focus more on carriage, core networks, and customer service/experience; this might be a better approach. Today, UK mobile communications are not just near the bottom among OECD countries — even if you compared them to some African nations, they are generally lagging behind, which is truly embarrassing.</p><ul><li><strong>Enhancing Openness and Innovation:</strong> Adopting a federated/interoperable architecture supports multi-cloud mixing, open-source components, and strict “sovereignty by design” (data residency + local governance + technical transparency). This attracts more developers and balances “control” with “innovation,” avoiding the lag caused by closed systems.</li><li><strong>Better Risk Diversification and Resilience:</strong> Neutral governance reduces dependence on a single operator. Policies can uniformly establish compliance, certification, and switching mechanisms (referencing the EU Data Act’s emphasis on portability). For a small nation like the UK in terms of geography and population, this allows limited resources to be concentrated on challenges regarding power, cooling, and the supply chain.</li><li><strong>Economic and Strategic Benefits:</strong> Sovereign data centres can drive economic growth and employment while serving as the “Guardian of UK Digital Infrastructure,” enhancing the nation’s bargaining power. The EU’s 2026 €180m sovereign cloud tender employs multi-provider diversification (rather than a single provider) precisely to avoid lock-in and fragmentation, which could serve as a policy reference for the UK.</li><li><strong>Long-term Healthy Development:</strong> This promotes competition rather than local market monopolies, preventing “operator clouds” from repeating the failures of the past decade in competing with hyperscalers (due to strategic bias, lack of scale, and cultural inertia). A neutral model is more likely to form a healthy coexistence with hyperscalers (sensitive workloads handled locally + general workloads handled globally), supporting UK AI economic growth rather than internal friction.</li></ul><p><strong>Potential Challenges and Pragmatic Recommendations:</strong></p><ul><li>Governance of a neutral model is complex (Who operates it? How is true neutrality ensured?) and prone to bureaucracy. This requires DSIT and Ofcom to take the lead in policy and regulation.</li><li>It requires strong government leadership + Public-Private Partnerships (operators like BT can participate as critical infrastructure providers rather than building their own separate platforms).</li><li>The UK could learn from the EU Cloud Sovereignty Framework or the French/German partnership models, prioritising the establishment of a “UK Sovereign AI Hub” as a shared national facility, with operator sovereign solutions serving as supplements for specific vertical industries.</li></ul><h4>Overall Conclusion:</h4><p>The BT-style path can respond quickly to regulatory demands in the short term, capture public sector contracts, and provide the UK with a “peace of mind” sovereign option. However, if all operators follow suit, the negative impacts will be more prominent for a “small nation” like the UK: fragmentation, high costs, and lagging innovation. Ultimately, this may turn local clouds into a “high-priced niche” rather than a strategic asset.</p><p>Establishing neutral, shared, and federated sovereign cloud and AI centres is more in line with long-term healthy development — it concentrates resources, reduces internal friction, and enhances competitiveness while maintaining necessary control and flexibility. This is not a binary choice of “all local or all hyperscaler,” but rather the construction of a layered ecosystem: neutral national infrastructure as the foundation + operator/enterprise specific extensions + global hyperscaler supplements.</p><p>If the UK government can promote such neutral facilities under the UK-Cloud and UK Sovereign AI frameworks, it will be more conducive to releasing the £18bn potential and avoiding the “fragmentation vs. dependence” dilemma repeatedly seen in European sovereign cloud discussions. Actual implementation needs to balance speed with sustainability; it is time for the government to demonstrate its level of strategic management capability.</p><p><strong>Missing the AI era could make the next generation of the UK slaves to other AI empires, and let them just dream the past vanished Imperial UK.</strong></p><p>Apr.29, 2026</p><h3>Jimmy CHOW</h3><blockquote><strong>Professional Bio / Collaboration Offer</strong></blockquote><blockquote><strong>Strategic Network Transformation &amp; Integration Expert</strong></blockquote><blockquote>I am currently open to new professional opportunities and consulting engagements, yes, just want to do sth., otherwise I will do bricklayer works.</blockquote><blockquote>If your organisation is navigating the complexities of <strong>multi-generation network integration</strong> (2G/3G/4G/5G/6G) for a future-proof <strong>cloud-native core network</strong> or managing the intricate network architectures and fragmented service platforms that arise from <strong>mergers and acquisitions (M&amp;A)</strong> aiming for an<strong> Autonomous Network Architecture</strong>, I offer the expertise to streamline your operations.</blockquote><blockquote>My focus is on driving operational efficiency and fostering competitive advantage by aligning current infrastructures with future-ready network evolution. Whether working independently or embedded within your existing team, I provide end-to-end support — from <strong>situational analysis and strategic roadmap design to hands-on implementation and seamless transition</strong>.</blockquote><blockquote>Let’s collaborate to transform your legacy complexities into a lean, agile, and future-proof digital foundation.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e7dc69b431d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[6G Core: Can We Build an Autonomous Network Without a Nervous Breakdown?]]></title>
            <link>https://medium.com/@mountdavistechnologies/6g-core-can-we-build-an-autonomous-network-without-a-nervous-breakdown-bcf698e59de4?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/bcf698e59de4</guid>
            <category><![CDATA[3gpp]]></category>
            <category><![CDATA[6g]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Tue, 28 Apr 2026 14:15:56 GMT</pubDate>
            <atom:updated>2026-04-28T16:32:27.117Z</atom:updated>
            <content:encoded><![CDATA[<p>Jimmy CHOW</p><p>Based on <a href="https://www.linkedin.com/pulse/6g-maltais-3gpp-learning-lessons-from-5g-emil-olbrich-4l9pc/"><em>Emil Olbrich</em></a>’s insights from the <strong><em>3GPP Malta 6G standards discussions,</em></strong> this analysis explores various 6G Core Network (CN) proposals from a professional perspective, drawing on decades of personal R&amp;D &amp; other various experience in the field/industry.</p><p>The quote below is from Emil’s article for convenience here.</p><blockquote><strong>Part 4: The Core Network Moves From Debate to Architecture (SA2)</strong></blockquote><blockquote>AI for 6G Architecture (KI#18): The Big One Got Approved</blockquote><blockquote>This is the first concrete foothold for “AI-native architecture” in the 6G core TR, and the range of competing architectural philosophies on the table is itself the story:</blockquote><blockquote><strong>NVIDIA</strong> is proposing a dedicated <strong>6G AI Plane</strong> network architecture with its own interfaces into the (optional) agentic 6G Core NFs. This is the most ambitious structural proposal on the table.</blockquote><blockquote><strong>Ericsson, T-Mobile USA, and Deutsche Telekom</strong> are proposing an optional <strong>6G CN AI domain</strong> that sits alongside the 6G CN PS domain and can evolve independently. That’s a more conservative, add-on architectural model.</blockquote><blockquote><strong>China Mobile and Huawei</strong> are pushing a <strong>6G AI Agent Network Architecture</strong> with intent interpretation built in.</blockquote><blockquote><strong>Jio Platforms</strong> is proposing an <strong>INCF / GGF / AIAF</strong> framework — Intelligent Network Control Function for intent translation, Governance and Guardrail Function for pre-execution validation, AI Audit Function for real-time auditing and execution sandboxing. That’s a governance-first take that should get attention from regulated operators.</blockquote><blockquote><strong>Lenovo</strong> is proposing an <strong>Agentic Interface Function (AIF)</strong> that translates agentic-protocol requests into SBA service invocations — the interesting angle here is that it treats the core NFs as the target of agentic actions rather than embedding agents in them.</blockquote><blockquote>Nobody knows what the heck “AI-native” means. Everyone agrees the core network should be AI-aware. The fight is whether AI lives in a parallel plane, in a bolt-on domain, distributed across agents, or embedded in every NF. That fight is going to shape every architectural decision for the rest of the study.</blockquote><p>Before global operators “rolled up their sleeves” to rapidly deploy 5G, I said that large-scale commercialization would fail before 5G CN matured (cloud-native technology was immature). Many years later, now no one doubted this.</p><p><a href="https://medium.com/@mountdavistechnologies/after-the-hustle-and-bustle-only-reality-remains-e82065de81ce">After the hustle and bustle, only reality remains</a></p><p>Today, the industry is welcoming the arrival of the next “G,” starting to formulate grand dreams. This time, based on the lessons of 5G failure and the hype surrounding AI, it’s even harder to reach a consensus on what 6G CN will look like. Perhaps, ultimately, 6G CN will again be an unattainable goal.</p><p>The clash of the five architectural solutions mentioned by <em>Emil </em>can be seen as <strong>five scenarios for “AI’s coronation”</strong>: radical disruption (Nvidia), conservative old-school (Ericsson), AI enthusiast (Huawei), timid and cautious (Jio), and the nosy newcomer (Lenovo).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2tdLrKD8qKZQafIBhuY5SA.png" /></figure><h3>1. NVIDIA</h3><p>The <strong>Dedicated AI Plane</strong> is the most radical: AI is not a plug-in but a <strong>“third plane” </strong>equal to the <strong>User Plane </strong>and <strong>Control Plane.</strong> Computing is the network. This favours the deep integration of GPU resources within the network.</p><p><strong>Analysis</strong>: The 6G AI Plane proposed by NVIDIA genuinely breaks the “siloed” architecture of traditional telecommunications. If we liken the traditional Control Plane (CP) to a “traffic warden”<em> (responsible for directing traffic and allocating seats)</em> and the User Plane (UP) to a “motorway” <em>(responsible for carrying data)</em>, then NVIDIA’s AI Plane (AP) is more like a <strong><em>“city’s smart brain”. </em></strong><em>what a Beautiful concept!</em></p><h4><strong>Deconstruction</strong>:</h4><h4><strong>1. Core Objective of the AI Plane: Shifting from “Process-Driven” to “Model-Driven”</strong></h4><p>Traditional 3GPP architectures (such as SBA) are based on<strong><em> “state machines” </em></strong>and <strong><em>“signalling flows”.</em></strong> When the network becomes extremely complex (e.g., 1024 antennas, ISAC, MTN, ISC,etc.), signalling flows explode. The goal of the AI Plane is not merely Autonomous Networks Level 4, but rather to become the <strong><em>“inference foundation” </em></strong>of the network:</p><ul><li><strong>Centralised Inference Decoupling</strong>: Previously, every network element (CU/DU/UPF) had to learn AI independently, which was extremely inefficient. The AI Plane provides a unified pool of computing power and models, where network elements only need to request “predictions” or “decisions” via an API.</li><li><strong>Cross-Domain Data Fusion (Data Siphoning)</strong>: It is the sole, legitimate <em>“network-wide data vacuum”. </em>It can simultaneously acquire signalling from the CU, channel metrics from the DU, and perception data from ISAC, thereby making more accurate judgements than a single network element.</li><li><strong>Bearing Physical Layer AI</strong>: NVIDIA’s core interest lies in GPUs. The AI Plane can take algorithms previously hard-coded in low-level hardware (FPGA/ASIC) — such as channel estimation and equalisation — and implement them in real-time using AI on cloud-based GPU clusters.</li></ul><h4>2. How to Communicate with CU/DU/RAN? (Implementation of Vertical Architecture)</h4><p>This is why NVIDIA proposes it as a “plane” rather than a “network function”. It requires a completely new set of vertical interfaces:</p><ul><li><strong>Data Collection Link (Perception Side)</strong>: Obtains raw IQ data or CSI information from the DU.</li><li><strong>Model Update/Inference Link (Execution Side)</strong>: Directly delivers trained beamforming weights to the antenna array.</li><li><strong>Control Feedback Loop (Control Side)</strong>: Communicates with the CU to adjust handover strategies.</li></ul><h4>3. Why Make it an Independent “Plane”?</h4><p>NVIDIA’s move is very much that of a “merchant”, but it is also highly rational:</p><ul><li><strong>Normalisation of Computing Power</strong>: If AI is scattered across various network elements, you need to purchase many types of proprietary chips. If it is independent as a plane, operators can buy large-scale GPU clusters (NVIDIA’s home turf) to process AI tasks for the entire network in a unified manner.</li><li><strong>Breaking Protocol Lock-in</strong>: Traditional signalling protocols (such as RRC) are extremely slow to update. If the AI Plane is independent, models can be updated daily like App versions, without waiting for the lengthy 3GPP standardisation cycles.</li><li><strong>Supporting True “Semantic Communication”</strong>: It does not merely transmit bits; it is also responsible for understanding the semantics behind those bits.</li></ul><h4>4. Its Relationship with AN L4: Executor and Commander</h4><p>TMF Autonomous Network Level 4 (AN L4) is a <strong>business classification objective,</strong> while the <strong>AI Plane is the infrastructure</strong> to achieve this goal.</p><p><strong>AN L4 is the “Objective”</strong>: It refers to the network’s ability to automatically process intent and self-heal from faults.</p><p><strong>The AI Plane is the “Means”</strong>:</p><ul><li>With only a Control Plane, you can only achieve rule-based (if-else) automation (L2/L3).</li><li>With an independent AI Plane, the network possesses large-scale, real-time deep inference capabilities, thereby supporting true L4.</li><li><strong>Example</strong>: When an intent is issued, the AI Plane calculates the optimal path and then directs the Control Plane to modify routing via an “AI-Control interface”.</li></ul><p><strong>Summary: What exactly is the AI Plane?</strong></p><p>It is not merely a functional module of AN L4; it is the <strong>“operating system kernel of 6G”</strong>. In NVIDIA’s vision:</p><ul><li><strong>The Control Plane</strong> is responsible for “compliance” and “connection establishment”.</li><li><strong>The User Plane</strong> is responsible for “fast forwarding”.</li><li><strong>The AI Plane</strong> is responsible for “how to optimise” and “how to perceive”.</li></ul><p><strong>In-depth Analysis</strong>: The ambition behind NVIDIA’s proposal is <strong>“de-telecomisation”</strong>. If the AI Plane becomes the core, then the communication network of the future is essentially a distributed data centre, and the 3GPP protocol is merely a “peripheral interface” for this data centre.</p><h4>Deployment:</h4><p>The independent plane proposed by NVIDIA does not imply that all AI will run in core network data centres thousands of miles away. To resolve latency issues, the AP plane must be physically distributed:</p><ul><li><strong>Near-Edge AP</strong>: Deployed within the base station (gNB) or on the O-RAN DU/CU side. It handles microsecond-level (μs) tasks.</li><li><strong>Regional AP</strong>: Deployed in edge data centres, handling millisecond-level (ms) tasks (such as multi-cell coordination and medium-to-long-term resource scheduling).</li><li><strong>Central AP</strong>: Responsible for model training, intent parsing (RDF/OWL), and network-wide strategies.</li></ul><h4>Unspecified Risks:</h4><p><strong>The Compute Black Hole: Does every base station require a GPU?</strong></p><p>This is precisely NVIDIA’s strategic gambit. If we follow NVIDIA’s AI-RAN vision, the answer is yes, though the form it takes may vary:</p><p><strong>A. Software-Defined Hardware (Software-Defined HW)</strong></p><p>Traditional base stations use dedicated chips (ASIC/FPGA) for signal processing. NVIDIA considers this too rigid. If every base station possessed a Grace Blackwell-level compute unit:</p><ul><li><strong>During Communication</strong>: It uses AI to perform channel estimation and interference cancellation.</li><li><strong>During Idle Time</strong>: It transforms into an edge compute node, leased to nearby enterprises for AI inference (returning to our previous discussion on <strong>Monetisation</strong>).</li></ul><p><strong>B. Compute Pooling (Pooling)</strong></p><p>Through the Open-RAN architecture, computing power does not need to be allocated 1:1 to every antenna. Instead, the processing requirements of several sites are aggregated into a small <strong>Edge Data Centre (Edge DC)</strong> via high-performance fronthaul networks. This allows for statistical multiplexing, reducing the “depth” of the black hole.</p><p><strong>C. How to resolve latency-sensitive services? (AI vs. Determinism)</strong></p><p>The current solution involves <strong>“Hierarchical Loops”</strong>:</p><ul><li><strong>Ultra-Fast Path</strong>: This remains within the hardware (e.g., Tensor Cores). For instance, the application of beamforming weights must be completed in nanoseconds; this cannot wait for AI inference results.</li><li><strong>Predictive Path</strong>: The AI Plane does not “operate physical quantities in real-time” but rather “predicts them in real-time”.</li><li><em>Example</em>: <em>The AI predicts that the channel will collapse in the next millisecond and proactively delivers a set of parameters to the physical layer cache. When that moment arrives, the physical layer switches directly without waiting for confirmation from the AP.</em></li></ul><p><strong>D. The Deep Gambit: The Fear of Operators</strong></p><p>This is why Ericsson, T-Mobile USA, and Deutsche Telekom preferred an <strong>“add-on” or “optional” architecture</strong> at the Malta meeting, rather than NVIDIA’s “independent plane”:</p><ul><li><strong>Cost Fear</strong>: The cost of upgrading GPUs across the entire network is enough to bankrupt operators.</li><li><strong>Energy Efficiency Fear</strong>: The power consumption of a GPU is significantly higher than that of dedicated communication chips. If saving 15% of algorithmic energy consumption results in an additional 50% for cooling, it is logically unsound.</li><li><strong>Supply Chain Fear</strong>: If the 6G architecture is entirely based on the AI Plane, and the AI Plane is highly dependent on GPUs, then operators will be deeply locked into a single vendor (NVIDIA) (<strong>Vendor Lock-in</strong>).</li></ul><h3>Rational Summary: Is this AP plane reasonable?</h3><p>Mathematically and in terms of performance, it is extremely reasonable. The efficiency of AI in processing non-linear channels is indeed higher than traditional mathematical formulas.</p><p>From a commercial and engineering perspective, it is a massive gamble. My previous articles analysed scenarios like CSI-RS, ISAC, and ISC, which traditional statistics-based algorithms cannot support; Machine Learning certainly has its place.</p><p><a href="https://medium.com/@mountdavistechnologies/ai-or-not-ai-a-question-from-the-soul-of-6g-ran-1ceda98ee2aa">AI or Not AI, a Question from the Soul of 6G RAN</a></p><h4><strong>My Judgement</strong>:</h4><p>NVIDIA’s Dedicated AI Plane may eventually be pared/crippled down into a<em> </em><strong><em>“Logical Plane”</em></strong><em>. </em>That is to say: the standard will define a set of AI interfaces, but the underlying hardware will not necessarily be a GPU.</p><ul><li><strong>Huawei</strong> may promote the use of its own <strong>Ascend</strong> chips.</li><li><strong>Qualcomm</strong> will push for AI to run on the <strong>Hexagon DSP</strong>.</li><li><strong>Operators</strong> will choose to deploy GPUs in critical “hotspot areas” to gain performance, while continuing to use low-power traditional solutions in vast rural areas.</li></ul><p><strong>For operators, it has come to a soul-searching question:</strong></p><p><em>Do you want </em><strong><em>an expensive “AI-everywhere” network</em></strong><em>, or </em><strong><em>an economical “intelligence-on-demand” network</em></strong><em>? Particularly for those companies whose share prices have already plummeted to single digits.</em></p><h3>2. Ericsson, T-Mobile, and DT: The “Evolutionary Add-on” Model</h3><p>While Huawei is busy installing a “brain” and NVIDIA is “demolishing and rebuilding,” the proposal from Ericsson, in collaboration with T-Mobile USA and Deutsche Telekom, resembles a <strong>“carefully planned sunroom extension”</strong> to an existing villa.</p><h4>1. Core Logic: Decoupling and Independent Evolution (The Optional AI Domain)</h4><p>The essence of the Ericsson proposal is the introduction of an <strong>Optional 6G CN AI Domain</strong>. This domain sits alongside the traditional Packet Switched (PS) domain, ensuring they coexist without interference.</p><ul><li><strong>Dual-Track Architecture:</strong> AI capabilities are encapsulated within a dedicated domain rather than being forcibly embedded into every individual Network Function (NF).</li><li><strong>Independent Evolution:</strong> This allows AI algorithms to be updated and compute power to be scaled at the pace of the AI industry, without having to overhaul fragile core network signalling processes for every upgrade.</li></ul><h4>2. Reconstructing AN L4: From “Full-Time Surgeon” to “Consultant Specialist”</h4><p>In Ericsson’s vision, AN L4 (Autonomous Networks Level 4) does not need to rewrite the genetic code of the core network. Instead, it establishes a <strong>“high-level advisory board”</strong> on the side.</p><ul><li><strong>On-Demand Invocation:</strong> The core network handles standard traffic (via the PS domain) as usual. Data is only offloaded to the “AI Domain” for inference when complex optimisations — such as energy saving or massive network slicing scheduling — are required.</li><li><strong>Risk Isolation:</strong> The greatest advantage of this “pluggable” structure is <strong>fault tolerance</strong>. If the AI domain “goes rogue,” the operator can simply disable the AI plugin, and the core network continues to function as a traditional communication network. For Tier-1 operators like T-Mobile and DT, who are notoriously risk-averse, this is the safest path to L4 implementation.</li></ul><h4>3. Deep Insight: Why This is the “Commercial Comfort Zone”</h4><ul><li><strong>Legacy Protection:</strong> Operators have yet to fully recoup their massive investments in 5G Core. Ericsson’s proposal allows them to transition to 6G via “plug-ins” without dismantling their existing architecture.</li><li><strong>The Multi-Vendor Balancing Act:</strong> Unlike the deep “vendor lock-in” potentially caused by Huawei’s native-agentic approach, Ericsson’s “Independent Domain” makes it easier to define standardised interfaces. This ensures operators maintain bargaining power between AI providers and equipment vendors.</li></ul><p><strong>Conclusion:</strong> The Ericsson proposal represents the <strong>pinnacle of pragmatism</strong>. It may not chase the millisecond-level extreme performance of 6G native intelligence, but it removes the psychological barrier of “fear to use” for operators. It is the path of least resistance for L4 to achieve large-scale commercialisation.</p><h3>3. Huawei and CMCC (China Mobile)</h3><p>The <strong><em>“6G AI Agent Core Network Architecture”</em></strong> proposed by Huawei and China Mobile (CMCC) at 3GPP exhibits fundamental differences from NVIDIA’s “Compute-centric” school of thought.</p><p>Whilst NVIDIA seeks to<strong> “brute-force” </strong>the physical layer using GPU compute power, Huawei and CMCC are using <strong>AI Agents</strong> to “redefine” control logic.</p><h4><strong>Deconstruction:</strong></h4><h4>1. Core Architecture: AI-Native Agentic Architecture</h4><p>The core of Huawei’s proposal (as outlined in their white paper <em>Vision and Key Technologies of 6G Network Architecture</em>) lies in <strong><em>“Hierarchical Closed-loops”</em></strong><em> and </em><strong><em>“Knowledge-driven”</em></strong><em> mechanisms.</em></p><ul><li><strong>Logical Entities: From Network Functions (NF) to Agents</strong></li></ul><p>Huawei’s proposal evolves traditional <strong><em>5G Network Functions </em></strong>(e.g., AMF, UPF) into <em>self-aware </em><strong><em>Agents</em></strong>.</p><p>Each Agent not only executes instructions but also carries its own <strong><em>Knowledge Base</em></strong><em>.</em></p><ul><li><strong>Three-Layer Intelligent Architecture:</strong></li></ul><ol><li><strong>Node AI:</strong> Handles tasks with extremely high real-time requirements (such as physical layer scheduling). These are completed via embedded AI directly within the node, rather than through external attachments.</li><li><strong>Network AI:</strong> Manages coordination across network functions. This is resolved through <strong>Agent-to-Agent communication</strong> (no longer rigid signalling flows) and negotiation.</li><li><strong>Service AI:</strong> Responsible for interpreting complex <strong>Intents</strong> as defined by the TMF.</li></ol><h4>2. Key Proposal Points: Intent Interpretation &amp; Logical Inference</h4><p>This is the most significant point of divergence between the Huawei/CMCC proposal and NVIDIA’s: they place immense emphasis on <strong>“Semantics”</strong> and <strong>“Intent”</strong>.</p><ul><li><strong>Built-in Intent Parser:</strong> Huawei proposes deploying an <strong>“Intent Management Function (IMF)”</strong> at the entrance of the core network. Rather than merely parsing JSON, it utilizes capabilities similar to <strong>Large Language Models (LLMs)</strong> to map natural language intents to network primitives.</li><li><strong>Logical Inference vs. Computational Inference:</strong> NVIDIA emphasises TFLOPS (computing power), whereas Huawei emphasises <strong>Reasoning (logic)</strong>. Within Huawei’s architecture, the network uses logical rules to determine whether an “intent is legitimate” and whether “conflicts exist”.</li></ul><h4>3. Comparison with NVIDIA: Distributed vs. Centralised</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/781/1*plfu5PD-NomGrAaKaYYbdA.png" /><figcaption>Picture credit: Jimmy</figcaption></figure><h4>4. “Guardrails” in Huawei’s Proposal: Shadows of the Jio Approach</h4><p>Huawei has proposed a four-stage closed-loop: <em>“Sensing-Decision-Execution-Evaluation”. </em>Before execution, a simulation must be verified through a <strong>Digital Twin</strong> layer. This coincides with Jio’s “Governance and Guardrail Function (GGF)”, both of which aim to resolve the issue of “AI misinterpreting intent and causing network paralysis”.</p><h4>5. Why does China Mobile (CMCC) support the “Agentic” architecture?</h4><p>As an operator, CMCC’s pain points are very clear:</p><ul><li><strong>Avoiding Hardware Lock-in:</strong> If NVIDIA’s solution were followed, the entire network would require GPUs, making OPEX prohibitively high. Huawei’s Agent architecture can run on general-purpose CPUs and self-developed NPUs, offering greater flexibility.</li><li><strong>Solving Extreme Complexity:</strong> 6G scenarios (satellite, industrial, ISAC) are too numerous for hard-coded signalling flows, which are already breaking down. <strong>Self-negotiation between Agents</strong> (akin to human social interaction) is the only way to keep the system concise.</li></ul><p><strong>Current Architectural Contention:</strong></p><p><em>Do we truly need an independent “AI Plane”?</em> <em>Or do we simply need to </em><strong><em>add “Agent” attributes</em></strong><em> to the existing “Control Plane” and “Management Plane”?</em></p><p>Combining this with <strong>TMF AN L4 (Autonomous Networks)</strong>, Huawei’s proposal resembles a <strong>“highly logically self-consistent AI Social Network”</strong>. In this architecture, if you — as an OSS maintenance engineer — issue an incorrect intent, the Agent within the system might question you like a colleague: <em>“Are you sure you want to do this? It would violate the energy-saving rules defined by OWL.”</em> In this regard, it aligns more closely with L4 objectives than a network relying solely on brute-force calculation.</p><p>Just guessing if Huawei can really delivery it.</p><h3>Differences or Restructuring Regarding TMF AN L4</h3><p>Huawei and CMCC’s “Agentic” proposal is essentially restructuring the <strong><em>“responsibility boundaries”</em></strong> of AN L4. In the traditional TMF AN framework, AN L4 is often perceived as an <strong><em>external, centralised management layer </em></strong><em>(much like a senior executive in a command centre). </em>However, Huawei’s approach <strong><em>embeds these capabilities directly into the Cloud-Native Functions (CNFs).</em></strong></p><p>This will inevitably lead to <strong>three</strong> <strong>profound transformations for AN L4</strong>:</p><h4>1. From “External Closed-loop” to “Native Closed-loop” (Closed-loop Reconstruction)</h4><ul><li><strong>Old Model (TMF AN L2/L3):</strong></li></ul><p>Collect data -&gt; Send to external OSS/Brain -&gt; Brain makes decision -&gt; Issue command to network element.</p><p><strong>Problem:</strong> High latency; the external brain may lack real-time visibility into the network element’s specific details.</p><ul><li><strong>Huawei Model (6G AI-Native CNF):</strong></li></ul><p>The closed-loop is completed within the network element (Agent). The element senses, decides, and optimises autonomously.</p><p><strong>AN L4 Transformation:</strong> Centralised L4 no longer handles granular daily operations but evolves into a <strong>“Policy Distributor”</strong> and <strong>“Final Arbiter”</strong>. It only oversees whether the “Intent” is correct, rather than adjusting specific “Parameters”.</p><h4>2. Semanticisation of Interface Language (Semantic Interop)</h4><p>If L4 functions are built into the CNF, the way CNFs communicate with upper management systems must change.</p><ul><li><strong>Past:</strong> L4 management systems issued specific commands (e.g., “Set PRB threshold = 80%”).</li><li><strong>Present:</strong> Since the CNF is now an <strong>“Intelligent Agent”</strong>, it does not accept micro-management. The management system simply issues an <strong>“Intent”</strong> (e.g., “Ensure VIP user experience”).</li><li><strong>AN L4 Transformation:</strong> The core function of the management system becomes <strong>“Intent Translation”</strong> — transforming business language into logic understandable by the Agent <em>(hence Huawei’s push for logic-driven architecture).</em></li></ul><h4>3. Challenges in Governance and Observability</h4><p>This is the change that gives operators the greatest headache.</p><ul><li><strong>Black-box Risk:</strong> If AN L4 decision logic is hidden inside Huawei’s CNF, how can an operator ensure the Agent isn’t “lying”? Or that it hasn’t sacrificed safety in pursuit of performance?</li><li><strong>New Tasks for AN L4:</strong></li></ul><p><strong>Audit:</strong> Monitoring the Agent’s decision-making process.</p><p><strong>Guardrails:</strong> This is the GGF mentioned in Jio’s proposal. Even if L4 capabilities are pushed down, the upper layer must have an independent physical tier to intercept unsafe Agent behaviour in real-time.</p><p><strong>Conflict Management:</strong> When internal Agent 1 (Energy Saving) and Agent 2 (Performance) within a CNF clash, the upper L4 must have a global perspective to set the final tone.</p><p><strong>Summary of Decentralisation:</strong></p><p>Huawei’s proposal is effectively driving the decentralisation of AN L4. Shifting from an independent software platform (OSS/EMS) to a distributed, embedded model, processing speeds will increase from minutes/seconds to <strong>milliseconds</strong> (due to being inside the element). Traditional L4 work, such as rule configuration and command orchestration, will change; instead, Agentic AN L4 within the CNF will handle intent arbitration, knowledge distribution, and compliance monitoring — functions previously held by the external AN L4.</p><h4>Risks</h4><p>Similar to TR 38.843’s risks, <strong><em>if AN L4 decision intelligence is “dissolved” within the respective CNFs </em></strong>of Huawei, Nokia, or Ericsson, <strong><em>the traditional “parameter-delivery” interoperability model will indeed collapse</em></strong>. Based on current TMF AN specifications and 3GPP SA2 evolution trends, <em>to achieve multi-vendor interoperability, the industry must undergo a paradigm shift from </em><strong><em>“Command Interoperability”</em></strong><em> to </em><strong><em>“Semantic Interoperability”</em></strong><em>.</em></p><h3>Possible Implementation Methods (Based on Current Specifications)</h3><h4>1. Standardised Intent Interfaces (Intent-driven API / TMF921)</h4><p>Currently the most mainstream solution. Since the internal workings of the network element are a black box, we ignore the internal calculations and unify only the “Requirement Description”.</p><ul><li><strong>Implementation:</strong> Regardless of the underlying CNF vendor, the upper management system (OSS) sends intents via the standard TMF921 Intent API.</li><li><strong>Interoperability Point:</strong> As long as Vendor A’s element understands the intent “Latency &lt; 10ms” and reports “Achieved”, interop is successful.</li><li><strong>Bottleneck:</strong> <strong>Depth and granularity of intent.</strong> If Huawei’s Agent supports “ISAC Intent” but Nokia’s only supports “Traffic Intent”, the management system will face severe “functional inconsistency”.</li></ul><h4>2. Unification of Knowledge Models and Ontologies (The RDF/OWL Path)</h4><p>If AI is built into elements, AI from different vendors must share a “common sense”.</p><ul><li><strong>Implementation:</strong> 3GPP standardises a universal Information Model and Ontology.</li><li><strong>Interoperability Point:</strong> Huawei’s and Nokia’s Agents both follow the same RDF framework. When one Agent says “Network Congestion”, the other understands exactly what this implies via OWL logic, enabling coordination.</li><li><strong>Status:</strong> <strong>Extremely difficult to achieve.</strong> Vendors treat their model training data and logic as core Intellectual Property (IP), making unified semantic transparency nearly impossible.</li></ul><h4>3. Mandatory Guardrail Mediation (Guardrail Mediation)</h4><p>Referencing Jio’s proposal: Since I cannot control your internal decisions, I will set up a <strong>“Security Checkpoint”</strong> at the interface.</p><ul><li><strong>Implementation:</strong> Define a standardised <strong>Governance Function</strong>.</li><li><strong>Interoperability Point:</strong> Any operational request from an Agent must be converted into a standard “Operational Primitive” and submitted to the checkpoint. The checkpoint audits these against unified policies defined by the operator.</li><li><strong>Conclusion:</strong> This approach does not seek to “understand” the peer, but merely to <strong>“constrain”</strong> them. This is likely the most realistic commercial compromise for multi-vendor environments.</li></ul><h4>4. The Possibility of Failure: Towards “Full-stack Silos”</h4><p>If the above standardisation fails, we face a dire situation: <strong>Full-stack Lock-in</strong>.</p><ul><li><strong>Vertical Monopoly:</strong> If Huawei’s L4 functions are too deeply coupled with its CNF, an operator might find they cannot control Huawei’s Agents using a Nokia management system.</li><li><strong>Consequence:</strong> To ensure the efficiency of the AN L4 closed-loop, operators would be forced to use a single vendor’s products in specific geographic regions or business domains. <strong><em>The “Decoupling” dream pursued by Open-RAN would be re-locked by “Intelligence”.</em></strong></li></ul><h4>Deep Insight</h4><p><strong>Current AN L4 specifications (like those published by TMF) are actually very fragile</strong><em>(I said many times, AN L4 will be a dream 10–15 years later, and those L4 show-case actually are for show purpose, will never be massive deployed in production environment)</em><strong>. </strong>Most are stuck at the <strong><em>“what the interface looks like” </em></strong>stage, without touching the core of<em> </em><strong><em>“how Agents interact and negotiate”</em></strong><em>. (I lack membership access; if anyone can share TMF L4 operator cases documents, it would help clarify the current maturity of those cases, nearly nothing can be hidden from my eyes).</em></p><p>If 6G truly moves towards Huawei’s proposed Agentic architecture, we need something akin to an<em> </em><strong><em>“Agent Social Protocol”</em></strong><em>.</em> Much like people of different nationalities communicating, they do not need to exchange DNA (internal algorithms), but they must <strong><em>collaborate through a common language </em></strong>(English/Standard Intent) and <strong><em>a shared social contract </em></strong>(Governance Guardrails).</p><p><strong><em>Poor operators once again face a dilemma:</em></strong> <em>pursue the ultimate performance of L4 and tolerate “Vendor Lock-in”, or lower the network’s IQ for the sake of “Multi-vendor Flexibility”?</em> However, I have always believed that only the operators’ own technical capabilities and their understanding of their own networks can truly identify what they need.</p><h3>4. Reliance Jio’s Proposal (INCF/GGF/AIAF)</h3><p>We previously touched upon some of Jio’s requirements. The Reliance Jio proposal feels more closely aligned with <strong>AN L4</strong> (Autonomous Networks Level 4); indeed, one might argue they should lead the discussions at the TMF (TeleManagement Forum). Their focus has shifted away from the debate over “whether to use AI, where to place it, or how to write the algorithms” towards <strong>“how to cage AI within a framework of rules”</strong>.</p><p><em>If Huawei’s proposal is about </em><strong><em>giving the network a “brain</em></strong><em>”, Ericsson’s is about </em><strong><em>“extending the back garden”,</em></strong><em> and NVIDIA’s is a </em><strong><em>“complete demolition and rebuild”</em></strong><em>,</em> then Jio’s proposal is about establishing a <strong><em>“judicial system”</em></strong> for the network. In a complex multi-vendor environment, this is often the final piece of the puzzle that gives operators the confidence to actually deploy AN L4.</p><h4>1. Jio’s Proposal: A “Battle-Hardened” Version of AN L4</h4><p>The current TMF AN L4 specifications define “objectives” and “levels”, but remain relatively thin on<strong><em> “how to ensure safety” and “how to conduct cross-domain auditing”. </em></strong>The three core functional modules proposed by Jio effectively fill the void in the commercial implementation of L4:</p><ul><li><strong>INCF (Intent Network Control Function): The “Translator” of L4.</strong></li></ul><p>It translates vague business intents into executable network policies and manages intent mapping across different vendors. This directly bolsters the intent-processing capabilities of L4.</p><ul><li><strong>GGF (Governance and Guardrail Function): The “Security Guard”.</strong></li></ul><p>This is the most critical component. It performs <strong>Pre-execution Validation</strong> before any AI decision is implemented.</p><p><em>Scenario</em>: If a Huawei Agent suggests shutting down a specific base station to save energy, the GGF will intercept this based on the operator’s “absolute rules” (e.g., “This base station is near a hospital and must never be deactivated”). This compensates for the lack of deterministic control in current L4 specifications.</p><ul><li><strong>AIAF (AI Audit Function): The “Auditor”.</strong></li></ul><p>It records the “thought process” of the AI in the background. Should something go wrong in the network, it is used to “settle accounts”. This addresses the most practical issues of compliance and accountability in L4 deployment.</p><h4>2. Deconstruction of the Strategy</h4><p><strong>A. High Compatibility with AN L4</strong></p><p>Jio’s proposal is essentially <strong>tailor-made for a multi-vendor L4 environment</strong>. In such a network, it is impossible to unify the AI logic of every vendor. Jio provides a “greatest common denominator” architecture: regardless of whose network elements are at the bottom, as long as they connect to a unified GGF (Guardrail) and AIAF (Audit), the operator can achieve “secure and controllable autonomy” at the L4 level.</p><p><strong>B. A “Dimensional Reduction” Enhancement of Existing Specifications</strong></p><p>Existing AN specifications discuss “interfaces” and “models” in broader terms, whilst Jio introduces the concepts of <strong>“Sandboxing”</strong> and <strong>“Real-time Auditing”</strong>.</p><ul><li>This takes AN from “laboratory automation” to a <strong>“regulated production environment”</strong>.</li><li>It transforms L4 from a “technical indicator” into a <strong>“manageable business process”</strong>.</li></ul><h4>Summary: The Entry Ticket for Large-Scale Commercialisation</h4><p>From the perspective of an operations engineer (OSS) or an operator, the primary concern is not whether the AI is clever enough, but rather <strong>“what happens if it goes rogue?”</strong> or <strong>“what if it misunderstands the intent?”</strong>.</p><p>Jio’s proposal resolves the question of <strong>“whether one dares to use it”</strong> through the <strong>GGF (Guardrails)</strong> and the question of <strong>“who takes the blame”</strong> through the <strong>AIAF (Audit).</strong> This aligns far more closely with the commercial logic of L4 than a mere debate over the precision of AI algorithms.</p><h3>5. <strong>Lenovo</strong></h3><p>Actually they should not come to waste time and $ on it.</p><p>Apr.28,2026</p><h3>Jimmy CHOW</h3><blockquote><strong>Professional Bio / Collaboration Offer</strong></blockquote><blockquote><strong>Strategic Network Transformation &amp; Integration Expert</strong></blockquote><blockquote>I am currently open to new professional opportunities and consulting engagements, yes, just want to do sth., otherwise I will do bricklayer works.</blockquote><blockquote>If your organisation is navigating the complexities of <strong>multi-generation network integration</strong> (2G/3G/4G/5G/6G) for a future-proof <strong>cloud-native core network</strong> or managing the intricate network architectures and fragmented service platforms that arise from <strong>mergers and acquisitions (M&amp;A)</strong> aiming for an<strong> Autonomous Network Architecture</strong>, I offer the expertise to streamline your operations.</blockquote><blockquote>My focus is on driving operational efficiency and fostering competitive advantage by aligning current infrastructures with future-ready network evolution. Whether working independently or embedded within your existing team, I provide end-to-end support — from <strong>situational analysis and strategic roadmap design to hands-on implementation and seamless transition</strong>.</blockquote><blockquote>Let’s collaborate to transform your legacy complexities into a lean, agile, and future-proof digital foundation.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bcf698e59de4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[BT Sovereign AI: The Rise of UK Sovereign Cloud and AI — Can It Carve a New Path Amid Hyperscaler…]]></title>
            <link>https://medium.com/@mountdavistechnologies/bt-sovereign-ai-the-rise-of-uk-sovereign-cloud-and-ai-can-it-carve-a-new-path-amid-hyperscaler-720d109e6ecf?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/720d109e6ecf</guid>
            <category><![CDATA[hyperscaler]]></category>
            <category><![CDATA[private-cloud]]></category>
            <category><![CDATA[ai-sovereignty]]></category>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Sat, 25 Apr 2026 10:30:31 GMT</pubDate>
            <atom:updated>2026-04-25T10:30:31.241Z</atom:updated>
            <content:encoded><![CDATA[<h3>BT Sovereign AI: The Rise of UK Sovereign Cloud and AI — Can It Carve a New Path Amid Hyperscaler Dominance?</h3><p>Jimmy CHOW</p><p>On 23 April 2026, British telecommunications giant BT announced a partnership with Nscale to deploy up to 14MW of NVIDIA full-stack AI infrastructure across existing strategic sites in the UK <a href="https://newsroom.bt.com/bt-to-work-with-nscale-to-accelerate-uk-sovereign-ai-capability-powered-by-nvidia/"><em>(click link)</em></a>, while simultaneously launching its <strong>Sovereign Cloud</strong><em> (private cloud)</em> platform. This marks a significant step forward in the UK’s <strong>Digital Sovereignty</strong> strategy. BT has become the first UK operator to offer a complete sovereign services suite (<strong>connectivity + voice + cloud + AI</strong>), aimed at providing <strong>public sector organisations, NHS healthcare, financial institutions, and critical infrastructure with a secure environment where data remains within the UK.</strong></p><p>In an era of geopolitical tension and increasingly strict data localisation regulations, BT’s move is not an isolated case but <strong>a typical example of European national operators responding to the dominance of US hyperscalers.</strong> Based on the latest publicly available information, this article analyses the details of BT Sovereign AI, provides a detailed comparison with the sovereign offerings from<strong> AWS, Azure, and Google, </strong>examines its strategic impact on the UK market, and assesses the market opportunities and risks.</p><h3>1. Details of BT Sovereign AI and Sovereign Cloud</h3><p>The core principle of BT’s Sovereign platform is “<strong>UK-hosted, UK-operated, UK-cleared</strong>” — meaning all data, access rights, operational management, and legal responsibility remain entirely within the UK, managed exclusively by British-based teams who have undergone security clearance.</p><ul><li><strong>Sovereign Cloud</strong>: This is a private/managed cloud platform. The underlying infrastructure is provided by <a href="https://www.datacenterdynamics.com/en/news/nscale-partners-with-bt-for-more-uk-data-centers/"><strong>Rackspace Technology</strong>’s UK data centres <em>(Rackspace Tech. is an USA company too),</em> delivering compute, storage, and backup capabilities. </a>Rackspace handles hosting, migration, day-to-day operations, and ongoing compliance support. The platform emphasises seamless hybrid cloud integration with legacy systems and hyperscalers, and eliminates egress fees, making costs more predictable. It primarily targets organisations handling sensitive or regulated workloads, such as NHS clinical data, financial transaction systems, and government platforms.</li><li><strong>Sovereign AI</strong>: In deep partnership with<a href="https://newsroom.bt.com/bt-to-work-with-nscale-to-accelerate-uk-sovereign-ai-capability-powered-by-nvidia/"> Nscale (a UK AI cloud company founded in 2024 and backed by NVIDIA), </a>BT is building up to <strong>14MW</strong> of AI data centre capacity at three of its existing strategic sites using NVIDIA full-stack AI infrastructure (including high-density GPUs). Nscale is responsible for modular deployment, while BT provides power, nationwide fibre backbone connectivity, and system integration. This enables customers to run AI training, inference, and large language models securely within the UK without sending data abroad.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AanDvfPRlV70dl-l" /><figcaption>Picture Credit: <a href="https://content.presspage.com/uploads/2429/426a5c59-77a3-4aa4-9a63-ada614c4fe4e/1920_ai-artificial-intelligence-block.jpg?10000">BT</a></figcaption></figure><p>BT has also introduced Sovereign Connectivity and Sovereign Voice, creating a “one-stop” sovereign services portfolio. According to BT research, sovereign-related AI opportunities could <a href="https://newsroom.bt.com/bt-becomes-first-uk-provider-to-offer-full-suite-of-sovereign-services--as-research-unveils-18bn-sovereign-ai-opportunity-for-uk/">unlock up to <strong>£18 billion</strong> in annual productivity </a>value for the UK economy.</p><p>BT’s approach is a <strong>pragmatic hybrid model</strong>: leveraging international advanced technologies (<em>NVIDIA and Rackspace, both from USA</em>) while maintaining firm control over local operations to rapidly address the UK’s urgent demand for sovereign AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sBIMutkKhuTzOivlq_pHBQ.png" /></figure><h3>2. Detailed Comparison with AWS, Azure, and Google Sovereign Cloud/AI Offerings</h3><p>BT’s Sovereign solution emphasises <strong>“pure UK control”, </strong>whereas the hyperscalers’ European or localised sovereign offerings provide richer ecosystems and greater scale, though their sovereignty is often criticised as <strong>“sovereignty washing”.</strong> The following sections offer a detailed comparison across multiple dimensions:</p><h4><strong>Ownership and Legal/Operational Jurisdiction (The Most Critical Difference)</strong></h4><p>BT Sovereign Cloud is a <strong>purely British</strong> solution with no US parent company involvement. Data and operational responsibility fall entirely under UK law and are not subject to foreign enforcement risks such as the US CLOUD Act.</p><p>In comparison:</p><ul><li><strong>AWS European Sovereign Cloud</strong> (generally available in Brandenburg, Germany, since January 2026): Physically and logically isolated from global AWS regions, operated by an EU entity, but with ultimate ownership by US-based Amazon, leaving potential exposure to US law.</li><li><strong>Azure Sovereign Cloud</strong>: Offers Sovereign Public Cloud (enhanced controls on existing European regions) and Sovereign Private Cloud (Azure Local, supporting air-gapped disconnected mode). While it stresses EU data boundaries and local governance, the ultimate parent remains US-based Microsoft.</li><li><strong>Google Cloud Sovereign Offerings</strong> (including Data Boundary, Dedicated, and Distributed Cloud GDC air-gapped): Can be operated through local partners for greater independence; in air-gapped mode, Google cannot access the environment remotely, yet the software stack remains from US-based Google.</li></ul><p><strong>BT’s Advantage</strong>: It offers genuine reassurance for UK public sector bodies and highly regulated industries that require zero foreign influence.</p><p><strong>Hyperscalers’ Advantage</strong>: They significantly mitigate legal risks through EU operating entities, data boundaries, and local partner models, although they cannot fully eliminate concerns over parent company jurisdiction.</p><h4><strong>Infrastructure and Deployment Models</strong></h4><p>BT uses a <strong>private/managed cloud</strong> model, relying on Rackspace UK data centres to deliver compute, storage, backup, and full managed services. It has a lower migration barrier and is well suited to hybrid cloud environments.</p><ul><li><strong>AWS European Sovereign Cloud</strong>: An independent, full-featured AWS partition offering near-identical services, APIs, and SLAs to global AWS, with multi-Availability Zone design for greater scale and resilience.</li><li><strong>Azure Sovereign Cloud</strong>: Sovereign Public Cloud enables rapid deployment on existing Azure regions; Private Cloud supports hybrid, on-premises, or fully disconnected deployments with high flexibility.</li><li><strong>Google Distributed Cloud (GDC)</strong>: Can be deployed in customer facilities, supporting both connected and fully air-gapped configurations based on Kubernetes, ideal for edge and highly isolated scenarios.</li></ul><p><strong>BT’s Advantage</strong>: Transparent pricing (no egress fees) and strong integration with existing systems. <strong>Hyperscalers’ Advantage</strong>: Globally consistent ecosystems and larger-scale infrastructure, better suited to organisations seeking standardisation.</p><h4><strong>AI and High-Performance Computing</strong></h4><p>BT Sovereign AI currently plans 14MW of NVIDIA full-stack capacity with modular deployment, specifically designed for sensitive UK-based AI workloads.</p><ul><li><strong>AWS</strong>: Inherits mature AI services such as Bedrock and SageMaker, with far greater scale than BT and strong alignment with global AWS innovation.</li><li><strong>Azure</strong>: Through Azure Local and Foundry Local, it supports running large multimodal AI models even in fully disconnected environments, with richer governance and productivity tools.</li><li><strong>Google</strong>: GDC air-gapped mode brings advanced AI capabilities such as Gemini and Vertex AI to local environments. It has already been deployed in high-security scenarios including NATO and UK MoD contracts, demonstrating leading AI maturity.</li></ul><p><strong>BT’s Advantage</strong>: Strong focus on local control and certainty that data remains in the UK. <strong>Hyperscalers’ Advantage</strong>: More complete AI toolchains, larger compute scale, and faster innovation cycles.</p><h4><strong>Use Cases, Compliance, and Cost</strong></h4><p>BT is particularly well suited to scenarios with rigid requirements for “data not leaving the UK”, such as NHS, financial institutions, government, and utilities, offering native alignment with UK regulations and local support.</p><p>Hyperscaler solutions cover a broader range of regulated industries across Europe (healthcare, finance, defence) with rich compliance certifications, but usually carry a sovereignty premium. BT offers more predictable costs (no egress fees), while hyperscalers provide greater elasticity for large-scale deployments. Many organisations ultimately adopt a <strong>hybrid strategy</strong>: placing the most sensitive workloads with BT and sovereign AI or general innovation workloads with hyperscalers.</p><h4><strong>Other Practical Differences</strong></h4><p>BT offers relatively smooth migration and relies on partners for its ecosystem; hyperscalers (particularly Azure and Google) have highly mature and consistent ecosystems, although some advanced services may require additional configuration in sovereign modes. BT’s solution is still in its early deployment phase, whereas AWS ESC, Azure Sovereign, and Google GDC already have more real-world implementations (such as Google’s defence contracts).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/922/1*B1jItNCuSeUaU5Dc9dpI5Q.png" /><figcaption>Table: <strong>Comparison Summary Table for better visual reference</strong></figcaption></figure><h3>3. Strategic Impact of BT’s Sovereign AI on the UK Market</h3><p>BT’s initiative is having a multifaceted impact on the UK market, accelerating domestic digital sovereignty while reshaping the competitive landscape.</p><ul><li><strong>Impact on Hyperscalers</strong>: BT is likely to divert some high-compliance, high-sovereignty workloads, particularly sensitive AI projects in the public sector and regulated industries. This will erode AWS, Azure, and Google’s share in these UK niche segments and force them to accelerate localisation efforts (for example, AWS expanding EU operating entities, Azure strengthening air-gapped capabilities, and Google deepening GDC deployments). However, hyperscalers will not be displaced — they will continue to dominate general-purpose cloud and global innovation thanks to their ecosystems and mature tools. BT will largely play a complementary rather than substitutive role.</li><li><strong>Impact on Other UK Operators</strong>: Leveraging its nationwide network advantage and existing sites, BT has quickly positioned itself as “the UK’s first complete sovereign suite provider”. Operators such as Virgin Media O2 and Vodafone/Three currently focus more on connectivity and infrastructure sharing, with relatively limited deep investment in sovereign cloud and AI. BT’s success may encourage them to pursue similar partnerships or accelerate their own efforts, but it could also widen industry polarisation — large operators transforming into “digital infrastructure providers” while smaller ones risk marginalisation. Overall, this boosts confidence in the UK’s domestic cloud ecosystem and drives telecom operators to evolve from traditional “pipes” to higher-value services.</li><li><strong>Broader Market Impact</strong>: BT’s strategy aligns with the UK Government’s push for domestic AI capability and digital sovereignty, making it well positioned to win public sector contracts. It also demonstrates a successful “telco + partner ecosystem” model: BT provides the network and trust, Rackspace supplies managed services, and Nscale + NVIDIA deliver AI compute. This offers UK businesses and public organisations more trustworthy local alternatives and reduces over-reliance on overseas cloud providers.</li></ul><h3>4. Market Opportunities and Risk Assessment</h3><p><strong>Opportunities</strong>:</p><ul><li>Significant economic potential: BT research indicates that digital sovereignty could unlock £18 billion in annual AI productivity value and drive rapid growth in the sovereign cloud market.</li><li>Hybrid cloud benefits: Customers can flexibly combine local sovereign solutions with global hyperscalers to create complementary ecosystems.</li><li><strong>Policy and trust advantages</strong>: As a domestic player, BT is well placed to secure government support, while also driving job creation, data centre investment, and development of the local AI supply chain.</li><li>Global trend tailwinds: The European sovereign cloud market is experiencing strong growth, and BT’s UK implementation could serve as a regional benchmark.</li></ul><p><strong>Risks</strong>:</p><ul><li><strong>Scale and maturity constraints:</strong> The 14MW AI capacity is relatively small compared with hyperscalers’ single-region deployments. The ecosystem depends on external partners, innovation speed may lag, and real-world deployment cases still require time to validate.</li><li><strong>Capital and technology dependencies</strong>: Although using existing sites reduces upfront costs, high-power AI facilities pose new challenges for electricity supply, cooling, and supply chains. Non-UK components (NVIDIA, Rackspace) are pragmatic but require ongoing management of geopolitical risks.</li><li>Intensifying competition: Hyperscalers will strengthen their localisation countermeasures; other operators or new entrants may fragment the market. Pricing premiums and migration friction could also slow adoption.</li><li><strong>Regulatory and external uncertainties: </strong>Policy changes or escalating international tensions could amplify risks. If “sovereignty” is perceived by the market as more marketing than substance, customer trust could be damaged.</li></ul><p>Overall, the risks appear manageable, but BT must accelerate proof-of-concept validation, continue expanding capacity, and deepen ecosystem partnerships to convert its first-mover advantage into long-term market leadership.</p><h3>Conclusion: A Pragmatic Path for UK Digital Sovereignty</h3><p>BT Sovereign AI is not an idealistic pursuit of 100% domestically developed technology but a realistic approach that prioritises control while embracing pragmatic international cooperation. In an era dominated by hyperscalers, it provides the UK with genuinely trustworthy local options and is helping to steer the wider market toward hybrid, sovereignty-friendly models.</p><p>For decision-makers, the key is not choosing entirely between BT and hyperscalers, but building the right combination according to workload sensitivity: the most critical sensitive data and AI should reside with BT, while scale and cutting-edge innovation can leverage AWS, Azure, or Google.</p><p>The UK digital sovereignty wave is only just beginning. <strong>Whether BT’s initiative can truly unlock the £18 billion opportunity will depend on execution speed, continued policy support, and market adoption. Over the coming years, we will witness more collisions and fusions between domestic solutions and global hyperscalers — and this may well be the key to the UK maintaining its digital competitiveness in the AI era.</strong></p><p>Apr.25, 2026</p><h3>Jimmy CHOW</h3><blockquote><strong><em>Professional Bio / Collaboration Offer</em></strong></blockquote><blockquote><strong>Strategic Network Transformation &amp; Integration Expert</strong></blockquote><blockquote>I am currently open to new professional opportunities and consulting engagements, yes, just want to do sth., otherwise I will do bricklayer works.</blockquote><blockquote>If your organisation is navigating the complexities of <strong>multi-generation network integration</strong> (2G/3G/4G/5G/6G) for a future-proof <strong>cloud-native core network</strong> or managing the intricate network architectures and fragmented service platforms that arise from <strong>mergers and acquisitions (M&amp;A)</strong> aiming for an<strong> Autonomous Network Architecture</strong>, I offer the expertise to streamline your operations.</blockquote><blockquote>My focus is on driving operational efficiency and fostering competitive advantage by aligning current infrastructures with future-ready network evolution. Whether working independently or embedded within your existing team, I provide end-to-end support — from <strong>situational analysis and strategic roadmap design to hands-on implementation and seamless transition</strong>.</blockquote><blockquote>Let’s collaborate to transform your legacy complexities into a lean, agile, and future-proof digital foundation.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=720d109e6ecf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Telecom Cloud-Native ? or Cloud-Naïve ?]]></title>
            <link>https://medium.com/@mountdavistechnologies/telecom-cloud-native-or-cloud-na%C3%AFve-7c7ae2dc6a1d?source=rss-82c3f2323425------2</link>
            <guid isPermaLink="false">https://medium.com/p/7c7ae2dc6a1d</guid>
            <category><![CDATA[6g]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[5g]]></category>
            <category><![CDATA[cloud]]></category>
            <dc:creator><![CDATA[Mount Davis Technologies]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 09:44:11 GMT</pubDate>
            <atom:updated>2026-04-16T09:44:11.101Z</atom:updated>
            <content:encoded><![CDATA[<p>Jimmy CHOW</p><h3>Introduction</h3><p><a href="https://www.linkedin.com/in/tom-nolle-truth-teller/?lipi=urn%3Ali%3Apage%3Ad_flagship3_feed%3Be18dRwNTRmKLTffL2IyGnQ%3D%3D">TOM Nolle</a>’s latest analysis article <a href="https://www.linkedin.com/posts/tom-nolle-truth-teller_one-of-the-problems-with-hype-driven-markets-share-7449784901333135361-Nos7?utm_source=social_share_send&amp;utm_medium=member_desktop_web&amp;rcm=ACoAAAJT_98BapiFtvukcxVj6CgWsAszQxLr-Ms"><em>Why We Need to Rethink “Cloud-Native”</em></a><em> </em>is really good to read and rethink by peers in this telecom industry. The author’s key point is that <strong><em>“cloud-native” is seriously misunderstood in the telecom industry, and true cloud-native is not about breaking down old monolithic systems into microservices, but about fundamentally adopting an event-driven, state-modeled architecture.</em></strong></p><p>Tom believes that the problem with the concept of “cloud-native” is abused in this industry, many telcos call themselves cloud-native by breaking down legacy system into microservices, but this doesn’t bring elasticity, scalability, or agility….</p><p><strong>TOM’s true cloud-native principle:</strong> event-driven + state model. Event is the driving forces, sales, billing, resource failure, etc. should all be considered events. Everything is a model, customers, resources, bills, services, etc. should be modeled as a state machine. and Event+State = microservice trigger point, each state/event combination corresponds to a microservice that can be started on demand. So, what does this mean ? There will be no more traditional OSS/BSS/NMS, all are simply the result of the events flowing between different model, not a fixed application systme. and the system is as resilient as “breathing”, microservices are launched on demand, resource promise are scalable and contractible, truly achieving cloud resiliency. Finally, Tom thinks architecture determines capability.</p><p>I personally support TOM’s opinions. The cloud-native is misused in this industry for years, sometimes I thought unless he/she is an system architect engineer , R&amp;D engineer, or with this background in this domain, most of those with cloud-native in their talks actually not really understand what she/he is talking about. The cloud-native in the mind of people with IT background is completely different with those with CT background.</p><p>That’s why I said in about 2018/2019 that the widely commercial development of 5G NSA would be a business failure, most of them thought it was small issue, and many years later it proved to be true. The simple reason I said in many old articles was the 5G Core was not mature at all in that time. The cloud-native 5G core is still not matured even today, so, all the promises of the 5G nearly completely failed but with a different RAN only, and many CxOs also stepped down because of the wrong strategic investments &amp; poor ROI/ROA.</p><p><a href="https://www.linkedin.com/posts/jimmychow_after-the-hustle-and-bustle-only-reality-activity-7438345728051150849-toDQ?utm_source=social_share_send&amp;utm_medium=member_desktop_web&amp;rcm=ACoAAAJT_98BapiFtvukcxVj6CgWsAszQxLr-Ms">Revisiting 5G Value Proposition and Business Model | Jimmy CHOW posted on the topic | LinkedIn</a></p><p>So, what does cloud-native really mean ?</p><h3>The Cloud-Native &amp; Telecom Cloud-Native Definition</h3><p><a href="https://github.com/cncf/toc/blob/main/DEFINITION.md">The cloud-native’s definition by <strong>CNCF</strong>:</a></p><blockquote><strong>Cloud native</strong> practices empower organizations to develop, build, and deploy workloads in computing environments (public, private, hybrid cloud) to meet their organizational needs at<strong> scale in a programmatic and repeatable manner</strong>. It is characterized by<strong> loosely coupled systems</strong> that interoperate in a manner that is<strong> secure, resilient, manageable, sustainable, and observable.</strong></blockquote><blockquote><strong>Cloud native technologies </strong>and architectures typically consist of some combination of <strong>containers, service meshes, multi-tenancy, microservices, immutable infrastructure, serverless, and declarative APIs</strong> — this list is non-exhaustive.</blockquote><p>But actually in telecom industry, there is <strong>NO single, universally accepted “telecom cloud-native” definition,</strong> those SDOs (3gpp, etsi, gsma, o-ran, and hyperscalers, etc.) converges on a practical interpretation of it. so, telecom cloud‑native means designing network functions so that <strong>compute is stateless, state is externalized, scaling is horizontal, lifecycle management (LCM) is declarative, and operations are fully automated</strong>, while still meeting <strong>carrier‑grade performance, reliability, and latency</strong> requirements.</p><p>It is a <strong>CNCF cloud‑native with Telecom‑grade constraints</strong>. so, we can break it down to many features which match what need in the telecom networks:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iLYaG6AKFEWVyOsIiuKFmA.jpeg" /></figure><h4>1. Stateless Compute, Stateful Network — Cleanly Separated</h4><p>Telecom <strong>workloads </strong>(AMF, SMF, UPF, gNB‑CU, etc.) must run as <strong>stateless service instances</strong>, while all <strong>business state</strong> <em>(UE context, sessions, policies, counters) </em>is stored in outside DB, not in containers with instances as now do:</p><ul><li>Distributed data layers (Redis, Cassandra, etc.)</li><li>NF‑internal state services</li><li>3GPP-defined repositories (UDM, UDR, PCF)</li></ul><p>This enables:</p><ul><li>Fast failover</li><li>Rolling upgrades</li><li>Horizontal scaling</li></ul><h4>2. Event‑Driven Architecture + State Machines</h4><p>Telecom NFs are inherently <strong>stateful</strong> at the business level <em>(UE registration, PDU session, handover).</em> Cloud‑native telecom should solve this by:</p><ul><li>Using <strong>event-driven microservices</strong></li><li>Implementing <strong>explicit state machines</strong></li><li>Persisting <strong>state externally</strong></li></ul><p>This is why AMF/SMF/UPF cloud‑native design always includes <strong>event buses</strong>, <strong>state stores</strong>, and <strong>idempotent handlers</strong>, they all should be decoupled completely.</p><h4>3. Horizontal Scalability</h4><p>Telecom cloud‑native requires:</p><ul><li>Scale-out, not scale-up</li><li>Per‑NF autoscaling based on load<em> (UE count, PDU sessions, throughput,etc.)</em></li><li>Independent scaling of control plane and user plane</li></ul><p>UPF is the strongest example: multiple UPF instances at edge, regional, and core.</p><h4>4. Immutable Infrastructure + Declarative Lifecycle Management</h4><p>Telecom CNFs must be deployable via:</p><ul><li>Helm charts</li><li>GitOps</li><li>Declarative APIs</li><li>Automated CI/CD pipelines</li></ul><p>No SSH, no CLI, no SNMP, no manual config, no mutable VMs.</p><h4>5. Carrier‑Grade Reliability (Five‑9s)</h4><p>Telecom cloud‑native adds strict requirements:</p><ul><li>Hitless upgrades</li><li>Zero‑packet‑loss failover (UPF)</li><li>Deterministic latency</li><li>Multi‑site redundancy</li><li>Geo‑redundant state replication</li></ul><p><strong>This is where telecom differs sharply from enterprise cloud‑native.</strong></p><h4>6. High‑Performance Data Plane</h4><p>Telecom cloud‑native must support:</p><ul><li>DPDK</li><li>SR‑IOV</li><li>eBPF</li><li>NUMA‑aware scheduling</li><li>CPU pinning</li></ul><p>Because UPF and O‑DU/O‑CU workloads require line‑rate packet processing.</p><h4>7. Multi‑Site, Edge‑Aware Deployment</h4><p>Telecom cloud‑native assumes:</p><ul><li>Thousands of distributed edge sites</li><li>Heterogeneous hardware</li><li>Constrained environments (far edge)</li><li>Centralized management</li></ul><p>This is not typical in enterprise cloud‑native.</p><h4>8. Standardized APIs &amp; Interoperability</h4><p>3GPP SBI (HTTP/2 + JSON) is the backbone. Telecom cloud‑native requires:</p><ul><li>API versioning</li><li>API backward compatibility</li><li>Service discovery</li><li>Observability (metrics, logs, traces), all CNF should support telemetry.</li></ul><p>So, put all of them together:</p><p><strong>Telecom cloud‑native</strong> refers to designing and operating network functions as <strong>stateless, horizontally scalable, event‑driven microservices</strong> running on <strong>declarative, automated, Kubernetes‑based infrastructure</strong>, with <strong>externalized state</strong>, <strong>carrier‑grade reliability</strong>, <strong>high‑performance data planes</strong>, and <strong>multi‑site edge deployment</strong>, while conforming to <strong>3GPP service‑based architecture </strong>(SBA).</p><h3>The Big difference between CNCF &amp; Telecom Cloud-Native</h3><h4>1. Life’s Purpose</h4><p><strong>CNCF cloud‑native</strong> — Built for <em>general enterprise workloads</em>: web apps, APIs, data services, SaaS.</p><p><strong>Telecom cloud‑native</strong> — Built for <em>5G/6G network functions</em>: AMF, SMF, UPF, gNB‑CU/DU, O‑RAN, IMS.</p><p>So, telecom cloud‑native must meet <strong>carrier‑grade SLAs</strong>, not just enterprise reliability. One can restart in every hour, but one need be reliable with 5nines.</p><h4>2. State Management</h4><p><strong>CNCF: </strong>Encourages stateless services. State externalized to databases, caches, object stores</p><p><strong>Telecom:</strong></p><ul><li><strong><em>Business logic is inherently stateful</em></strong> (UE context, PDU sessions, mobility, timers, etc.)</li><li><strong>Compute must be stateless</strong>, but <strong>state machines + externalized state</strong> are mandatory</li><li>Requires <strong>ultra‑fast, distributed state stores</strong></li></ul><p>Telecom NFs must survive failover <em>without losing subscriber sessions</em>.</p><h4>3. Scalability Model</h4><p><strong>CNCF: </strong>Horizontal scaling based on CPU/memory or custom metrics <strong>Telecom:</strong></p><ul><li>Scaling based on <strong>UE count, PDU sessions, throughput, QCI/5QI mix</strong></li><li>Control plane and user plane scale independently</li><li>UPF requires <strong>multi‑site scale‑out</strong> (edge, regional, core)</li></ul><p><strong>Telecom scaling is domain‑specific and far more complex.</strong></p><h4>4. Performance Requirements</h4><p><strong>CNCF: </strong>Optimized for throughput and elasticity</p><p><strong>Telecom:</strong></p><ul><li>Requires <strong>line‑rate packet processing</strong></li><li>Uses DPDK, SR‑IOV, eBPF, CPU pinning, NUMA awareness</li><li>Must meet deterministic latency (sub‑millisecond in some cases)</li></ul><p><strong>UPF performance is orders of magnitude beyond typical CNCF workloads.</strong></p><h4>5. Reliability &amp; Availability</h4><p><strong>CNCF: </strong>High availability, but graceful degradation is acceptable</p><p><strong>Telecom:</strong></p><ul><li><strong>Five‑9s availability</strong></li><li>Hitless upgrades</li><li>Zero‑packet‑loss failover</li><li>Geo‑redundancy across thousands of sites</li></ul><p><strong>Telecom cannot tolerate packet loss or service interruption.</strong></p><h4>6. Architecture Patterns</h4><p><strong>CNCF:</strong></p><ul><li>Microservices</li><li>Service mesh</li><li>Declarative APIs</li><li>GitOps</li></ul><p><strong>Telecom:</strong></p><ul><li>All CNCF patterns <strong>plus</strong>:</li><li>Event‑driven state machines</li><li>3GPP Service‑Based Architecture (SBI)</li><li>Multi‑cluster, multi‑site topology</li><li>Edge‑aware orchestration</li></ul><p><strong>Telecom adds domain‑specific architectural constraints.</strong></p><h4>7. Deployment Environment</h4><p><strong>CNCF: </strong>Public cloud, private cloud, hybrid</p><p><strong>Telecom:</strong></p><ul><li>Central cloud + regional cloud + thousands of MEC/edge sites</li><li>Heterogeneous hardware</li><li>Constrained far‑edge environments</li></ul><p><strong>Telecom cloud‑native is inherently distributed at massive scale.</strong></p><h4>8. Lifecycle &amp; Operations</h4><p><strong>CNCF: </strong>CI/CD, GitOps, observability, automation</p><p><strong>Telecom:</strong></p><ul><li>All CNCF practices <strong>plus</strong>:</li><li>Closed‑loop automation (self‑healing, self‑optimization)</li><li>Intent‑based orchestration</li><li>SLA‑driven scaling</li><li>Telecom‑grade alarms, KPIs, counters</li></ul><p><strong>Telecom operations are more automated and SLA‑driven.</strong></p><h4>Summary: The Core Difference</h4><blockquote><strong>CNCF cloud‑native</strong> is about agility, automation, and scalability.</blockquote><blockquote><strong>Telecom cloud‑native</strong> is about agility + deterministic performance + carrier‑grade reliability + massive edge distribution.</blockquote><p><strong>Telecom cloud‑native = CNCF cloud‑native plus telecom‑specific constraints.</strong></p><h3>The most confused items: STATEful &amp; STATELESS</h3><p>I think most of the peers in CT domains actually don’t really understand them well.</p><h4>1. Cloud-native “stateless” = compute node stateless (infrastructure level)</h4><p>The statelessness emphasized by cloud native refers to:</p><ul><li>Microservice instances can be destroyed and rebuilt at any time</li><li>Does not depend on local disks</li><li>Do not save business state in containers</li><li>You don’t need to migrate your local state when you scale out</li></ul><p>That is: <strong>the stateless is the “Service instance”, not the Business itself.</strong></p><p>The elasticity of Kubernetes, Serverless, and container orchestration all rely on this “stateless instance”.</p><h4>2. “State Model” of Telecom Service = State of Business Logic (Domain Level)</h4><p>Telecom services are inherently strong:</p><ul><li>subscription states(active / suspended / terminated）</li><li>resource states(allocated / reserved / released）</li><li>network event states(fault / recovery）</li><li>charging states(rated / billed / invoiced）</li></ul><p>These states MUST exist or the business cannot function.</p><p>The “state model” emphasized by TOM refers to: <strong>Business objects (customers, resources, services) should be modeled as state machines, event-driven state changes.</strong></p><p>This belongs to the <strong>business architecture layer</strong>, not the runtime infrastructure layer.</p><h4>🔗 Why don’t the two conflict?</h4><p>Because the right thing to do with cloud native is:</p><ul><li><strong>The service instance is stateless</strong></li><li><strong>business states is externalized state</strong></li><li><strong>Event sourcing</strong></li></ul><p>This is completely consistent with the TOM’s opinion.</p><p>In fact, the <strong>event-driven architecture (EDA) + state machine model </strong>is typical of cloud-native:</p><ul><li>Kafka / Pulsar：event streams</li><li>Redis/etcd/DB: External state store</li><li>Microservices: Decide the next action based on the event + current state</li></ul><p>This is exactly <strong>the combination of “stateless computing + stateful business”</strong> in telecom network.</p><h3>5G Core is Cloud-native ready ?</h3><p>The architecture of 5G SA itself is <strong>closer </strong>to cloud-native than 4G, but to be “truly cloud-native”, it needs to be reconstructed from four levels: <strong>architecture, state management, interface mode, and deployment method</strong>.</p><p>What now called/claimed cloud-native by some vendors actually are <strong>containerized network function (CNF)</strong> but <strong>NOT</strong> the required <strong>Cloud-Native Function (CNF)</strong>. To make them conform the core characteristics of cloud native, still need to do: <strong>(1) The control plane/user plane function can be scaled horizontally</strong> <strong>(2) the instance is stateless (or weak)</strong> <strong>(3) stateful external + event-driven. </strong>Nothing reached by the Containerized Network Function (CNF) now.</p><p>Even in the 3GPP Spec. of 5G CN, the AMF, UPF, SMF etc. are still based on interface, and CP/UP also coupled tightly.<strong> so, be fooled ? I said many years ago be not so aggressive, just wait the core be mature and then consider deployment 5G.</strong></p><h4>How to make AMF cloud-native ?</h4><p>🚀<strong>Outboard the UE Context (the most important).</strong></p><p>The largest state of AMF is the <strong>UE context (registration status, NAS security context, session information, etc.)</strong>. To be cloud-native, you must:</p><ul><li>AMF instances do not save UE state (stateless instance)</li><li>UE states stored outside in KV/DB(Redis、etcd、Cassandra)</li><li>Any AMF instance can handle the next message of the same UE</li></ul><p>This allows AMF to:</p><ul><li>Scale out</li><li>Instances are destroyed/rebuilt at any time</li><li>no session sticky</li></ul><p>🚀<strong>Event-driven AMF (replacing traditional “process machines”).</strong></p><p>The essence of AMF’s business is events:</p><ul><li>Registration Request</li><li>Service Request</li><li>Handover Required</li><li>Paging</li><li>Deregistration</li><li>….</li></ul><p>Each <strong>event triggers a state machine migration</strong> instead of the traditional “flowchart + large logic block”.</p><p>🚀<strong>AMF microservice (split by function).</strong></p><p>Instead of “blind dismantling microservices”, they are dismantled by event/state boundaries:</p><ul><li>NAS Decode/Encode</li><li>NGAP Handler</li><li>UE Context Manager</li><li>Mobility Manager</li><li>Timer Manager</li><li>SMF Interaction Service</li></ul><p>Each microservice is <strong>stateless </strong>and only handles events + updates external state.</p><p>🚀<strong>AMF’s cloud-native deployment capabilities</strong></p><ul><li>Kubernetes deployment</li><li>HPA automatically scales up</li><li><strong>Sidecar (optional) for logs, metrics, and link tracing (telemetry for AN L4 purpose)</strong></li><li>Grayscale upgrade, rolling upgrade</li><li>Multi-instance load balancing (SBI based on HTTP/2)</li></ul><h4>How to make UPF cloud‑native？</h4><p>UPF is the data plane, much more difficult than AMF, but the principle is the same:</p><p>🚀<strong>UPF session state is external or recoverable</strong></p><p>The status of the UPF includes:</p><ul><li>PDR/FAR/QER/URR</li><li>TEID mapping</li><li>QoS rules</li></ul><p>To becloud‑native：</p><ul><li>State Outer (KV/DB)</li><li>Or the state can be quickly rebuilt (SMF issues full rules)</li><li>UPF instances are stateless or weak</li></ul><p>This allows UPF to:</p><ul><li>Scale out</li><li>Instance failure recovery is automatic</li><li>No session stickiness is required</li></ul><p>🚀<strong>UPF data plane acceleration is compatible with cloud native</strong></p><p>UPF must be high-performance, but still cloud-native:</p><ul><li>DPDK + SR-IOV(high performance)</li><li>eBPF/XDP（more cloud-native）</li><li>SmartNIC/DPUs（no need GTP-U）</li><li>User and Control Plane Separation (CUPS)</li></ul><p>🚀<strong>UPF Distributed Deployment (Edge + Core).</strong></p><p>UPF is a natural distributed component:</p><ul><li>Edge UPF (Low Latency)</li><li>Core UPF (Billing, Policy)</li><li>Multi-UPF selection (SMF takes care of selection)</li></ul><p>Cloud-native’s UPF can:</p><ul><li>Automatically scale up and down by region</li><li>Automatic scheduling by flow</li><li>Instances are automatically assigned by service type (URLLC/eMBB).</li></ul><p>🚀<strong>UPF microservice (limited).</strong></p><p>UPF is not suitable for breaking down into dozens of microservices, but it can be broken into:</p><ul><li>CP（PFCP Handler）</li><li>DP（DPDK/eBPF Pipeline）</li><li>Session Manager</li><li>Usage Report Service (charging/reporting)</li></ul><p>The data plane maintains high performance and the control plane is cloud-native.</p><p>Finally, <strong>the RAN also need to the same thing,</strong> some works are done by the open-ran/o-ran initiative. but still not enough.</p><p>So, now, you know what’s the telecom cloud-native really means.</p><p>Apr.16, 2026</p><h3>Jimmy CHOW</h3><blockquote><strong>Professional Bio / Collaboration Offer</strong></blockquote><blockquote><strong>Strategic Network Transformation &amp; Integration Expert</strong></blockquote><blockquote>I am currently open to new professional opportunities and consulting engagements, yes, just want to do sth., otherwise I will do bricklayer works.</blockquote><blockquote>If your organisation is navigating the complexities of <strong>multi-generation network integration</strong> (2G/3G/4G/5G/6G) for a future-proof <strong>cloud-native core network</strong> or managing the intricate network architectures and fragmented service platforms that arise from <strong>mergers and acquisitions (M&amp;A)</strong> aiming for an<strong> Autonomous Network Architecture</strong>, I offer the expertise to streamline your operations.</blockquote><blockquote>My focus is on driving operational efficiency and fostering competitive advantage by aligning current infrastructures with future-ready network evolution. Whether working independently or embedded within your existing team, I provide end-to-end support — from <strong>situational analysis and strategic roadmap design to hands-on implementation and seamless transition</strong>.</blockquote><blockquote>Let’s collaborate to transform your legacy complexities into a lean, agile, and future-proof digital foundation.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7c7ae2dc6a1d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>