The Other Side of the AI Stack
A note from the author: With over 15 years of experience in supply chain and logistics, I have dedicated my career to building the foundational operational systems that drive efficiency and quality. During my tenure at Google, I focused on the data foundations of reverse logistics for consumer hardware — building the data infrastructure and analytics pipelines that bridged the gap between product operations and product quality. By integrating customer feedback and repair depot insights directly into the product lifecycle, I worked to influence design decisions and enhance overall repairability. Currently at Rivian, I continue this focus by leading the development of advanced logistics and supply chain applications that connect physical operations to digital intelligence.
I am not a hardware engineer or a data center architect. I am a supply chain enthusiast genuinely curious about systemic operational problems — the kind that sit in the gap between what a system reports and what is physically true. This article exists because I believe the AI infrastructure industry is standing at exactly that gap right now. And almost nobody is talking about it.
The $725 Billion Question Nobody Is Asking
What the largest capital investment cycle in technology history means for hardware lifecycle management — and why the industry needs to pay more attention.
The numbers that defined 2025 and 2026
The AI infrastructure investment cycle is unlike anything the technology industry has seen before. The four largest hyperscalers — Amazon, Google, Meta, and Microsoft — spent approximately $410 billion on capital expenditure in 2025. In 2026, that number is projected to reach $725 billion, a 77% year-over-year increase. Goldman Sachs now projects a combined $5.3 trillion in hyperscaler CapEx between 2025 and 2030. The baseline aggregate estimate across compute, data centers, and power rises to $7.6 trillion between 2026 and 2031.
To put that in context: $725 billion in a single year exceeds the entire GDP of Saudi Arabia
A new class of infrastructure builder has entered the race
What makes this moment particularly significant is that the buildout is no longer the exclusive domain of the traditional hyperscalers. A new and structurally distinct category of infrastructure investor has emerged — and their presence changes the reverse logistics calculus entirely.
The AI model companies — from compute renters to infrastructure owners. Anthropic, OpenAI, and xAI began as pure model companies, renting compute from AWS, Google Cloud, and Microsoft Azure. That model is changing rapidly. Anthropic has committed $50 billion to American computing infrastructure, partnering with Fluidstack to build custom data centers in Texas and New York, with the first sites going live in 2026. Beyond construction, Anthropic has signed more than a dozen letters of intent for direct data center leases totaling over 1 gigawatt, and has secured a contract with xAI for exclusive access to Colossus 1 in Memphis — a facility housing over 220,000 NVIDIA GPUs drawing approximately 300 megawatts of power, at $1.25 billion per month through May 2029.
OpenAI, through the Stargate joint venture with SoftBank and Oracle, is targeting $500 billion in data center infrastructure. xAI built its own hybrid data center and power-generation plant in South Memphis — and Elon Musk has noted that gas turbines are sold out through 2030.
These companies are not renting infrastructure. They are building it. And unlike hyperscalers that have spent decades developing the operational systems to manage what they build, the AI model companies are acquiring physical infrastructure at extraordinary speed without the institutional knowledge of how to sustain it.
NVIDIA — from chipmaker to infrastructure architect. NVIDIA’s transformation from GPU manufacturer to infrastructure investor may be the most structurally significant development of 2026. In a single year, NVIDIA has committed more than $40 billion across cloud providers, data center operators, optical networking firms, manufacturing partners, and AI model developers — including stakes in CoreWeave, Nebius, Coherent, Marvell, Corning, and IREN. NVIDIA has also partnered with Bechtel to modularize a 1-gigawatt data center design, directly shaping how physical infrastructure gets built at scale.
“Your data center used to be a data center for files. It’s now a factory for tokens.” — Jensen Huang, NVIDIA CEO, Computex 2026
The neocloud layer — speed without operational depth. CoreWeave ended Q1 2026 with more than 1 gigawatt of active power across nearly 50 data center locations, with customers including Meta, OpenAI, and Anthropic. Its contracted revenue backlog reached $99.4 billion as of March 31, 2026. Nebius posted Q1 revenue growth of 684% year-over-year. Lambda is preparing its own IPO after securing a deal with Microsoft and $1.5 billion in funding.
Get Krishna’s stories in your inbox
Join Medium for free to get updates from this writer.
These companies are growing at extraordinary rates, deploying hardware at hyperscaler density without the decades of operational infrastructure that hyperscalers have built. Their reverse logistics functions are, in most cases, improvised. A company with $99 billion in contracted future revenue and $35 billion in annual CapEx cannot afford to manage the lifecycle of that hardware as an afterthought.
What goes into that $725 billion — the full stack
Understanding what the CapEx actually buys matters for understanding the reverse logistics challenge it creates. This is not monolithic spending on a single type of hardware. It is a layered stack of components, each with different failure modes, different lifecycle characteristics, and critically different reverse logistics requirements.
- Layer 1 — Compute: GPUs, TPUs, and custom accelerators: NVIDIA H200 and B200 GPUs are the headline components. H100s currently retain $15,000-$20,000 in secondary market value on the certified refurbishment market. A single eight-GPU server configuration represents $120,000-$160,000 of hardware. GPU-dense server chassis weigh 50–70 pounds, include water-cooling manifolds in liquid-cooled configurations, and require trained technicians to safely disconnect. Beyond NVIDIA, each hyperscaler deploys proprietary silicon — Google’s TPUs, Amazon’s Trainium, Meta’s MTIA, Microsoft’s MAIA — each with distinct failure characteristics and disposition requirements.
- Layer 2 — Memory: the invisible bottleneck: High Bandwidth Memory (HBM) — the stacked DRAM AI accelerators depend on — is manufactured by only three companies (SK Hynix, Samsung, Micron) using processes that cannot be ramped quickly. HBM demand has grown roughly 5x between 2023 and 2026. Memory failures are a primary driver of GPU returns, and memory is also the component most sensitive to mishandling during reverse logistics. H100 GPUs retain $12,000-$18,000 in secondary value in early 2026 largely because HBM supply constraints make older-generation hardware still valuable for inference.
- Layer 3 — Networking: the invisible infrastructure: InfiniBand high-speed networking fabric connects GPU clusters and enables the parallel processing required for large model training. Networking failures are among the hardest to diagnose — they often manifest as intermittent performance degradation rather than clean failure events, making triage complex and RMA rates difficult to calibrate. A failed transceiver in the wrong location can take down an entire rack of compute.
- Layer 4 — Power and cooling: the civil infrastructure layer: Generators, switchgear, transformers, UPS systems, and cooling infrastructure are the physical foundation the compute stack sits on. These components have completely different failure modes, different repair ecosystems, and fundamentally different logistics requirements than IT hardware. A generator does not ship via FedEx. It requires specialized heavy freight, oversize permits across every state it transits, crane coordination at the destination, and site-level receiving infrastructure that most IT-oriented logistics functions are simply not equipped to handle.
- Layer 5 — Storage and edge compute: NVMe-dense storage nodes where training datasets and model checkpoints accumulate at petabyte scale carry distinct data security obligations at end of life. A storage node retiring from an AI training cluster may contain model checkpoints, training data, or intermediate representations that constitute sensitive intellectual property. Decommissioning without certified data destruction — documented to NIST SP 800–88 standards — is not just an environmental compliance failure. It is a potential intellectual property and competitive intelligence risk.
Each of these five layers requires a different operational playbook, different vendor relationships, different carrier requirements, and different documentation standards. Most organizations building AI infrastructure today have a coherent strategy for deploying all five layers. Almost none have a coherent strategy for sustaining and recovering them.
The hardware reality nobody is publishing
Meta’s Llama 3 cluster experienced a component failure every three hours. That is not a dramatic event — it is normal operating conditions for AI infrastructure at scale: 419 unexpected interruptions across a 16,384-H100 cluster over 54 days of training, with roughly half of those interruptions attributable to the GPUs themselves or their HBM3 memory.
Industry analysts including Gartner and Dell’Oro Group identify AI accelerator replacement timelines of 2–3 years for enterprises running inference and fine-tuning workloads at scale, compared to the 5–7 year lifecycles that traditional server refresh planning assumes. Each GPU generation since 2022 has delivered a fundamental memory architecture upgrade. The NVIDIA H100 shipped with 80GB of HBM2e. The H200 replaced that with 141GB of HBM3e — nearly double the memory bandwidth on an entirely different architecture. These are not upgrades procurement departments can defer.
The first real wave of GPU-dense AI servers deployed between 2022 and 2024 is likely to enter decommissioning cycles between 2026 and 2029. The global data center decommissioning market expanded from $12.12 billion in 2025 to $12.95 billion in 2026, growing at 7.37% annually — driven almost entirely by AI infrastructure acceleration.
Do the math. If 75% of $725 billion — roughly $544 billion — is hardware, and that hardware has an effective service life of 18–36 months, the industry could be generating an estimated $180–270 billion annually in hardware requiring active lifecycle management by 2028. When a 10-rack AI cluster retires, it generates 80–120 GPU accelerators in a single event, each with distinct data sensitivity profiles and varying HBM memory architectures. Standard ITAD pickup schedules were not designed for this disposal density or per-device compliance complexity.
That wave is coming. The question is whether the industry will be ready for it.
The sustainability of the CapEx cycle
The five largest hyperscalers are on track to spend between $700 billion and $900 billion on capital expenditures in 2026, with Amazon projected to turn free cash flow negative this year as a result. The hyperscalers are increasingly leaning on debt markets to bridge the gap between rapidly rising AI CapEx budgets and internal free cash flow — transforming historically cash-funded business models into ones utilizing leverage.
Goldman Sachs projects that the big four hyperscalers will face combined annual depreciation approaching $400 billion by 2030 — more than their combined profits in 2025. This cannot continue indefinitely. The CapEx cycle will not collapse — the competitive dynamics make underinvestment too dangerous — but it will mature.
Goldman’s own base case for 2027 sits at approximately $1.1 trillion in hyperscaler CapEx, well above the $920 billion Wall Street consensus. Goldman argues this deceleration in growth from 77% in 2026 to just 22% in 2027 is too conservative. The bank estimates that token consumption will increase 24 times through 2030 — largely led by the rise of enterprise AI agents — which would sustain elevated infrastructure spend through the end of the decade.
As the cycle matures, the conversation will shift from “how fast can we build” to “how efficiently can we operate what we’ve built” and “how much value can we recover from what we’re retiring.” That shift is where reverse logistics, service operations, and hardware lifecycle management move from operational afterthought to strategic imperative. Organizations that treat lifecycle management as a cost center will find that the depreciation math works against them. Organizations that treat it as a value recovery and operational intelligence function will find that it changes the economics of the entire infrastructure investment.
The question that changes everything
Every conversation about AI infrastructure investment focuses on the same metrics: gigawatts of capacity, petaflops of compute, terabytes of storage, milliseconds of latency.
Almost nobody is tracking the metric that will ultimately determine whether the $725 billion investment holds its value: what percentage of deployed hardware is operating at full capacity at any given moment, and how fast does failed hardware return to service?
Hardware utilization efficiency is entirely a function of service operations and reverse logistics. And right now, across an industry deploying capital at a rate that has no historical precedent, it is being managed as an afterthought.
The operational disciplines required to get this right are not new. They have been built and refined in consumer electronics, automotive, telecommunications, and defense. What is new is the scale at which the AI infrastructure industry now needs to apply them — and the urgency with which organizations at every layer of the stack need to build the capability.
In the next issue, I will go inside the specific operational disciplines that determine whether reverse logistics becomes a cost center or a strategic value recovery function — drawing on what I learned building the data foundations of reverse logistics at Google, and what the AI infrastructure industry needs to build now before the first wave of hardware decommissioning hits.
Coming in Issue 2: The anatomy of a reverse logistics system — from RMA triage to depot diagnostics, and how data transforms hardware returns into product quality intelligence.
Originally published at https://kg12803700.substack.com on July 12, 2026.
