<?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 David Schaerf, North Star Labs on Medium]]></title>
        <description><![CDATA[Stories by David Schaerf, North Star Labs on Medium]]></description>
        <link>https://medium.com/@david_46352?source=rss-d7e0085c3ed0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*ACBGFszgIBlQhU65</url>
            <title>Stories by David Schaerf, North Star Labs on Medium</title>
            <link>https://medium.com/@david_46352?source=rss-d7e0085c3ed0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 06:59:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@david_46352/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[The AI a hospital buys and the AI its clinicians use are two different systems]]></title>
            <link>https://medium.com/@david_46352/the-ai-a-hospital-buys-and-the-ai-its-clinicians-use-are-two-different-systems-89c5fb47ab64?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/89c5fb47ab64</guid>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Wed, 16 Sep 2026 13:59:57 GMT</pubDate>
            <atom:updated>2026-09-16T13:59:57.721Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/839/1*pxJ9klQcQ1T4UkAAJOhFyw.png" /></figure><p>Most doctors now use AI at work. Almost none use the version their hospital actually stands behind. That gap is not a scandal and it is not a story about undisciplined clinicians. It is a structural fact about how demand and supply have moved at different speeds, and it shows up in every health system where anyone has bothered to measure it.</p><p>I ran a small survey to find out how wide the gap really is. Forty-four clinicians in German-speaking hospitals, asked two plain questions: what AI do you actually use in your work, and what has your institution sanctioned. The sample is small and self-selected, so it describes a pattern rather than a prevalence. But the pattern was stark enough, and consistent enough with what other markets are reporting, that it is worth laying out in full.</p><h3>What the survey found</h3><p>Ninety-three percent of respondents use AI in their clinical work. Only eighteen percent have a sanctioned tool available at their hospital. The rest improvise, on personal accounts and personal devices, outside anything the institution controls or can see.</p><p>Underneath those headline numbers, the detail is where the argument lives. Of the clinicians who use AI, most use it only through privately procured tools, not through anything their hospital provides. A large share report that their institution has no clear rule governing the use at all, and the most common reason the use stays private is the same everywhere: uncertainty about data protection and compliance. A smaller group described using AI in tasks that touch clinical decisions directly, and nearly all of them were doing so with private tools in institutions that had set no rules.</p><p>There is one caveat I want to state plainly, because it is the difference between a defensible finding and an overclaim. The survey captured the task a clinician was using AI for, not the content they typed into it. Someone who says they use AI to help with a discharge letter might mean drafting anonymised prose, or might mean pasting in a specific patient’s findings. For the data-protection question those are entirely different acts, and the survey cannot tell them apart. So this is not a claim that patient data is leaving hospitals. It is a claim that a category of clinical work is happening on ungoverned tools, and that whether sensitive data is involved is exactly the thing no one is currently in a position to know.</p><h3>It is not a German problem</h3><p>My first assumption was that this was a peculiarity of the German system, which is often characterised as cautious and slow on digital health. The assumption did not survive contact with the other markets.</p><p>In the United States, surveys put professional AI use among physicians above eighty percent. The difference is not the level of use, it is that a clinical-grade tool actually won the market. OpenEvidence, which answers from the medical literature and cites its sources, is now the most widely used clinical AI tool among US physicians, ahead of every alternative by a wide margin according to independent surveys. Its exact penetration and how often physicians reach for it are contested, and the company’s own frequency figures should be treated as company figures. But the direction is not in doubt. A clinical-grade default emerged in the US, and it emerged from the bottom up, through individual adoption, not because any institution mandated it.</p><p>The United Kingdom shows the same shape as Germany. A large share of physicians report reaching for consumer tools, ChatGPT and Copilot, to answer clinical questions, while only a small minority use a medical-specific tool. Heavy use, thin sanctioned supply.</p><p>Look further out and the picture gets stranger rather than simpler. China closed the gap by force, with the state pushing AI into hospital workstations directly. It is governed, but from the top down rather than by choice. South Korea shows the opposite failure: a national survey of more than two thousand doctors found heavy use, much of it clinical-grade imaging AI, alongside almost no institutional governance, with only a tiny fraction of institutions having any AI guidelines at all. Israel took the institutional route, building on decades of unified health records, with hospital-wide sanctioned deployments and home-grown safety frameworks that international bodies now cite as benchmarks. Governance ahead of most of the world, built on data infrastructure almost nobody else has.</p><p>Five systems, five different answers to the same question. None of them accidental, and none of them a story about individual clinicians misbehaving.</p><h3>Why the gap exists, and why it persists</h3><p>The clean way to see this is that the AI a hospital buys and the AI its clinicians actually use are usually two different systems. One is narrow, validated, procured, and largely invisible, embedded in a radiology workflow or a documentation module. The other is broad, personal, and ungoverned, opened in a browser tab between patients. The clinician lives in the second system while the institution assumes it has handled the first.</p><p>The reason the gap does not close on its own is a matter of speed and incentive. Procuring a new clinical tool takes a health system somewhere between twelve and eighteen months. Opening a consumer AI tool on a personal phone takes zero days. Policy cannot out-run that gap. It can only decide what happens after the first incident traces back to one of those personal accounts, and by then the use is long established. This is not new behaviour. Clinicians moved to consumer messaging apps years before hospitals offered a compliant alternative, and procurement never won that race on speed, only on liability once a breach forced the switch.</p><p>Which raises an uncomfortable question about whether the forcing event will even arrive this time. The messaging switch happened because a breach produced a discrete, dated, visible incident that IT could point to. Ungoverned AI use fails differently. There is no server log, no breach notification, no single moment. The harm, if it comes, surfaces slowly and individually, one patient and one outcome at a time, and it may never aggregate into the kind of visible event that forced the earlier switch. The institutions waiting for the incident that tells them to act may be waiting for something that never happens.</p><h3>What a sanctioned option actually has to do</h3><p>If banning the tool does not work, and buying a tool nobody opens does not work either, then the real question is what would make a clinician choose the sanctioned option over the one already in their pocket. The public conversation around this survey, which drew in clinicians, informaticians, and operators across several health systems, converged on a set of conditions that are worth stating as a framework. A sanctioned tool wins only when it satisfies all of them, and most current efforts satisfy none.</p><p>It has to be as good as the consumer tool, not merely compliant. This is the condition everyone underrates. A safe, approved, clunky internal tool loses to a fast, useful consumer one every time. Safe is table stakes. Good enough that people actually choose it is the entire game, and it is far harder to procure than it is to specify.</p><p>It has to be close to the moment of need. Adoption follows proximity. Whichever tool sits closest to the point where a clinician has a question becomes the default, and a tool that requires opening a separate system, logging in, and waiting has already lost to the browser tab. Notably, the tool that won the US market needs no access to the patient record at all, which is precisely why it spread without procurement. The tasks that require touching the record are the ones stuck in the long procurement cycle.</p><p>It has to be grounded in current, traceable sources. A tool that answers confidently from an outdated source, presented as authoritative, is more dangerous than an obvious guess, because it looks trustworthy. Sanctioned has to mean provenance-checked, not just permitted. The difference between a clinical-grade tool and a confident chatbot is whether you can see where the answer came from and how current it is.</p><p>It has to restore visibility, not remove it. The deepest cost of ungoverned use is not any single exposure, it is institutional blindness. When clinicians work on outside tools, the institution loses the ability to learn which inputs shaped a decision, which means it cannot audit, cannot improve, and cannot even see what is failing. The blindness compounds, because every ungoverned interaction is data the institution never gets to learn from, which makes the sanctioned option it is trying to build worse than it needs to be. A sanctioned tool is partly how an institution gets its own risk surface back into view, but only if it is good enough that people use it.</p><p>And underneath all of it sits the reason the framework matters at all. The gap is not an abstraction about governance efficiency. It is a patient in front of a clinician who is using an ungoverned tool, and no one chose that on the patient’s behalf. That is why “safe” cannot be a checkbox and “good” cannot be optional. The two have to arrive together, or the sanctioned option simply does not get used, and the patient keeps covering the gap.</p><h3>Where this goes</h3><p>The demand side of this is settled. Clinicians have adopted AI, in every market anyone has measured, and no policy is going to reverse that. What remains open is entirely on the supply side: whether anyone builds a sanctioned option good enough, close enough, grounded enough, and visible enough that clinicians choose it over what they already use.</p><p>That is a harder instruction than it sounds. “Build a clinical-grade tool” is something a health system knows how to procure. “Match the usefulness and friction of consumer AI while adding provenance, safety, and an audit trail” is a product capability most institutions cannot buy off a shelf and cannot easily build. The gap between those two sentences is the whole problem, and it is why the sanctioned option keeps losing.</p><p>The survey behind this piece was small and deliberately exploratory, meant to test a suspicion rather than settle it. A larger, multi-country study is now taking shape to measure the gap properly, pairing what clinicians actually use against what their institutions have sanctioned, across health systems rather than within one. The pattern already looks robust enough to be worth measuring at scale. The question that follows from it, who builds the version clinicians would actually choose, is the one worth spending the next few years on.</p><p><em>The working paper from the pilot survey is available on request. Figures cited for the US, UK, and other markets are drawn from published third-party surveys and are the reported figures of their respective sources.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=89c5fb47ab64" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Absorption Line.]]></title>
            <link>https://medium.com/@david_46352/the-absorption-line-19c2a1326305?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/19c2a1326305</guid>
            <category><![CDATA[health-technology]]></category>
            <category><![CDATA[digital-health]]></category>
            <category><![CDATA[ehr]]></category>
            <category><![CDATA[ai-in-healthcare]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Wed, 08 Jul 2026 16:36:17 GMT</pubDate>
            <atom:updated>2026-07-08T16:36:17.920Z</atom:updated>
            <content:encoded><![CDATA[<h4>Every frontier model release raises the water. What sits below the line stops being a product.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VlvEi7d6uq48jTG7KAMRKw@2x.jpeg" /></figure><p>What holds above it is all that survives.</p><p>Two years ago, clinical transcription was a product category. Companies raised hundreds of millions to build AI that could sit in an exam room, listen to a clinician talk to a patient, and turn that conversation into structured documentation. It was hard to do, it was valuable, and it commanded serious pricing.</p><p>That is over. Epic now ships a native AI scribe inside its EHR. athenahealth bundles one at no additional cost. Doximity gives one away for free to its network of over two million clinicians. The capability that entire companies were built on is now a checkbox that the platform vendors include in the base product.</p><p>This is not a story about scribes. Scribes are the first domino. The pattern underneath is structural, and it applies to every category in clinical AI.</p><p>I call it The <em>Absorption Line.</em></p><p><strong>The waterline</strong></p><p>Picture a volcanic island surrounded by a rising sea. The island is the clinical AI market. Every company on it occupies a position at some elevation. The sea is frontier model capability, and it rises with every release. Whatever capability the sea reaches stops being differentiated. The platforms absorb it and give it away for free, the same way Epic absorbed transcription.</p><p>The Absorption Line is the current waterline. Everything below it has already been commoditised. Everything above it is still defensible, for now. The critical word is “still,” because the line does not stay where it is. It climbs.</p><p>The transcript went under in 2025. Coding and revenue cycle management are next. You can see the water moving toward them already. Ambience raised $373 million and positioned itself squarely in coding and RCM, but that is the strip of land directly above the current waterline, the next territory the platforms will contest. Whether Ambience holds that ground depends on how fast it converts the position into something the platforms cannot replicate. The clock is the rising water.</p><p><strong>The mistake everyone makes</strong></p><p>Most founders respond to the rising line by shipping faster. More features, more releases, a longer roadmap. This feels like the right response, but it is a trap. Shipping more features is competing with the frontier on capability, and the frontier will always win that race. It has more compute, more data, more capital, and no margin pressure. Speed decides how long you last above the waterline, not whether you survive.</p><p>The companies that understand this are not running faster. They are digging deeper. The question is not “how do I stay ahead of the next model release” but “what do I hold that the next model cannot give away for free.”</p><p>That question has two fundamentally different answers, and confusing them is the most expensive strategic mistake in health AI right now.</p><p><strong>Bedrock</strong></p><p>Some moats sit below the waterline permanently. The sea cannot reach them no matter how high it rises. I call these bedrock moats.</p><p>Proprietary outcome data is the strongest. When your product runs inside a clinical workflow and captures real patient outcomes from that deployment, you accumulate a dataset that no frontier model can scrape, buy, or synthesise. It compounds with every encounter. It never enters anyone else’s training set. And it gets more valuable the longer you run, which means the moat widens over time rather than narrowing. This is the rarest asset in health AI and the one that matters most.</p><p>Regulatory clearance and clinical liability form the second bedrock position. A frontier model ships general capability and takes zero clinical responsibility. Someone has to own the clearance, carry the liability, and be the accountable entity that a hospital can contract with and an insurer can underwrite. That accountability is a moat by law, not by product. It cannot be absorbed because the platforms do not want to carry the risk.</p><p>Bedrock moats are permanent by nature. They do not erode with the next release. They do not depend on being faster than the competition. They exist because they are structurally unreachable.</p><p><strong>Seawalls</strong></p><p>Then there are seawalls. These are moats that hold back the water, sometimes for years, but they erode with every frontier release. They buy time. They do not buy safety.</p><p>Deep embedding is the most common seawall in health AI. Your product is wired into the medical record, the billing system, the compliance stack, the clinical workflow, so deeply that removing it would be surgery. This is a real moat, but only when the entanglement is yours. If you are a thin feature layer renting Epic’s surface, then Epic holds the embedding, not you. And when Epic ships a native version of what you do, your embedding advantage disappears overnight because it was never yours to begin with.</p><p>Operational process is another seawall. The human plus AI machine that turns raw model output into a guaranteed, paid, accountable clinical outcome. The model handles the pattern matching. The human handles the judgment, the edge cases, the relationship. Together they deliver something neither can do alone, and that is genuinely valuable. The problem is the half life. Every frontier release makes the model’s share larger and the human’s share smaller. The seawall gets lower every year.</p><p>Habit and trust form the third seawall. The reflexive reach, the clinician who opens your app without thinking, the patient relationship that a competitor cannot re-point with a better product demo. Habits are sticky. But a good enough free alternative bundled inside the platform the clinician already uses can redirect a habit faster than most founders expect. Doximity is running exactly this play right now, giving away a free scribe into a network it already owns, betting that the habit follows the platform.</p><p><strong>Who understands this</strong></p><p>Watch the companies that are converting seawalls into bedrock while the water is still low enough to build.</p><p>Abridge is the clearest example. It took roughly 30 percent of the AI scribe market according to KLAS data, the largest share in the category. But instead of defending that position by shipping more transcription features, it moved up into revenue cycle intelligence and real-time prior authorisation through a partnership with Availity. The transcription share was the seawall. Revenue cycle with proprietary outcome data from thousands of clinical deployments is the bedrock it is building underneath. The $5.3 billion valuation prices that trajectory, not the current product.</p><p>Doximity went the other direction, and it is worth understanding why. Doximity is not trying to hold ground above the waterline. It is joining the tide. By giving away a free scribe to its two million clinician network, it is using the commoditised capability as a distribution weapon to protect a completely different business: its physician marketplace and its pharmaceutical advertising revenue. Doximity is not drowning. It is arming the tide. That is a legitimate strategy, but it is not the same game as the companies trying to survive on the island.</p><p>Ambience sits in the most exposed position of any well-funded company in the space. It raised $373 million and built a strong product in ambient documentation with a move into coding and RCM. But coding and revenue cycle are the next strip of land the waterline is heading for. Ambience’s window to convert that position into bedrock, through proprietary outcome data or regulatory clearance or both, is open now and closing. The next two years will determine whether Ambience becomes a permanent summit resident or the next capability the platforms absorb.</p><p>Below the waterline, the picture is already clear. Suki holds roughly 10 percent market share as a clinical assistant, but its depth in any single workflow is limited. Nabla and Freed each sit at roughly 4 percent, competing on price and speed in a category where the floor price is approaching zero. These companies are not failing because their products are bad. They are failing because the capability they sell is no longer scarce.</p><p><strong>The European lag</strong></p><p>In Europe the water is lower, for now. Fragmentation across national health systems and regulatory complexity slow the tide. CompuGroup and Dedalus, the two largest European EHR platforms, cannot bundle what Epic bundles. They do not have the AI infrastructure, the integration depth, or the developer ecosystem to absorb clinical AI capabilities the way the American platforms do.</p><p>That lag is real, and it creates a genuine window for European health AI founders. The absorption that is already underway in the US market has not started in most European markets. The seawall time that American companies have already burned is still available in Europe.</p><p>But the window is closing. EHDS is forcing European EHR systems to expose standardised APIs, which accelerates the platform layer’s ability to absorb third party capabilities. And the frontier models do not respect geography. When a model can transcribe, code, and summarise a clinical encounter for free, it does not matter whether the platform absorbing it is Epic in Wisconsin or CompuGroup in Koblenz.</p><p>Most European health AI founders are spending their lag time the same way their American counterparts did: shipping features. The ones who survive will be the ones who used it to reach bedrock.</p><p><strong>The test</strong></p><p>The chart accompanying this article maps ten companies against the Absorption Line. Four are already under. Two hold defensible positions above the dashed line marking where the water goes next. The rest sit on contested ground.</p><p>The test for your own company, your own investment, or your next procurement decision takes one question.</p><p>What are you holding that the tide cannot reach?</p><p>If the answer is proprietary outcome data that compounds with every deployment, or regulatory clearance that carries real clinical liability, you are standing on bedrock. If the answer is speed, features, embedding in someone else’s platform, or the habits of users who have not yet seen a free alternative, you are standing on a seawall. Seawalls are not worthless. They buy time. But they are not safety, and the water is rising.</p><p>The only strategic move that matters is conversion. Use the time your seawalls buy to dig down to bedrock. Turn a process advantage into a proprietary outcomes dataset. Turn embedding into being the system of record. Turn a habit into a data loop that compounds independently of the platform underneath.</p><p>Most companies spend that time shipping features instead, and the water takes them anyway.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19c2a1326305" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Winner Takes Next: US Edition]]></title>
            <link>https://medium.com/@david_46352/winner-takes-next-us-edition-462aed86ebdb?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/462aed86ebdb</guid>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[health]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[ai-in-healthcare]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Wed, 17 Jun 2026 10:19:05 GMT</pubDate>
            <atom:updated>2026-06-17T10:19:05.028Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r0xKMoRFHUdyEdEBrGZ86A.jpeg" /></figure><h3>Two Markets, Same Thesis, Two Different Playbooks</h3><p>Abridge raised $300 million in Series E four months after a $250 million Series D. The company is deployed in over 150 health systems and valued at $5.3 billion. It has won Best in KLAS for ambient AI two years running.</p><p>It started as a clinical documentation tool. It is now expanding into revenue cycle management, billing, and coding.</p><p>The scribe is becoming the clinical workflow platform.</p><p>If you read the European edition of Winner Takes Next a few weeks ago, this pattern is familiar. Win one category. Embed so deeply that no competitor can displace you. Then use that position to take the next one.</p><p>The thesis is the same in the US. The landscape and the playbook are different. And the differences matter if you are building, investing, or competing across both markets.</p><h3>The US landscape</h3><p>The US healthcare AI ecosystem runs hotter, faster, and with more capital than Europe. The companies are bigger, the rounds are larger, and the category leaders are expanding earlier.</p><p>Abridge is the clearest example. But the pattern runs across the chart.</p><p>Ambience Healthcare raised $243 million in Series C from a16z and the OpenAI Startup Fund. Documentation was the wedge into Cleveland Clinic, UCSF Health, Houston Methodist, and dozens of other systems. Clinical decision support and workflow automation are the expansion. Their investors described ambient AI as a “wedge into essential clinical workflows over time.” That is Winner Takes Next stated as an investment thesis.</p><p>Nuance DAX Copilot, now a Microsoft product, is embedded in 77% of US hospitals. Its integration with Epic is so deep that the line between point solution and infrastructure has disappeared. DAX is no longer a tool that sits on top of Epic. It is becoming part of what Epic is.</p><p>Amazon One Medical is running the reverse play. Instead of starting inside the clinical system and expanding outward, Amazon started with the consumer and is moving inward. Pharmacy, primary care, and AI in one vertically integrated stack. The traditional clinic platforms watching this approach from across the market are underestimating the speed.</p><p>These are all Winner Takes Next moves. Different starting positions, different expansion vectors, same underlying logic. Win the first category. Lock in the workflow. Climb.</p><h3>The structural advantage the US has over Europe</h3><p>The US market moves faster than Europe for a structural reason that most people outside the American healthcare system do not fully appreciate.</p><p>The 21st Century Cures Act, passed in 2016 and enforced through its information blocking provisions, required every certified EHR vendor in the United States to open standardized FHIR APIs. It made information blocking illegal. Any clinical AI company that wants to connect to Epic, Oracle Health, or athenahealth can do so through a standardized interface.</p><p>The door is open. The cost of walking through it is low. The barrier to winning the first category is technical integration, and that barrier is largely gone.</p><p>In Europe, the situation is reversed. KIS and HIS vendors have no regulatory obligation to open their APIs. Integration is negotiated vendor by vendor, hospital by hospital. Getting inside a European hospital system takes months of technical and commercial negotiation. Every integration is custom. Every connection is fragile.</p><p>This creates two fundamentally different strategic playbooks.</p><h3>The US playbook</h3><p>Land fast through open FHIR APIs. The integration layer is standardized, so the race is to clinical adoption. Get into the health system, get physicians using the product every day, and build the workflow dependency that makes switching painful. Then expand into adjacent categories before competitors who connected just as easily can build the same habits.</p><p>Integration is commodity. Clinical adoption is the moat.</p><p>Abridge is the template. It did not win because its Epic integration was technically superior. Every ambient AI company can connect to Epic. Abridge won because physicians trust it, use it daily, and have changed how they document. Copying the API connection takes a week. Replicating the clinical habit takes years.</p><p>The risk of the US playbook is speed of competition. Because the door is open for everyone, the window between landing and embedding is short. A company that connects to Epic and does not immediately drive deep clinical adoption will find three competitors connected to the same system within months.</p><h3>The European playbook</h3><p>Negotiate the integration. Accept the slower timeline. Build deep technical embedding that competitors cannot easily replicate. The moat in Europe is the connection itself, because every integration is hard won and every new competitor faces the same painful negotiation.</p><p>The European playbook rewards patience and relationship depth over speed. The companies that have spent years building integrations with Dedalus, CompuGroup, or Orbis have a structural advantage that no amount of capital can quickly overcome. A well funded competitor entering the German hospital market still has to negotiate the same KIS integration from scratch.</p><p>Integration is the moat. Clinical adoption follows because the competition never made it through the door.</p><p>Recare and Doctolib are templates for this approach. Both built deep integration networks across European health systems over years. Both are now using those networks to expand into adjacent categories. The expansion is slower than in the US, but the defensibility at each step is higher.</p><h3>The convergence</h3><p>Here is where the comparison gets strategically interesting.</p><p>Europe is moving toward open APIs. The European Health Data Space regulation is in effect, with compliance deadlines in 2029 and 2031. FHIR implementation guides aligned with EHDS were published in late 2025. The implementing acts specifying the European Electronic Health Record Exchange Format are expected in early 2027.</p><p>The intent is clear. Europe is building toward a version of what the 21st Century Cures Act created in the US: mandatory interoperability through standardized APIs.</p><p>The current reality is just as clear. Only 27% of European hospitals achieve full cross border data exchange today. 54% of hospitals cite lack of standardized FHIR APIs as their top constraint. The Nordic and Benelux regions lead at 47% connectivity. Southern Europe sits at 14%.</p><p>Europe is years behind the US on open API infrastructure. But the direction of travel is set.</p><p>When the European market eventually reaches a meaningful level of API openness, the competitive dynamics shift. The companies that built their moat on integration difficulty face a world where that moat erodes. New competitors can connect more easily. The hard won exclusivity of a custom KIS integration becomes less defensible.</p><p>The European companies that will be best positioned for that transition are the ones building clinical adoption moats alongside their integration moats. The ones that physicians use every day and would resist switching away from, regardless of how easy it becomes for a competitor to technically connect.</p><p>The integration moat buys you time. The clinical adoption moat is what survives when the APIs open.</p><h3>Revenue and reimbursement: the US category with no European equivalent</h3><p>The US chart has a column that the European edition did not need. Revenue and reimbursement is a massive standalone category in the United States, built on a billing complexity that simply does not exist in single payer or social insurance systems.</p><p>Waystar is public. R1 RCM processes billions in claims. AKASA, Candid Health, Cedar, and SmarterDx are building AI native revenue cycle platforms from the ground up. The category is large enough and complex enough to sustain a full ecosystem of companies that do nothing else.</p><p>This matters for the Winner Takes Next thesis because revenue cycle is where several clinical AI companies are expanding. Abridge moving from documentation into billing is a direct example. The documentation captures the clinical encounter. The same data, structured correctly, feeds the billing workflow. The wedge creates the bridge.</p><p>European companies watching this should pay attention. The billing complexity may be uniquely American, but the pattern of using clinical data capture as a bridge to financial workflow is universal. The European equivalent is smaller in scale but structurally similar: documentation feeding referral workflows, feeding insurance authorization, feeding reimbursement. The companies that own the clinical data capture in Europe will eventually face the same expansion question.</p><h3>Payer AI: the open frontier</h3><p>The most interesting column on the US chart is the one with no clear winner.</p><p>Payer AI. Utilization management, risk prediction, claims intelligence. Cohere Health is building prior authorization AI. Oscar, Clover, and Devoted Health are approaching it from the insurance side. But nobody dominates. Nobody has built the moat.</p><p>This is the least consolidated category in US healthcare AI and arguably the highest value territory still open on the map.</p><p>The reason it matters for the Winner Takes Next thesis is the expansion question. The companies that have already won clinical categories are accumulating data, relationships, and workflow integrations that position them to move into the payer space. Abridge capturing clinical encounters at 150 health systems has a data asset that is directly relevant to payer decision making. Whether Abridge or a company like it makes that move is one of the most consequential strategic questions in US healthcare AI over the next three years.</p><p>From the other direction, the payer companies that build strong AI capabilities have an expansion path into provider workflow. The boundary between payer and provider AI is artificial. It exists because the business models are separate. The data and the workflow are continuous.</p><p>The company that bridges that boundary first has a Winner Takes Next opportunity that dwarfs anything else on the chart.</p><h3>Two markets, two playbooks, one thesis</h3><p>The US landscape is moving faster, with more capital, on open infrastructure. The European landscape is moving more deliberately, with deeper integration moats, on infrastructure that is still largely closed but transitioning.</p><p>Both markets reward the same underlying behavior: win a category deeply enough that expansion becomes a natural extension of the position you already hold.</p><p>The strategic question for a company operating in both markets is whether to run the same playbook in both or adapt to the structural realities of each. The answer, based on every company I have watched succeed in both, is adaptation. Speed in the US. Patience in Europe. Clinical adoption everywhere.</p><p>The companies that figure out how to run both playbooks simultaneously are the ones that will look very different in three years. The ones running a single playbook in both markets are leaving value on the table in at least one of them.</p><p>Winner takes next. The question is which version of next your market structure actually allows you to reach.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=462aed86ebdb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Intelligence Stack]]></title>
            <link>https://medium.com/@david_46352/the-intelligence-stack-ecbc512ea59e?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/ecbc512ea59e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[apple]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Wed, 10 Jun 2026 09:31:00 GMT</pubDate>
            <atom:updated>2026-06-10T09:45:49.977Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oZsdfdZfmQnE_xLeKXTsiA.png" /></figure><p>AI tokens right now are priced like Facebook ads in 2015.</p><p>If you were spending on Facebook acquisition in 2015, you remember the economics. A click cost thirty cents. A qualified lead cost a few dollars. The ROI was so obvious that the only question was how much budget to move in. Then the window closed. Costs tripled. Then tripled again. By 2020 the same lead cost ten times what it cost five years earlier. The companies that built organic distribution early survived. The ones that stayed dependent on paid acquisition got squeezed.</p><p>The same pattern is playing out in AI right now. Frontier model pricing is artificially low. Not because inference is cheap. Because the hyperscalers are buying market share and eating the loss.</p><p>And just like Facebook ads in 2016, the prices are already starting to move.</p><h3>The subsidy window is closing</h3><p>On June 1, GitHub switched every Copilot user from flat-rate to usage-based billing. Developers started burning through a month of credits in two hours. GitHub’s own explanation was plain: the pricing needed to reflect actual compute costs.</p><p>This was not one company making a strange decision. Enterprise AI software prices are up 20 to 37% on renewal across the board, according to Tropic’s procurement data from March 2026. Anthropic, OpenAI, and Google have all raised prices. GitHub is killing annual plans entirely.</p><p>Frontier AI models have been priced at a loss since the beginning of the current cycle. The hyperscalers financing them were buying market share and training moats, not running a sustainable business. It worked. Enterprises built workflows on those prices, hired around them, and assumed costs would keep falling.</p><p>Inference costs for standard tasks did fall over the past two years. But frontier model pricing was never on a real downward trajectory. It was suppressed while the market was being captured.</p><p>That suppression is ending. The floor is becoming the ceiling.</p><p>And here is the thing that makes this moment particularly sharp. The expensive end of the stack is getting more expensive at the exact time the cheap end is getting dramatically cheaper.</p><p>I think of this as the cost inversion. It is the central dynamic of the next three years. And most organizations have not noticed because they are still looking at last year’s API pricing page.</p><h3>Four tiers, four cost structures</h3><p>There are four tiers of AI deployment available right now. They have different economics, different capability ceilings, and different ROI profiles. Understanding the difference is the most important strategic decision in enterprise AI in 2026. I call the framework The Intelligence Stack.</p><p>On-device is where the volume goes. Apple committed to this at scale during WWDC this week, rebuilding its entire AI architecture around a System Orchestrator that routes simple tasks to on-device models running on Apple Silicon and sends only complex requests to the cloud. When iOS 27 ships in September, that architecture reaches 2.5 billion active devices. On the open hardware side, NVIDIA’s RTX Spark arrives this autumn with a Blackwell-class GPU and 128GB of unified memory inside a laptop. The on-device tier is no longer a compromise for lightweight tasks. It is serious compute at zero marginal cost per query.</p><p>On-premise is the quiet winner for regulated industries. Llama 3.1 70B, Mistral Large, fine-tuned open source variants running on your own infrastructure. No per-token billing. Full data sovereignty. The cost per useful output keeps dropping as the infrastructure investment spreads across more workloads. If your organization has compliance constraints, this tier gets more attractive every quarter.</p><p>Cloud mid-tier is the workhorse nobody talks about. Claude Haiku, GPT-4o mini, Gemini Flash. Scalable, predictable cost, strong enough for most enterprise tasks. Proposal drafting. Report generation. Code completion. Structured data extraction. This is where the ROI is clearest right now, and most organizations are barely using it because they jumped straight to the top.</p><p>Frontier cloud is where most of the money is going and most of the waste is happening. Claude Opus 4, GPT o3, Gemini Ultra. Extraordinary reasoning capability. Extraordinary cost. The right tier for complex legal analysis, novel engineering problems, multi-step research that a cheaper model genuinely cannot do. But the gap between frontier and mid-tier has to be material to the outcome. If it is not, you are paying for horsepower you are burning in traffic.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H8tD77yWHSOnQ2IxzNJJqg.png" /></figure><h3>The cost inversion</h3><p>Look at where the four tiers are heading.</p><p>Frontier pricing is at its lowest relative point right now. The subsidy window is still open, barely. GitHub’s shift, the 20 to 37% renewal increases, the Anthropic and OpenAI price adjustments: these are early signals of a normalization that accelerates from here.</p><p>At the bottom of the stack, costs are moving in the opposite direction. Apple is routing routine AI work on-device by default. NVIDIA is putting Blackwell-class inference into laptops. On-premise infrastructure amortizes steadily downward. The lower tiers get structurally cheaper every month.</p><p>The lines on the chart are crossing. Organizations that have already built tiered routing are on the right side of the inversion. The ones running everything through Tier 4 are on the wrong side and the gap is widening.</p><p>Uber burned through its full 2026 AI budget in four months. Microsoft cancelled Claude Code licenses across the company before the fiscal year closed. Every headline called these cost blowups. They were not. They were routing failures. Both deployed frontier intelligence on standard coding workflows and documentation. Tasks that belonged two or three tiers lower.</p><p>They were not outliers. They were the first organizations large enough and visible enough that the budget failure became public. The same misconfiguration is happening at hundreds of companies that are not making headlines yet.</p><h3>Where human judgment still compounds</h3><p>There is a second dimension to the routing decision that gets less attention. It is not just about which model tier. It is about where human judgment is still the thing that matters.</p><p>Sorting email, classifying documents, scheduling, compliance checks. High-volume, low-judgment tasks. They belong on Tier 1 or Tier 2. No human in the loop. Near-zero cost.</p><p>Board decisions, novel product strategy, a difficult conversation with a key client. Low volume, high stakes, high judgment. A model can draft the brief. The human makes the call.</p><p>The zone I find most interesting is where volume and judgment are both high. Portfolio analysis across hundreds of positions. Legal review across a thousand contracts. Governance at scale inside a hospital system. This is where human insight scales through AI reach, and where the compounding gets real. The human sets the direction. The agent executes. Neither gets there alone.</p><p>The mistake runs both ways. Deploying agents where you need human judgment gets you fast, wrong answers. Deploying humans on work that should have been automated years ago is the most expensive routing mistake in the stack. We just could not see it before agents made it obvious.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7Va0B_iKtcykpvk_-Fj0qw.png" /></figure><h3>The third cost pressure</h3><p>Two cost corrections are already underway. Subsidies are normalizing. Routing misallocation is burning budgets faster than anyone planned.</p><p>A third one is coming.</p><p>Governments will tax agent deployment. The logic is not complicated. Income tax and payroll contributions assume that economic output flows through people. When agents absorb that work, the revenue base erodes. The fiscal response is an agent levy of some kind. My hunch is that the timing is closer than most people expect.</p><p>When it arrives, cloud-tier costs shift upward again. On-device and on-premise sit below the threshold.</p><p>One more thing worth flagging. Hyperscaler AI infrastructure spend is approaching $650 billion in 2026 and the full supply chain number trends toward a trillion. All of that investment assumes a world where most AI work runs through centralized cloud compute. But Apple just committed 2.5 billion devices to on-device-first architecture. NVIDIA put Blackwell-class inference in a laptop. If the volume actually migrates to local tiers, the demand assumptions behind the biggest capex cycle in technology history are overstated.</p><p>I want to be clear that this is a hypothesis. But the economics and the architecture are both pointing the same way.</p><h3>The routing decision</h3><p>The subsidy window is closing. The cost lines are crossing. And the organizations that figured out the routing decision early are pulling away from the ones that did not.</p><p>This is not a technical question for the infrastructure team. It is the most consequential financial decision of the AI era.</p><p>The organizations getting this right are building a cost advantage that compounds as pricing normalizes. The ones running everything through Tier 4 are building on a foundation that is already cracking.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ecbc512ea59e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Winner Takes Next]]></title>
            <link>https://medium.com/@david_46352/winner-takes-next-34934b95190b?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/34934b95190b</guid>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[go-to-market-strategy]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Mon, 01 Jun 2026 15:12:04 GMT</pubDate>
            <atom:updated>2026-06-04T07:57:14.131Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2ThSR5dmyMC4iSB77DGA3w.png" /></figure><p>Sword Health used to be a musculoskeletal therapy app. Then it grew into a $3 billion chronic care platform.</p><p>The Portuguese-founded unicorn acquired its biggest European rival Kaia Health for $285 million and is expanding into pelvic health, women’s health, cardiometabolic, and mental wellbeing through Phoenix, its AI care specialist.</p><p>How a single-category company grows into a platform is the most important question in clinical AI right now. And it is being answered, in different forms, by category leaders that earned their position in one specific corner of healthcare and are now using it to take adjacent ones.</p><p>Doctolib is doing the same thing in clinical workflow. The €6.4 billion French company that runs the patient booking infrastructure for roughly 80 million Europeans launched an all-in-one practice management system in Germany, with three AI assistants built into it. A telephone assistant that handles calls around the clock and routes appointments. A consultation assistant that documents the visit in real time. The booking layer it started with, now one feature inside the platform. Germany is the proving ground. France and Italy are next. The pitch to the market is that the operating system of European primary care is up for grabs and Doctolib is building toward owning it.</p><p>Nelly has hovered up a meaningful share of the German dental market and raised €50 million from Cathay Innovation to do it. The patient intake tablet at the front desk has turned into an AI scribe, a treatment plan agent, factoring against private and statutory insurance, and patient payments by smartphone. The intake company is becoming the Financial Operating System of the dental practice.</p><p>Recare is doing the hardest version of the move. Hospital workflow is the most defended territory in healthcare. Procurement cycles run in years, the integration burden is enormous, the political surface inside a hospital is complex enough to kill most vendors before they finish a pilot. Recare won discharge management in roughly two thirds of German hospitals anyway, raised €37 million from DNV, and is now pushing its AI agent up into clinical documentation and down into aftercare coordination. The company that broke into hospital workflow once is using that position to break into adjacent workflows from inside the building.</p><p>Same patient, same workflow, more of the operation, every quarter. Same move. Different entry points.</p><p>This is the move worth paying attention to, and it is happening across the European healthtech landscape right now, in different forms, at different stages, by companies that won one specific thing and used that position to take the next one.</p><h3>The pattern, the chart, the two games</h3><p>The same playbook is running across every care setting on the continent.</p><p>Avelios looked at the legacy hospital substrate, decided it could not be climbed, and is rebuilding it from scratch with a €30 million Sequoia-led Series A. Tandem Health built a cross-border ambient scribe out of Stockholm, reached more than 200,000 NHS professionals through its Accurx partnership, and tens of thousands of clinicians across the UK, Germany, France, Spain, and the Nordics. Kinnevik led the $50 million Series A specifically to fund Tandem’s expansion from documentation into clinical coding, decision support, and care coordination. The scribe is becoming an AI-native operating system. Abridge owns ambient documentation in US clinics and is moving into clinical decision support and revenue cycle. Corti built clinical speech infrastructure in Denmark and is dropping further down to become the model layer that other healthcare AI companies will rent. imito started in hospital image documentation in Switzerland and is moving into wound diagnostics. Hippocratic AI won AI-driven patient engagement at scale, with 115 million patient interactions across 50 health system partnerships and a $3.5 billion valuation. The agents that started by handling pre-op outreach and post-discharge follow-up calls are now expanding into chronic care management and broader care delivery through the company’s new agent app store.</p><p><em>Same move. Different entry points.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c-GtftQBNPr9no_W0wUTnw.png" /></figure><p>Every arrow on the chart is doing one of exactly two things.</p><p>Some arrows expand across care. A company that owns one clinical function reaches into an adjacent one. Abridge into decision support. Doctolib taking three steps at once from scheduling. Tandem climbing from ambient documentation toward clinical coding and decision support. OpenEvidence crossing from the clinician’s desk to the consumer’s phone. Recare spreading along the patient journey in both directions from discharge. Sword expanding across adjacent therapy areas through Phoenix.</p><p>The other arrows become the infrastructure. A company drops out of the clinical surface and tries to own one of the rails everything runs on. Corti becoming the clinical model layer. Nelly taking over the financial operations of the practice. Avelios rebuilding the full-stack substrate so every other hospital AI product has to sit on top of it.</p><p>Two games. The chart shows who is playing which.</p><h3>Why the others died</h3><p>The first wave of healthcare AI was a flood of point solutions. Hundreds of narrow products, each solving one slice of a workflow, each promising to integrate cleanly into whatever clinical system the buyer already had. Most never did. The ones that died were the ones that stayed shallow. They were trapped in their niche, never indispensable to anyone, eventually absorbed or ignored.</p><p>The survivors did one thing differently. They picked one specific problem painful enough to drive deep adoption from day one, and then embedded until the customer could not function without them. Two years on Nelly and you do not switch. Five years on Doctolib and your entire practice is built around the booking flow. A hospital running Recare for discharge has its aftercare coordination flowing through Recare’s rails. Sword’s 600,000-plus active patients are not migrating to another musculoskeletal platform. The depth is the moat. It is also the entry ticket to everything that follows. There is no climbing without first being indispensable somewhere.</p><p>Depth in one category is the only thing that compounds. Breadth bought too early kills you. Depth bought first earns the right to breadth later.</p><p>What B2B SaaS founders call land and expand, European healthtech is now running at scale. The land was the first category. The expand is everything else. The difference is that in clinical AI the categories are not features inside a software product. They are sectors of healthcare, each with its own regulatory surface, its own buyers, its own integration depth. Which is why the move only works for the companies that earned the right by being indispensable in one category first.</p><h3>The deeper moat is becoming the infrastructure</h3><p>Expanding across care is real defensibility. But it is still defensibility inside a feature surface. The company that owns documentation in the clinic competes with the company that owns documentation across the hall. Both are surface plays, both are subject to feature wars, both can be undercut by a sharper product or a cheaper one or an EHR vendor that decides to absorb the category.</p><p>Owning a rail is different.</p><p>When you become the infrastructure, every company that plays on the surface above you has to pay rent in some form. Compute cycles. Transaction fees. A clinical model API call. A slot in your billing flow. Surface companies compete with each other. Infrastructure companies sit underneath the competition and take a cut regardless of who wins.</p><p>This is why the teal arrows on the chart matter more than the purple ones, even though there are fewer of them. Corti is no longer just a documentation product. It is building Europe’s first sovereign healthcare AI infrastructure, with its own clinical speech models, its own data, and partnerships with Wolters Kluwer and SPS Philips that route other vendors’ workflows through Corti’s stack. The CEO confirmed that the company will go public, just not in 2026. The private market, he said, is still appealing.</p><p>Nelly is doing the same move from the other side of the chart. Patient onboarding at the front desk was the surface play. Becoming the Financial Operating System is the infrastructure play. The German billing houses that have served practices for decades have not been disrupted by another billing company. They are being disrupted by a patient onboarding company that earned the right to expand into their territory and is now collecting against statutory insurance reimbursements through the same product.</p><p>Avelios is the most direct version of this move. Rather than climb on top of the legacy hospital substrate, the company decided to rebuild the substrate itself. Every documentation tool, every diagnostic AI, every decision support layer that a hospital running Avelios installs sits on Avelios’s foundation. The surface companies are tenants. Avelios is the landlord.</p><p>That is the second game. It is the bigger one.</p><h3>The contested zone</h3><p>The full-stack hospital layer is where the two games collide head on. It is also the most violent square on the board.</p><p>On one side, Avelios is climbing down from the AI layer to rebuild the substrate as an AI-native platform. On the other side, Epic, Oracle Health, Dedalus, and CompuGroup are doing the opposite move, climbing up out of their system of record positions to absorb the AI categories above them. Epic has roughly 125 generative AI features in active development. Oracle is rebuilding its EHR around an AI-first architecture. Eighty percent of US hospitals already use the AI their EHR vendor offers, against fifty-two percent using third-party AI.</p><p>The doom narrative for healthcare AI startups was that the incumbents would simply absorb everything. That story is half right. The incumbents are not standing still. They are climbing. But the assumption that this kills the startups is wrong. It only kills the shallow ones. The deeply embedded startups are climbing too, from the other end. The full-stack hospital layer will be owned by whoever embedded most deeply first, regardless of whether they started as substrate or surface. The startups that built genuine depth are credible challengers. The startups that did not are absorbed in passing.</p><p>The fact that Sequoia led the Avelios round matters. US tier-one capital is not in the habit of funding European healthcare infrastructure plays. The check is a vote that a credible challenger to Epic and Oracle can come out of Munich. That is a different kind of bet than the funding that flowed into US ambient scribes over the past three years.</p><h3>The playbook is starting to work in Europe?</h3><p>Europe has been working on digital health for years. The list of well-funded national initiatives that never reached commercial scale is long, and anyone in this industry can recite it. So the question is not why Europe is waking up. Europe has been awake. The question is why this particular set of companies is breaking through when so many before them did not.</p><p>The answer sits in the playbook the chart shows. Embed deeply in one category, then climb. The European market punishes shallow point solutions harder than the US market does, because the integration burden is higher, the procurement cycles are longer, and the regulatory edges are sharper. Companies that try to spread early die faster here. Companies that go deep first build moats that are nearly impossible to replicate, because the depth itself is the moat. The German practice that has been on Nelly for two years is not switching. The Danish call center running on Corti is not migrating. The French clinic on Doctolib has built its workflow around the platform. The NHS GPs reaching for Tandem inside their Accurx workflow are not pulling it out.</p><p>That is why international capital is showing up. Sapphire and Sozo into Sword. Kinnevik leading Tandem. Cathay leading Nelly. Sequoia leading Avelios. DNV doubling down on Recare. Doctolib already at €6.4 billion with General Atlantic and Eurazeo. These are not exploratory bets. They are recognitions that the embed-then-climb model works in Europe, and that the companies running it are uncatchable in their home markets in a way American capital cannot replicate by writing bigger checks.</p><p>The market structure that made Europe look slow for years is the same one producing companies with structurally deeper moats than their US peers. The capital is following the moats, not the marketing.</p><p>It is too early to call this a wave. Three years from now we will know whether the late-stage rounds turned into the durable European healthtech players the region has been waiting on, or whether some of these companies stalled at scale. But it is no longer too early to call it a pattern.</p><h3>What this means</h3><p>For founders, the question is no longer which category you are in. It is whether you have embedded deeply enough in one category to earn a move, and which of the two games you are playing when you make that move. A company with no credible next move is a point solution waiting to be absorbed.</p><p>For investors, the arrows are the diligence. A company expanding across care from a position of depth is a compounder. A company moving toward becoming the infrastructure is a potential category-shaper. A company that has not made a credible move, or cannot, is a feature waiting to be replaced.</p><p>The winners in clinical AI are not the ones who tried to take everything at once. They are the ones who concentrated their entire company on a single category, embedded so deeply that nobody else could fit in the same space, and only then started moving.</p><p>This is not a strategy question anymore. It is an arithmetic one. The companies that won a category have buyers who will not switch, integrations that took years to install, and political capital inside their customer’s organization that any competitor would need a decade and a fortune to replicate. From that position, the next category costs them a fraction of what it costs anyone starting from outside. From that position, becoming the infrastructure underneath the category is not an ambition. It is a year of product work and a credible distribution channel they already own.</p><p>The companies that did not win a category are running the opposite math. Every adjacent move they make is a fresh customer acquisition, a fresh integration, a fresh political fight inside the building. They are competing for the same expansion territory from a worse starting position. The cost compounds against them, just as the embedded leader’s compounds for them.</p><p>This is why the chart’s silences matter as much as its arrows. Every uncrowned cell is a category still in play. Every crowned cell without an arrow is a leader that has not yet started to move, which is a window that will close. Every company not on the chart at all is, in this market, increasingly hard to fund.</p><p>The next three years will not produce a thousand category-defining clinical AI companies. They will produce a dozen, and that dozen will overwhelmingly come from the leaders already embedded somewhere on this chart, or one very much like it. The race for the first category is mostly over. The race for the second one is the entire game.</p><p><strong>Winner takes next.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34934b95190b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Clinical API Layer]]></title>
            <link>https://medium.com/@david_46352/the-clinical-api-layer-1f844685e5b1?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/1f844685e5b1</guid>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Tue, 19 May 2026 12:35:33 GMT</pubDate>
            <atom:updated>2026-05-19T12:35:33.159Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PsQ9X72_VcPocGndXN-2-w.png" /></figure><h4>Regulated medical intelligence is becoming infrastructure.</h4><p>Corti does not run a hospital, does not sell an EHR, and most patients will never know the company exists. The Danish startup packages medical speech recognition, clinical coding, and documentation systems into APIs and SDKs that other healthcare companies embed into their own products. Telemedicine platforms integrate the transcription. Health systems integrate the ambient documentation. Developers integrate the coding assistance.</p><p>Corti is not trying to own the healthcare interface. It is trying to become part of the layer underneath it.</p><p>That distinction matters because a growing number of healthcare AI companies are making the same move. They are not selling standalone software so much as selling clinically validated intelligence as infrastructure that other people build on top of. At first glance, this looks like another wave of healthcare AI tooling, and the industry produces hundreds of those every year. But something more structural is happening underneath the surface noise.</p><p>Medical intelligence is separating from the application layer. The companies creating the intelligence increasingly do not own the customer relationship, and the companies owning the customer relationship increasingly do not build the intelligence. That separation, more than any individual product launch, is what is changing the economics of healthcare software.</p><h3>How the old model worked</h3><p>For most of modern healthcare IT, shipping a new clinical capability meant assembling three very different things yourself. You built the model. You secured the regulatory clearance. You built the distribution. Each of those efforts required different teams, different timelines, and different capital structures. The AI team moved at a different pace than the regulatory team. The regulatory team moved at a different pace than enterprise sales. Enterprise sales moved at a different pace than hospital procurement. That fragmentation shaped how healthcare innovation worked, and how slowly it worked, for years.</p><p>Those layers are now starting to collapse into a single product. A company can package the intelligence, the validation, the CE mark or FDA clearance, and the API or SDK, and sell the entire thing as infrastructure that someone else integrates. That changes who gets to build, and what they can build with.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Cf3iZMq3_uS4lYZpUQ0n8A.png" /></figure><p>A pharmacy app no longer needs to become a dermatology company in order to offer skin assessment. A hospital platform does not need an internal computer vision team in order to ship wound analysis. An EHR vendor does not need to spend five years building ambient documentation models from scratch. They can rent the capability instead, and increasingly they do.</p><h3>A market that has been forming for a decade</h3><p>The pieces of this market have existed for years, just never as a coherent layer. Infermedica started building diagnostic systems more than a decade ago. Autoderm exposed dermatology APIs before most people in healthcare cared about foundation models. Imaging marketplaces have been assembling parts of this ecosystem for the better part of the last five years. Until recently, though, these companies felt like isolated exceptions. They are starting to feel like members of the same stack.</p><p>The shift here is not just that healthcare AI exists. The shift is that clinically trusted intelligence is becoming modular, in a way that almost nothing else in healthcare software has been. Once you see the market through that lens, a lot of recent activity starts looking different. Suki embedding inside athenahealth, Autoderm powering skin assessment inside Boots and DocMorris, imito running inside Dedalus, clinical knowledge systems exposing APIs and MCP connectors for AI agents to call. Different products, different buyers, same underlying structural move.</p><h3>Who actually benefits</h3><p>The obvious answer is startups. Small teams can now ship clinically useful products without building every layer themselves, and infrastructure that once required years of development can increasingly be rented from companies that have already done the hard work. That story is true, but it is not the most important consequence of what is happening.</p><p>The larger shift is happening inside incumbents.</p><p>Healthcare incumbents were never structurally designed to become frontier AI labs. Dedalus was never going to train dermatology models. Oracle Health was never going to build speech foundation models. Epic was never going to assemble a computer vision research organization. These are companies that optimize around reliability, around integrations, around procurement cycles, around compliance, and around workflow continuity. Until recently, that orientation looked like a disadvantage in a market that suddenly seemed to be all about AI. It may now become an advantage again, because the intelligence layer no longer needs to be built internally. It can simply be integrated.</p><p>That changes the innovation cycle of healthcare incumbents in a way that is easy to underestimate. For years, healthcare software felt technologically frozen because every meaningful expansion required building genuinely new capability from scratch, on a clock and budget that the underlying business could not really support. Now incumbents can absorb innovation externally while keeping the distribution, the workflow ownership, and the enterprise relationships that took them decades to build. The distribution layer stays where it already is. The intelligence layer becomes modular. That, more than any single product, is the deeper shift.</p><h3>The hyperscalers arrive</h3><p>Even the recent hyperscaler healthcare push starts making more sense once viewed through this framework. Within a single quarter this year, OpenAI launched ChatGPT Health, Anthropic launched Claude for Healthcare, and AWS launched Amazon Connect Health. All three entered with compliant infrastructure, with EHR connectivity already wired in, and with enterprise healthcare customers already attached to the launch. Nine weeks, three of the most capitalized AI companies in the world, all entering the same layer of the same market.</p><p>That changes the floor of the market almost immediately. Generic compliant inference, which was an interesting capability eighteen months ago, is becoming abundant very quickly. Which means the moat moves elsewhere, away from things that are about to be commodities and toward things that are harder to replicate.</p><p>Not necessarily into the model itself. Not even permanently into regulatory clearance. Those advantages still matter, and they remain meaningful for verticals with deep clinical validation requirements like cardiology or radiology, where five years of FDA work and large patient validation studies cannot be conjured into existence by a frontier lab. But across the broader market, more and more clinically validated systems will come online every quarter, and the wrapper itself stops being the source of differentiation.</p><p>Defensibility starts moving outward. Into the data that accumulates from real clinical deployments at scale. Into the depth of workflow integration that took years to earn. Into clinician trust that no API can shortcut. Into enterprise distribution. Into the embedded position inside systems already in use by the people you want as customers. The companies with the strongest position are not always the ones with the best models. Often they are the ones already sitting inside the workflow.</p><p>The result is an unusual balance of power. Startups can build faster than they could before. Incumbents can evolve faster than they could before. Hyperscalers compress the value of generic infrastructure underneath both groups. The market that emerges from this is one where intelligence becomes increasingly modular while distribution remains highly concentrated.</p><h3>A pattern healthcare has seen before</h3><p>Healthcare has seen versions of this pattern before, just not for clinical intelligence itself. Cloud infrastructure centralized compute. App stores centralized distribution. EHRs centralized the clinical workflow. Each of those shifts looked, in the moment, like a story about a particular product or company, and only became visible as a structural shift in retrospect. This feels similar, except the thing being centralized this time is regulated intelligence.</p><p>And the category still does not really have a name. People call it healthcare AI infrastructure, or clinical AI APIs, or AI as a service. None of those phrases quite capture what makes this layer structurally different from the wave of AI tooling that came before it. The defining characteristic is the bundle, not any single component of it. The intelligence, the validation, the regulatory wrapper, and the access layer, delivered together as something another company can directly embed into its own product. Sold once, integrated once, governed once.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9SOuO4o4qVIr-iTNqzZucQ.png" /></figure><p>The most suitable name for this bundle is the Clinical API Layer. Not because APIs are the important part. Because the API is the delivery mechanism for something healthcare has historically struggled to modularize at all. Clinically trusted intelligence is what is finally becoming programmable, and that is the shift worth paying attention to.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1f844685e5b1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Black Box Just Opened]]></title>
            <link>https://medium.com/@david_46352/the-black-box-just-opened-78c7d23d96a0?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/78c7d23d96a0</guid>
            <category><![CDATA[brain-training]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[alzheimers]]></category>
            <category><![CDATA[drug-development]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Thu, 12 Mar 2026 06:11:00 GMT</pubDate>
            <atom:updated>2026-03-12T06:11:00.616Z</atom:updated>
            <content:encoded><![CDATA[<h4>How whole brain emulation changes what we thought we knew about human potential</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y2Qqms4XKtqP7zeVnR8g1g.png" /></figure><p>Last week a team of scientists copied the entire brain of a fruit fly into a computer. Every neuron, every connection, 125,000 cells and 50 million synapses. Then they dropped it into a simulated body.</p><p>It walked. It ate. It groomed itself.</p><p>Nobody programmed any of that. The wiring was enough.</p><p>I’ve been thinking about this since I saw it, and I can’t stop. Not because of the science fiction implications everyone else is reaching for. But because of something quieter and more personal. For the first time in the history of medicine, we didn’t observe a brain from the outside and guess what was happening inside. We built one from the blueprint and watched it run.</p><p>That’s new. And I’m not sure we’ve understood yet what it changes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*S1Gkk5gPS4-IpiIV" /></figure><h3>What We’ve Been Missing</h3><p>Every tool we’ve ever had for studying the brain has been a tool for watching it from the outside.</p><p>We’ve mapped regions. We’ve measured electrical activity. We’ve observed what happens when parts of it go wrong. We’ve built statistical models that approximate how it behaves across large populations. And we’ve used all of that to develop treatments for the diseases that break it.</p><p>The results have been sobering. Neurological disease sits at the center of medicine’s longest and most expensive failure streak. The diseases where we understand the brain least are the diseases where we fail most often. Alzheimer’s. Parkinson’s. ALS. Decades of research, billions in investment, and the fundamental problem has remained the same: you cannot reliably fix something you cannot accurately replicate.</p><p>What Eon Systems demonstrated is structurally different from everything that came before it. Previous attempts to model brains either mapped the wiring without giving it a body to run in, or built bodies controlled by machine learning rather than actual biological architecture. This was the first time a complete biological connectome, copied neuron by neuron from electron microscopy data, was placed in a simulated body and produced behavior through its own dynamics.</p><p>The fly wasn’t trained. It wasn’t programmed. The architecture generated the behavior.</p><p>That distinction matters more than it might seem. It means the wiring itself carries the intelligence. Not the data the brain was exposed to. Not the patterns it learned over time. The structure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lZ64feRGCxMF3cpoMZpsUw.png" /></figure><h3>The Mouse Is Next</h3><p>Eon’s roadmap is explicit. The fruit fly is the proof of concept. The mouse brain is next, at 70 million neurons, roughly 560 times the complexity of the fly. The mouse is also the standard preclinical model sitting at the center of almost every neurological drug trial running today.</p><p>When a digital mouse clears the same threshold, the trial pipeline changes. You run your experiment in simulation before a single animal is involved. Unlimited cohorts. No variability between subjects. You induce the disease state on demand, test your intervention a thousand times, and arrive at the physical trial already knowing what you’re looking for.</p><p>The 90% failure rate in CNS drug development is not inevitable. A significant part of it is a modeling problem. We’ve been testing interventions on biological systems we couldn’t fully replicate, which means we’ve been flying partially blind at the most expensive stage of the process.</p><p>That starts to change. Not immediately, not completely, but directionally and irreversibly.</p><h3>The Question I Keep Coming Back To</h3><p>Here’s where I want to be honest about what’s actually driving my thinking.</p><p>I work in healthcare and healthtech. I build and back companies in this space. The drug development implications matter to me professionally. But that’s not why I haven’t been able to put this down.</p><p>What I keep thinking about is my own brain.</p><p>Not a diseased one. The one I’m running right now, today, as I write this. The one that will be running in ten years, twenty years, thirty years, if things go reasonably well.</p><p>We talk about cognitive decline as though it’s inevitable. A slow erosion that begins somewhere in middle age and proceeds in one direction. We accept it the way we accept other facts of aging, as a thing that happens to us rather than a process we might understand and influence.</p><p>But if the architecture generates the behavior, and we can now read the architecture with genuine precision for the first time, the question changes.</p><p>It stops being: how do we slow the decline?</p><p>It becomes: what is this brain actually configured to do, and are we anywhere close to running it at its potential?</p><p>I’m not talking about nootropics or biohacking or any of the optimization culture that has attached itself to this space. I’m talking about something more fundamental. The gap between the brain you have and the brain you’re using might not be a motivation problem or a habit problem. It might be a configuration problem. And configuration, unlike willpower, is something you can actually study.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rKfTS_QqHkEWVUvO7oJ0Gg.png" /></figure><h3>What Becomes Visible</h3><p>Once you can run a brain in simulation and watch the architecture operate, a set of questions becomes answerable that weren’t before.</p><p>Why do certain patterns of thought persist even when we consciously want them to change? What’s running in the background during decision making that we have no conscious access to? Why do some people maintain sharpness and storage capacity deep into old age while others experience significant decline on a similar timeline?</p><p>These aren’t philosophical questions anymore. They’re architectural ones.</p><p>The implications for how we think about learning, memory, attention, and cognitive maintenance across a lifetime are significant. Not in the sense that we’ll all be uploading ourselves to servers by 2040. But in the sense that the model we’ve been using, the brain as a black box we observe from the outside and treat when it breaks, is being replaced by something with much higher resolution.</p><p>For the first time, we might actually be able to see what’s happening in there.</p><h3>The Uncomfortable Question</h3><p>Every major shift in how we understand biology has revealed that we were modeling something wrong.</p><p>We thought the genome would explain disease. It explained less than we hoped and opened up an entirely new layer of complexity beneath it. We thought neural networks modeled on the brain would behave like the brain. They didn’t, and understanding why taught us things about both.</p><p>Now we can run the brain itself, not a statistical approximation of it, not a machine learning policy trained to mimic it, but a structural copy operating through its own dynamics.</p><p>If that changes what we understand about disease, about treatment, about the mechanisms behind neurological failure, it will almost certainly also change what we understand about the healthy brain operating within its normal range.</p><p>Which raises the question I can’t answer but can’t stop asking.</p><p>If we’ve been modeling the brain wrong this whole time, what else have we gotten wrong about what we’re capable of?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_hXerJVjQEcr8EHhMVqoOQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=78c7d23d96a0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Rocket Engine That Broke the R&D Sound Barrier]]></title>
            <link>https://medium.com/@david_46352/the-rocket-engine-that-broke-the-r-d-sound-barrier-8cc1891ff8f9?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/8cc1891ff8f9</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[industrial-design]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Fri, 13 Feb 2026 09:18:50 GMT</pubDate>
            <atom:updated>2026-02-15T13:27:50.822Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/780/1*yMUrt_57eMOamdR0xQCKJg.png" /><figcaption>Cross-section Leap71 rocket engine</figcaption></figure><p>It looks like a fossilized nervous system.</p><p>Copper-alloy skin. Swirling organic ridges. No straight lines. No symmetry. No trace of a human hand.</p><p>If you found it in a lab, you wouldn’t ask what machine it belongs to. You’d ask what planet it came from.</p><p>This object, the LEAP 71 rocket engine, was ignited for the first time in mid-2024. Cryogenic oxygen. Kerosene. Full thrust.</p><p>It worked. First attempt.</p><p>And no engineer drew it.</p><p>Not a general-purpose AI chatbot. A computational engineering model called Noyron, built by a company called LEAP 71, encoding physics, material science, and manufacturing constraints into a single deterministic system. No sketches. No CAD. No human spatial reasoning involved in the geometry at all.</p><p>This isn’t just an interesting engineering story. It’s evidence that we’ve crossed a threshold most people haven’t noticed yet — the R&amp;D sound barrier — the point where complexity and physics finally outpace what human cognition can manage alone.</p><h3>The Two-Week Engine</h3><p>Here’s what actually happened.</p><p>A company called LEAP 71 fed an AI system called Noyron a single input: the laws of physics it had to obey. Thrust requirements. Material limits. Thermal constraints. Then they let it run.</p><p>Noyron didn’t sketch. It didn’t iterate on a CAD screen. It grew a solution, treating thermal dynamics, structural integrity, and fluid flow as one unified mathematical problem instead of three separate departments arguing in a meeting room.</p><p>From blank screen to successful hot-fire test: less than two weeks.</p><p>A comparable NASA development cycle: years.</p><p>That gap is not a productivity improvement. It’s a category change. When Chuck Yeager broke the sound barrier in 1947, he didn’t fly slightly faster than before — he entered a regime where the old rules no longer applied. LEAP 71 just did the same thing to R&amp;D.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F6Xx1GXjRbMk%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D6Xx1GXjRbMk&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F6Xx1GXjRbMk%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/fe11f7c068bcb3e17e4021253f6955b9/href">https://medium.com/media/fe11f7c068bcb3e17e4021253f6955b9/href</a></iframe><h3>The Prison We Didn’t Know We Were In</h3><p>For 50 years, industrial R&amp;D has been trapped inside what I call the CAD Prison.</p><p>Not trapped by the laws of physics. Trapped by what a human brain can visualize in 3D software.</p><p>Engineers favor 90-degree angles, straight lines, symmetry. Not because these shapes are optimal — but because our brains and our legacy tools required them. When a single prototype mistake costs $100M and three years of your career, “radical” stops being a design philosophy and starts being a career-ending risk.</p><p>So we refined. We iterated. We moved the ceiling up one millimeter at a time.</p><p>This is what I’d call the Incremental Era — 50 years of steady, cautious, millimeter-by-millimeter progress. Valuable, yes. But fundamentally limited by the ceiling of human spatial cognition. We were always approaching the sound barrier. We just couldn’t break through it.</p><p>The LEAP 71 engine is proof that era is over.</p><h3>Why AI Is Not What You Think It Is</h3><p>Most companies are using AI to write emails faster and summarize meeting notes. That’s fine. But it completely misses what this technology actually is.</p><p>AI is not a productivity tool. It is a discovery engine.</p><p>The difference matters enormously. A productivity tool makes you faster inside the existing system. A discovery engine breaks you out of the system entirely — it operates in the regime beyond the sound barrier, where complexity stops being a constraint and becomes a free variable.</p><p>When the cost of testing a genuinely crazy idea drops to the price of electricity and a 3D print, something fundamental shifts. If a fractal, spiraling cooling channel is 1% more efficient than a straight one, the AI grows it. It doesn’t matter that it looks like something dredged up from a deep-sea trench. Physics don’t care about aesthetics.</p><p>This isn’t incremental improvement. It’s what comes after incremental improvement ends.</p><h3>The Engineer Isn’t Dead — The Engineer Is Upgraded</h3><p>I want to be clear about what this doesn’t mean.</p><p>It doesn’t mean engineers are obsolete. It means the job has changed in a way most people haven’t caught up to yet.</p><p>We’re moving from being the Builders to being what I’d call Architects of Intent. The human sets the constraints — the desired outcome, the material limits, the ethical lines that can’t be crossed. The AI explores the solution space, running through variations that would take a human team decades to test physically. Then the human chooses — the path that makes sense for safety, for society, for the actual mission.</p><p>The sound barrier analogy holds here too. Breaking it didn’t make pilots obsolete. It made them responsible for a fundamentally different kind of flight.</p><p>We aren’t losing engineering jobs. We’re turning engineers into people who can direct the laws of physics with a well-crafted brief. People no longer limited by what they can draw, only by the laws of physics they choose to encode.</p><p>That’s not a downgrade.</p><h3>Where This Is Already Happening</h3><p>The same logic cooling a rocket combustion chamber is being applied right now in other fields.</p><p>Healthcare is my industry and the one I keep returning to.</p><p>Bone implants are already being grown — not designed — with internal lattice structures tuned to a specific patient’s bone density. Geometries no human could draft by hand, optimized for how that individual heals. The AI wasn’t handed a template to improve. It was handed a problem — this patient, this material, this outcome — and it found a solution that was always physically possible but never within reach of a human team working the traditional way.</p><p>The same logic applies anywhere the constraints are known but the solution space is too large for human teams to search. Drug discovery. Device architecture. Delivery mechanisms. The biology isn’t missing. The search engine was.</p><p>In climate tech, carbon capture membranes are being grown with fractal surface areas far more efficient than current designs. In energy, fusion reactor components are being engineered with cooling geometries that handle heat loads that would destroy anything built with human-scale symmetry.</p><p>Each of these is the same story: a solution that was always physically possible, but cognitively unreachable — sitting on the other side of the sound barrier, waiting.</p><p>The “alien” aesthetic isn’t a quirk. It’s what answers look like when they’re no longer filtered through human spatial limits.</p><h3>The Real Question for the Next Few Years</h3><p>Here’s what I keep coming back to.</p><p>We’ve spent 50 years in the Incremental Era, moving carefully and slowly because the cost of being wrong was enormous. That era shaped how R&amp;D organizations are built, how they’re funded, how they think. Most of those organizations don’t yet realize the constraint that shaped them no longer exists.</p><p>The companies that define the next decade won’t be the ones with the most engineers. They’ll be the ones that figure out fastest how to operate as Architects of Intent — how to set the right constraints, ask the right questions, and trust the physics to find the answer.</p><p>By 2030, I think first-attempt hardware success will be an expectation, not a headline.</p><p>The LEAP 71 engine didn’t just pass a test. It proved that the sound barrier we’d been creeping toward for half a century was never a wall.</p><p>It was a door.</p><p>The LEAP 71 engine didn’t just pass a test. It proved that the sound barrier we’d been creeping toward for half a century was never a wall.</p><p>It was a door.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AQbxe26t968-oNY507K0rA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8cc1891ff8f9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Death of the Health Dashboard: Why Apple Went “Headless”]]></title>
            <link>https://medium.com/@david_46352/the-death-of-the-health-dashboard-why-apple-went-headless-b6dc386600ae?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/b6dc386600ae</guid>
            <category><![CDATA[health]]></category>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[business-strategy]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Mon, 09 Feb 2026 12:17:20 GMT</pubDate>
            <atom:updated>2026-02-15T13:48:56.207Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kVvVK_idLeKJ6LbcHQxV2g.png" /></figure><h3>The Death of the Health+ Coach App: Why Apple Went “Headless”</h3><h4>Apple didn’t kill its health+ coach app. It killed the idea that health needs an app at all.</h4><p>For the last decade, the strategy for consumer health tech rested on a single, flawed assumption: that quantification equals action. We believed that if we gave users enough data — tracking every heartbeat and every sleep stage — they would transform into their own clinical directors. Instead, we built massive cognitive load. Users are drowning in fragmented data, forced to play amateur data scientist just to decide if they should skip the gym.</p><p>We built bigger and bigger reservoirs. More sensors. More dashboards. More data to store. The entire industry optimized for capacity when it should have optimized for velocity.</p><p>Apple saw the plateau. And while the market waited for them to launch a standalone “Health Coach” app to unify the data, they did something far more radical. They killed the container entirely.</p><p>Reports of Apple scaling back Project Mulberry are being read as a retreat from health AI. The market is misdiagnosing it. Apple did not abandon the health coach. It killed the container. The standalone app was a structural bottleneck — a vertical silo in a world going horizontal. What Apple is building instead is a Headless Health OS: health intelligence unbundled from any single interface and embedded into the operating system itself.</p><h3>From Hardware to Intelligence</h3><p>This pivot tracks an organizational reshuffle. For years, Apple’s health strategy was an operational masterclass led by Jeff Williams, focused on the hardware sensors and the Watch as a physical hub. The goal was the “Watch as a Doctor.”</p><p>Today, the center of gravity has shifted to Services and Ecosystems under Eddy Cue. By folding the Health and Fitness teams under Dr. Sumbul Desai, Apple is signaling that “Activity” is no longer a separate app vertical. It is a data stream for a unified system. Cue’s Services mindset recognizes that a standalone coaching app like the proposed Mulberry project creates a structural bottleneck. Instead of competing app-for-app with Oura or Whoop, Apple is building the horizontal reasoning engine that those companies must eventually plug into.</p><h3>The Sensor Mesh</h3><p>The first stage of the pivot is the unbundling of the hardware. The Apple Watch is no longer the sole source of truth. It is now part of a distributed sensor mesh.</p><p>The Apple Ring, rumored for a 2026 launch, is the ultimate headless device. It is optimized for 24/7 passive biometrics — sleep and baseline temperature — and because a ring has no screen, it forces the interface to migrate entirely to the OS reasoning engine.</p><p>The more brilliant move is RF sensing in AirPods. Beyond using microphones to track respiratory rate, Apple is reportedly leveraging the radio frequency antennas in AirPods to detect biometrics. By measuring perturbations in the electromagnetic field around the ear, Apple can detect heart rate and stress-induced micro-gestures without adding a single new piece of hardware. The sensor is already in your pocket. Apple is unlocking the intelligence through software.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/825/1*76NUZNljfooATfmt8n7Xzw.jpeg" /></figure><h3>The Reasoning Engine</h3><p>If the dashboard is dead, where does the insight go? It moves into the horizontal intelligence layer. The true “head” of this headless strategy is Project Campos — the ground-up rebuild of Siri for iOS 27 using LLM-based infrastructure.</p><p>Using Assistant Schemas and App Intents, the system moves from answering questions to performing actions. A third-party health app can see that your HRV is low, but it has no idea why. The Headless OS reasoning engine performs a cross-domain logic check.</p><p>Imagine the system sees your HRV trending low and your sleep was fragmented. On its own, that is just data. But the reasoning engine also sees you have a high-stakes meeting at 2 PM. Now it has context. The user is physically depleted and heading into a high-cortisol event. The system triggers a Stress Recovery action — no app opened, no dashboard checked. Only the OS has the system-level permissions to access your calendar, real-time audio markers, and biometric history simultaneously.</p><h3>The Shortest Logic Loop</h3><p>If a user has to “check an app” for a health insight, the window for intervention has already closed. Apple is building the Shortest Logic Loop in the industry.</p><p>By bypassing the dashboard and delivering insights through the surface layer — proactive nudges in the Dynamic Island, haptic whispers on the Watch, audio interventions through AirPods — Apple becomes the trusted health advisor. Contrast this with engagement-hungry apps. A third-party app wants your attention. Apple’s headless advisor wants to give you your time back. The shift is from screen time to system actions.</p><h3>The Module Economy</h3><p>This is the section that matters most if you are building in health right now.</p><p>The era of the generalist dashboard is over. But the era of the specialized health module has just begun. The future of the App Store will be a marketplace of highly specialized intelligence modules — nutrition logic from one startup, metabolic modeling from another, physical therapy recovery protocols from a third. These will not be apps you open. They will be Intent Providers that plug into Apple’s reasoning engine. They will benefit from the platform’s deep access to biometric data, but they will surface their value through Siri’s voice or a system notification.</p><p>This changes the product architecture for every health startup. The moat is no longer the interface. It is the quality of the logic module. Founders who build for Intent Providers — who optimize their intelligence for the reasoning engine rather than for a standalone experience — will own the next layer. Those still building destination apps are building for an era that Apple just declared over.</p><p>The competitive question is no longer “Can you get the user to open your app?” It is “Can your reasoning module earn a permanent slot in the OS?”</p><h3>The OS Wins</h3><p>Apple’s bet is that the most valuable health company is not the one with the best dashboard. It is the one with the Shortest Logic Loop.</p><p>We spent twenty years building bigger reservoirs for health data. The next ten will be about Kinetic Intelligence — moving data into action before the user ever opens a screen. We did not need a better health coach. We needed a better operating system.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b6dc386600ae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Dam Breaks: How Upstream Intelligence Is Finally Unleashing Healthcare Data]]></title>
            <link>https://medium.com/@david_46352/the-dam-breaks-how-upstream-intelligence-is-finally-unleashing-healthcare-data-e27e5f1a4822?source=rss-d7e0085c3ed0------2</link>
            <guid isPermaLink="false">https://medium.com/p/e27e5f1a4822</guid>
            <category><![CDATA[data-lakehouse]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[api]]></category>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[David Schaerf, North Star Labs]]></dc:creator>
            <pubDate>Thu, 29 Jan 2026 11:52:25 GMT</pubDate>
            <atom:updated>2026-01-29T11:52:25.039Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RVH6yfF_VpNNr9qKdTRcSw.jpeg" /></figure><h4>For decades, legacy EHRs have acted as stagnant reservoirs. A new “Double Pincer” strategy — driven by Vertical AI and standardized consumer data — is finally opening the floodgates.</h4><p>For the last twenty years, the strategy for digital transformation in healthcare has centered on a single, monolithic assumption: that we must centralize everything into the Hospital Information System (HIS).</p><p>We poured billions into platforms like Epic, Cerner, and Orbis, effectively building massive digital dams. We treated data as something to be trapped, stored, and guarded.</p><p>The result? The EHR became a deep, stagnant reservoir of Health IT. While it is excellent at holding data for billing and compliance, it was never designed to let data flow. Once clinical insights enter the reservoir, they settle into sediment — unstructured text blobs and PDFs — trapped behind vendor firewalls and expensive interface fees.</p><p>But the pressure has finally built up. We are no longer waiting for the legacy vendors to open the cracks. We are building a bypass.</p><p>A structural shift is underway. A “Double Pincer” movement — driven by Vertical AI (from the provider side) and Standardized Consumer Protocols (from the patient side) — is decoupling intelligence from storage.</p><p>Here is how the dam breaks.</p><p><strong>The Problem: The Cost of Stagnation</strong></p><p>The fundamental limitation of the current architecture is Data Gravity.</p><p>In the legacy model, innovation is forced to happen inside the reservoir. If you want to deploy a sepsis prediction model or a patient recovery tracker, you have to ask the vendor for permission to access your own water.</p><p>This has created an ecosystem of “stagnation by design.” Third-party innovators are starved of data, and clinicians are drowning in administrative tasks because the system is designed for storage, not for kinetic action.</p><p>For years, we waited for regulation to force the gates open. But while we waited for policy, Technology found a new path.</p><p><strong>Trend 1: The Upstream Bypass (Vertical AI)</strong></p><p>The first channel of the bypass is Vertical Intelligence. We are now able to capture clinical data upstream — at the very source of the river (the patient-doctor interaction) — before it ever reaches the EHR’s gravity well.</p><p>The Arsenal: Capturing the Flow Modern AI point solutions are no longer just “scribes”; they are independent data generators.</p><p>Ambient AI (e.g., Nabla, Suki, Nuance DAX, Voize): These tools capture the raw conversation. Instead of just dumping a text note into the EHR, they extract structured metadata: specific Medications, Allergies, and Social Determinants of Health (SDoH) in real-time.</p><p>Radiology &amp; Pathology AI: These tools analyze images before they are filed away, generating discrete data points — tumor margins, fracture classifications, and quantitative density metrics.</p><p>The Strategic Pivot: Forward-thinking organizations are “forking the river.”</p><p><strong>Stream A (Compliance): </strong>The narrative PDF flows into the EHR reservoir to satisfy legal requirements.</p><p><strong>Stream B (Innovation):</strong> Rich, structured data flows directly into the hospital’s own Open Intelligence Layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PgPpHMNB3NkZiuPML9NaTQ.jpeg" /></figure><p><strong>Trend 2: The Downstream Flood (Standardized Consumer Data)</strong></p><p>The second channel comes from the patient side. For years, patient-generated data was a trickle of isolated, proprietary streams. But recent moves by tech giants, most notably Apple Health’s adoption of FHIR R4, have turned this trickle into a flood.</p><p><strong>The Arsenal: </strong>Real-World Evidence We are moving from “step counting” to clinical-grade streams:</p><p><strong>Continuous Vitals: </strong>Resting heart rate, HRV trends, and oxygen saturation.</p><p><strong>Digital Biomarkers:</strong> Walking asymmetry and double support time (critical for post-op recovery).</p><p>Because this data is now standardized (FHIR), it doesn’t need to be shoehorned into an EHR that wasn’t built for high-frequency streams. It flows naturally into the new Open Intelligence Layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gFQzx30_ISr0dqDYjhIKUQ.jpeg" /></figure><p><strong>The New Architecture: The “Clinical Lakehouse”</strong></p><p>Where does this diverted water go? It flows into a new architectural construct: the Clinical Data Lakehouse.</p><p>Think of this as the “Switzerland” of your data architecture — neutral, open, and owned by you. An active, controlled ecosystem that sits above the stagnant EHR reservoir.</p><p>The Structure: It utilizes open standards (PostgreSQL, Delta Lake, FHIR Stores) that you own and control.</p><p>The Function: This is where the “System of Intelligence” lives. It is where you run Python models, population health analytics, and agentic workflows — permissionless and fast.</p><p><strong>The Payoff: “Horizontal” Care</strong></p><p>When you combine Upstream Clinical Data and Downstream Consumer Data in this neutral Lakehouse, you unlock use cases that were previously impossible because the data was trapped in separate silos.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GjfJTqgEhy3hTlzCQwTDrg.jpeg" /></figure><p><strong>Use Case 1: The Orthopedic “Safety Net”</strong></p><p>The Old Way: A patient gets knee surgery. The surgeon writes a note. The patient goes home. The surgeon sees them 6 weeks later. If complications arise in week 2, nobody knows because the data points are disconnected.</p><p>The New Way:</p><ol><li>AI Scribe captures the specific surgical technique used (structured data).</li><li>Apple Health streams “Walking Asymmetry” data post-discharge to the Lakehouse.</li><li>The Insight: An algorithm in the Lakehouse detects a 20% spike in asymmetry on Day 4. It cross-references the surgical technique and flags the care team immediately, preventing a readmission.</li></ol><p><strong>Use Case 2: The Metabolic “Radar”</strong></p><p>The Old Way: A diabetic patient visits every 3 months. Decisions are made on 4 data points a year.</p><p>The New Way:</p><ol><li>Consumer Devices stream continuous glucose and sleep data.</li><li>Clinical AI analyzes the doctor’s previous voice notes to extract “patient reported barriers” (e.g., “struggles with cooking”).</li><li>The Insight: The system correlates poor sleep with glucose spikes and identifies the lifestyle barrier, triggering a nudge from a health coach agent — all without the doctor lifting a finger.</li></ol><p><strong>Conclusion: Prioritize Flow Over Storage</strong></p><p>The difference between a reservoir and a river is movement.</p><p>The “Dependency” on legacy vendors was actually a choice. We chose to prioritize storage (The Dam) over agility. The future belongs to those who prioritize flow.</p><p>The Double Pincer strategy allows healthcare leaders to tap into the kinetic energy of their data. By building an upstream bypass and an open Lakehouse, you ensure that innovation flows continuously — regardless of the static concrete block sitting in the center of your IT architecture.</p><blockquote><strong>Strategic Takeaway: Stop staring at the wall. Open the gates.</strong></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e27e5f1a4822" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>