· 24 min LONG READ #AI#Nvidia#strategy#infrastructure#Jensen Huang

The Moat That Cannot Be Coded: Nvidia, Frontier Labs, and Defensibility in the 21st Century

Tacit knowledge, CUDA as institutional memory, and who captures AI value when models commoditise.

TL;DR: in seven lines

  1. Institutional knowledge and relationships matter because they shape what gets observed, which failures are noticed, which experiments are run next and which counterparties will share information in the first place.
  2. Nvidia defends a 75% gross margin three years into the AI build-out, with its lead widening rather than narrowing.
  3. Writing kernels that run is not the same as writing kernels that saturate the hardware.
  4. General models absorb tasks whose context becomes standardised and cheap to obtain.
  5. Talent mobility functions as an immune system that maintains the health of an entire ecosystem.
  6. Innovation does not arrive from outside the institutional knowledge base; it arrives from outside the institutional structure that was holding it captive.
  7. Nvidia combines CUDA and its accumulated engineering practice with a deep-tech metabolism that forces that knowledge to be refreshed every product cycle, and a governance structure centred on Huang that has repeatedly allowed the company to reallocate against its current product.

Andy Grove’s only the paranoid survive has rarely felt truer than today. The question of what holds against displacement haunts everyone in the AI economy: large corporations watching new entrants reshape their markets in months, service providers watching their offerings dissolve into the next foundation model release, startups facing the possibility that their entire product becomes a public feature of a platform, investors trying to underwrite multi-year theses against quarterly capability jumps.

In our earlier framework on AI moats, we put displacement risk at the heart of the problem. The framework laid out the four-level hierarchy that AI companies need to think through: technical foundation, product depth, market positioning, sustainability. We named the major defences: domain delta and velocity dynamics at the technical layer, workflow embedding and execution autonomy at the product layer, capability thresholds and market structure at the positioning layer, and three sanity checks at the sustainability layer. We showed why traditional SaaS valuation tools systematically miss what creates and erodes AI value. And we argued that single moats erode quickly: sustainable advantage requires compound layered defences.

The current paper extends one specific dimension that the framework treated more briefly than it deserved: institutional tacit knowledge and the relationships it travels with. The accumulated practice that lives in the people of an organisation (engineers, commercial teams, operations, regulatory affairs, clinicians, support staff), in the processes they have refined together, and in the trusted relationships they have built with customers, suppliers, and regulators. Much of this is poorly documented and hard to transfer. That matters more as AI makes codified knowledge cheaper, but I would now draw the boundary less sharply than I did when I first wrote this paper. Models can learn from code, video, trajectories, tool use and other traces that capture practices people never wrote down explicitly. Agents can also investigate an environment when they are given access to it.

The more robust asymmetry is therefore not between knowledge that AI can and cannot ever learn. It is between actors who have access to the loop that produces and tests the knowledge, and actors who only see its outputs. Institutional knowledge and relationships matter because they shape what gets observed, which failures are noticed, which experiments are run next and which counterparties will share information in the first place. Nvidia is the most visible illustration.

There is an apparent paradox here that we should name immediately, because it shadows the rest of the paper. If institutional knowledge and relationships make such a strong moat, why does innovation almost always come from the outside? Microsoft did not build OpenAI from within. Google had to chase a lab it largely failed to commercialise. And the broader pattern is older than the AI cycle: even the most innovative tech companies have built much of their current product portfolio through acquisition rather than internal R&D (Alphabet1, Amazon2, Apple3, Meta4, Microsoft5). Across waves and across firms, large incumbents have been refreshed by entrants more often than by their own R&D. We come back to why this happens later in the paper. The short version, which the rest of this piece works towards, is that the paradox is only apparent.

What this paper extends in the existing moat conversation

A short word on where this paper sits relative to the existing moat conversation, and why we are not aiming for exhaustivity. Every general-purpose framework, from Porter’s five forces to Helmer’s 7 Powers to our own AI moats hierarchy, will name slightly different defences in slightly different ways. The same family of canonical objections gets recited and dismissed across all of them:

  1. Proprietary data, but everyone now reasons that scraping and synthetic generation will erode any data moat.
  2. Network effects, but the argument goes that they presuppose a stable user-to-user value loop that AI compresses or bypasses, since a single agent can replace many of the interactions a network used to mediate.
  3. Brand, but the dominant view is that brand is no defence against a step change in capability, and brand itself can be built quickly through aggressive hiring of creative talent and large marketing spend.
  4. Switching costs, but APIs have made switching almost trivial.

The list could go on (scale economies, regulatory barriers, ecosystem effects, etc.). The pattern is the same in every case: a real defence is mentioned, then waved away by a generic argument that often turns out to be too quick. Each of these defences deserves its own dedicated treatment, and we will probably write more of those over time.

The argument of this paper is not that those classical defences have collapsed. It is that one specific defence, institutional tacit knowledge and the relationships it travels with, is structurally stronger than most of the others against displacement, and the most under-appreciated. The rest of this paper uses Nvidia as the cleanest illustration to work the thesis out, then extends it across the AI economy. Nvidia defends a 75% gross margin three years into the AI build-out, with its lead widening rather than narrowing. Something is resisting the supposed commoditisation. The question is what exactly, and what the answer says about everyone else.

Nvidia as a case study: CUDA is institutional knowledge

Start with the cleanest illustration. Why can’t hyperscalers simply build their own GPU stack, given that they have the engineering budgets and the strategic motivation? Huang, Nvidia’s founder and CEO, uses a metaphor. A CPU, he says, is a Cadillac. Almost anyone can drive it at highway speed. An Nvidia accelerator is an F1 car: almost anyone can also drive it at a hundred miles an hour, but pushing it to its actual limit is a different sport.

The metaphor answers the question that has shadowed Nvidia’s margins since the first ASIC6 announcements: if hyperscalers can write their own kernels7, isn’t CUDA’s moat just inertia?

Writing kernels that run is not the same as writing kernels that saturate the hardware. Nvidia engineers, when they get inside a partner’s stack, have publicly delivered 2x or more on a meaningful share of kernels and up to 4x throughput in joint optimisation efforts8. Hyperscalers can technically build their own software. They cannot easily reproduce the engineers who have spent twenty years learning the racing line.

This kind of advantage has a name in the strategy literature. Hamilton Helmer, in 7 Powers (2016), calls it process power: an embedded set of company-specific activities that produce superior outputs as a result of long-term commitment to a process that competitors cannot shortcut even if they understand it perfectly. Helmer’s paradigmatic example is the Toyota Production System. CUDA is the software-stack equivalent. The code is open in many parts. The papers describing the techniques are public. What is not transferable is the layered intuition of thousands of engineers who have, over twenty cycles of GPU generations, learned which optimisations matter for which workload, which pitfalls to avoid, which non-obvious refactors unlock disproportionate gains.

The deeper concept here comes from the philosopher of science Michael Polanyi, who in The Tacit Dimension (1966, ★★★★★) coined the formula that has anchored a half-century of thinking on knowledge transfer: we know more than we can tell. The skilled worker knows things in their hands and their attention that they cannot articulate. The skilled engineer knows patterns of code-architecture interaction that they cannot fully document. This tacit dimension is unusually resistant to commoditisation because commoditisation works best on activities that can be specified, measured and reproduced. Polanyi’s point still holds: skilled practice contains regularities that the practitioner cannot fully articulate.

AI makes that boundary more porous. A model can infer patterns from behavioural traces that no engineer ever turned into a manual. Given tools and access, an agent can test hypotheses and gather more information. It would be a mistake to define the moat as “whatever AI cannot formalise”, because that frontier is moving quickly.

The stronger version of the argument is about the cost and continuity of capture.

Institutional knowledge is produced inside a stream of failed experiments, edge cases, supplier interactions, customer complaints and engineering decisions. The organisation observes those events as they happen, often before it knows which ones will later matter. A competitor or general model typically sees a much thinner record.

Three layers of reality: where institutional knowledge sits in the AI value stack, ordered by how costly each layer is to capture

Andler’s problem/situation distinction is still useful here. Formalising a situation is itself work. Better models can help with that work, sometimes dramatically, but they still need access to the situation and feedback on whether the formalisation was good enough.

Russell’s distinction between knowledge by acquaintance and knowledge by description can also be stated less absolutely. A model has access to descriptions unless it is connected to sensors, tools or people that let it investigate. Domain insiders already possess those channels. They know which measurement is suspicious, which supplier claim to question, which patient history changes the interpretation, which test to run next.

Coming back to our Reset Test9, a large capability jump in foundation models does not automatically erase Nvidia’s position because Nvidia’s advantage is not a static body of facts waiting to be learned. It is a continuously updated loop across hardware, software, workloads, suppliers and customers.

The moat is not knowledge that AI can never acquire. It is privileged access to the process by which the useful knowledge is generated, validated and refreshed.

Why better general models do not automatically capture the whole stack

The same logic applied above Nvidia gives a more useful version of the vertical-AI debate.

The strongest argument against vertical companies is obvious: general models improve quickly enough that a differentiated feature can become an API call a few quarters later. That risk is real. The opposite inference, that better general models therefore capture all of the downstream value, is much less obvious.

A model still has to be connected to a specific use context. In the examples below, that means proprietary data, instruments, permissions, regulation, feedback and people who know what to do when the formal description of the problem turns out to be incomplete.

Frontier labs can build those layers too, and increasingly do. There is no structural rule preventing a lab from moving deeper into health, defence or enterprise workflows. The question is how many situated loops a horizontal organisation can economically own at once, and whether it has the same access to the data and feedback as actors already inside them.

  • Human biology. A frontier model that reasons brilliantly across published clinical literature still depends, for any genuinely new prediction, on proprietary experimental data the major labs do not automatically have access to, but also on the institutional know-how to produce and interpret that data: which controls matter, which artefacts to discard, how to design the next experiment so it actually answers the question. This is why techbio companies cannot operate with AI talent alone: they need biologists who carry the tacit understanding that no published paper fully captures. As we showed in our piece on foundation models hitting the noise floor in biology, even the best perturbation-prediction models converge toward the irreducible experimental noise of the underlying biology. The moat is the loop: companies that own both the experimental infrastructure and the modelling layer can keep pushing that ceiling down. This is probably why Anthropic has begun to acquire in health. Companies like OneBiosciences10 (clinical-grade single-cell tumour profiling) and Raidium11 (the first radiology foundation model, trained on 130TB of real-world imaging) build precisely on that combination of data and human expertise.
  • Plant biology. The same logic applies in agritech, where the value sits not only in proprietary germplasm (the genetic material of cultivated plant lines, accumulated by seed companies and research institutions over decades) and matching multi-environment phenotyping data, but also in the long-running contractual relationships with breeders and research stations that produce that data, and in the institutional know-how to translate genomic signals into commercially relevant traits. Breeding cycles still run on multi-year horizons (8 to 12 years for major commodity crops, faster for vegetables like tomatoes), which makes the accumulated institutional position structurally hard to bypass. Living Models12 has built BOTANIC, a foundation model trained on 1,600 plant genomes across 43 species, designed to make this knowledge accessible to regional breeders and underserved crop systems (sorghum, cassava, millet) where climate adaptation matters most.
  • Defence and dual-use. The moat here lives in highly sensitive datasets that no public actor can access (sensor recordings, mission telemetry, sovereign-controlled training data), in the institutional relationships with procurement agencies, integrators, and operational forces that take years to establish, and in the operational know-how needed to build systems that work under extreme robustness, multi-modal sensor fusion, edge inference, and adversarial conditions. None of this is obtained simply by training a stronger model on internet-scale text and code. Harmattan AI13 is building in this space, where the contextual barriers are stark and the moat lives in mastering them.

These three are vertical plays sharing the same structure: deep contextual data and know-how that the labs cannot reach. As a side note before moving on, the vertical lens is not the only one that works. A startup can also compete with frontier labs across many verticals at once, provided it brings two things the labs do not: a specialised non-LLM architecture and proprietary data the LLM training corpus does not contain. PriorLabs14 has built TabPFN, a tabular foundation model that learns causal relationships from synthetic data and matches tuned tree-based ensembles at industry scale15. The same logic could apply to advanced manufacturing intelligence (AMI) or to world models for robotics and physical simulation. In each case, the moat lives in specialised technical know-how rather than in the customer’s situation, but the underlying logic is the same: the value can sit outside the general model itself, in specialised architectures, proprietary data and the loop that connects them to the problem.

There is a second constraint on frontier labs, although I would describe it as economic rather than structural. Their largest fixed costs and strongest distribution advantages push them toward horizontal model capability and very broad deployment. A vertical actor can instead concentrate its organisation on one workflow, and can route across model providers when that is useful.

That does not mean the labs are condemned to sell commodity tokens. They have distribution, consumer interfaces, proprietary capabilities, developer ecosystems and the option to integrate further into particular workflows. OpenAI, Anthropic and Google are already doing exactly that.

Nor does it mean a vertical company is safe merely because it knows an industry. As the general models improve, parts of today’s vertical product will repeatedly collapse into the model layer.

The market is more likely to move at both ends. General models absorb tasks whose context becomes standardised and cheap to obtain. Specialised models, vertical software and workflow companies retain or create value where performance still depends on proprietary data, unusual architectures, access, feedback or deep integration into the real world.

The strategic question is not whether the model or the vertical “wins”. It is which parts of the surrounding system become generic as the model improves, and which parts become more valuable.

Why mobility beats coercion

If the moat is institutional knowledge, the next question is where it circulates best. Counter-intuitively, the strongest ecosystems are not those that protect knowledge through legal coercion, but those that let it circulate. AnnaLee Saxenian’s Regional Advantage (1994) showed why Silicon Valley pulled ahead of Route 128 in the 1980s: California’s non-enforcement of non-competes (in place since 1872) meant engineers could leave one employer and join another the next morning, while Massachusetts protected individual firms at the cost of ecosystem dynamism. Each Boston company was more defensible. The Boston ecosystem compounded knowledge slower.

Talent mobility functions as an immune system that maintains the health of an entire ecosystem. Tacit knowledge migrates with the humans who carry it; each person changing jobs every three years (engineers, but also operators, regulatory specialists, commercial leaders, clinicians) carries years of accumulated practice into a new combination, where it compounds with what was already there. The same dynamic made medieval Venice the wealthiest city in Europe through its commenda contracts, and later destroyed it when the serrata of 1297 closed access to a hereditary elite (we worked this case study out in our digest of Why Nations Fail). Strict non-compete enforcement is the modern serrata.16 The collective moat is more defensible than the individual moats that erode through circulation.

The ultimate edge: combining institutional knowledge with continuous renewal

If institutional knowledge and relationships accumulate in people, then old companies should win mechanically because they have more memory. But that is emphatically not what we observe. Even the most technically sophisticated incumbents fail to innovate from within and have to acquire startups to refresh their position17. They are not worse than non-tech incumbents on this point; they are essentially in the same situation, despite the structural advantage one would expect from their software leverage.

Microsoft has effectively bought OpenAI. Google had DeepMind for over a decade and still had to rebuild much of its AI position with Gemini. IBM missed AI cloud entirely. Cisco has been buying its way into every adjacent technology for two decades. Why do most large tech companies have to outsource their innovation to startups they later acquire?

A first observation: the displacing startups themselves are rarely from scratch. Anthropic was founded by senior OpenAI alumni. Mistral grew out of Meta’s FAIR lab. Strong vertical AI companies are typically founded by former operators of the verticals they target, and they aggressively recruit from incumbents to capture operational knowledge, customer relationships, and regulatory savvy. Innovation does not arrive from outside the institutional knowledge base; it arrives from outside the institutional structure that was holding it captive.

David Teece’s Dynamic Capabilities and Strategic Management (1997) gives the frame: operational capabilities (doing well what you already know how to do) versus dynamic capabilities (reconfiguring competences in response to a changing environment). Most large companies excel at the first and atrophy on the second. The rare companies that combine both, Apple, Nvidia, and to some extent TSMC, share two structural traits:

  1. They operate under a deep-tech metabolism that forces organisational memory to compound. As we developed in our deployment series, the deep tech paradox (long feedback loops, slow iteration, deep interconnections, low reversibility) is not just a constraint, it is a strategic metabolism that rewards firms able to compress learning within physical limits. Each tape-out at Nvidia costs $50M and 3 years; TSMC’s fabs do not forgive approximation; Apple ships products customers hold in their hands. The cost of forgetting becomes prohibitive, and organisational memory compounds in ways that fast-iterating software shops structurally lose, against the dominant velocity-over-everything doxa in tech. The pattern weakens for the more software-anchored frontier firms (Meta primarily, Google and Microsoft sitting in between because of their hyperscale infrastructure), which is consistent with their heavier reliance on acquisition.
  2. They have a governance structure that lets them go beyond the innovator’s dilemma. Christensen’s The Innovator’s Dilemma (1997) showed that incumbents are structurally incapable of investing in innovations that cannibalise their current revenues, because every internal incentive pushes against it. The flip side, for new entrants, is sustaining a daring direction that contradicts dominant beliefs and available data, the “this is undoable, yet done” posture. Apple, Nvidia and TSMC have managed both forms over time. The presence of a founder still in operational command (Huang at Nvidia, Tim Cook carrying the Jobs lineage at Apple, Morris Chang at TSMC into his eighties) is one mechanism, but not the only one: long-term concentrated ownership, dual-class structures, mission-clarity at the board level can play a similar role.

Maintaining dynamic capabilities seems to require deep-tech metabolism plus a governance structure that authorises sustained reallocation against current consensus.

Conceptual toolbox

ConceptWhat it meansWhy it matters
Institutional knowledge and relationshipsAccumulated practice living in people, processes, data and trusted relationships with customers, suppliers and regulators. Some can be codified; much is costly to identify and continuously refresh.Advantage persists when an organisation has privileged access to the loop that generates and validates the knowledge.
Process power (Helmer)An embedded set of company-specific activities that produce superior outputs through long-term commitment to a process competitors cannot shortcut.Names the mechanism behind CUDA, the Toyota Production System, and moats that compound through repeated practice rather than through scale or IP alone.
Polanyi’s tacit dimension (★★★★★)We know more than we can tell. Embodied skill and situational judgement exceed what practitioners can fully state as rules.AI can learn some of these regularities from behavioural traces, but gathering the right traces and testing them remains an organisational problem.
Andler’s problem vs situationA problem comes with a representation and objective; a situation still requires deciding what matters and how to represent it.Better models can help formalise situations. Situated actors retain an advantage when they control access, feedback and the test of whether the formalisation works.
Russell’s acquaintance vs descriptionKnowledge through direct contact with a thing versus knowledge through propositions about it.Models can investigate when connected to tools and sensors, but domain insiders often begin with richer access and a better sense of where the description may be incomplete.

What this means for Nvidia

Nvidia combines CUDA and its accumulated engineering practice with a deep-tech metabolism that forces that knowledge to be refreshed every product cycle, and a governance structure centred on Huang that has repeatedly allowed the company to reallocate against its current product.

Those advantages reinforce one another. A new architecture creates production and software problems. Solving them generates new knowledge with customers and suppliers. Deployment exposes new workload bottlenecks. The next architecture begins with more information than the previous one did.

The strategy is not free. By July 2026, Nvidia had committed $279 billion to future supply and capacity, primarily memory and manufacturing capacity, alongside additional commitments to cloud services, data-centre leases and equity investments18. Some of those commitments are flexible and they should not be described as off-balance-sheet capex. What they show is that Nvidia’s “asset-light” model increasingly depends on coordinating large amounts of industrial capacity that sit on other balance sheets.

Whether that remains a source of leverage or becomes a source of rigidity is exactly the kind of multi-dimensional problem the next paper is about.


“We know more than we can tell.”

Michael Polanyi, The Tacit Dimension, 1966, ★★★★★.


Willy Braun signature

Footnotes

  1. [Context] Alphabet has built a substantial part of its current product portfolio through acquisition: Maps (Where 2 Technologies, Keyhole, ZipDash, 2004), Android (2005), YouTube (2006), DoubleClick (2007), Waze (2013), Looker (2019), Fitbit (2021), and many infrastructure layers underneath Gmail, Google Docs (originally Writely, 2006) and Google Spreadsheets (originally 2Web Technologies’ XL2Web, 2006). ↩

  2. [Context] Amazon’s portfolio includes Audible (2008), Zappos (2009), Kiva Systems (2012, now Amazon Robotics), Twitch (2014), Annapurna Labs (2015, the silicon team behind AWS Graviton and Trainium), Whole Foods (2017), Ring (2018), MGM Studios (2022), and iRobot (announced 2022, abandoned 2024). ↩

  3. [Context] Apple has acquired the foundations of much of what users perceive as native: Siri (2010), Beats (2014), Shazam (2018), Workflow (2017, which became Shortcuts), and most of the silicon design lineage that became the M-series chips (P.A. Semi 2008, Intrinsity 2010). ↩

  4. [Context] Meta’s most strategically important products are acquired: Instagram (2012), WhatsApp (2014), Oculus (2014, the foundation of its VR strategy), and CTRL-Labs (2019). ↩

  5. [Context] Microsoft has been one of the most acquisitive frontier tech companies: Skype (2011), Mojang (2014), LinkedIn (2016), GitHub (2018), Nuance (2022), Activision-Blizzard (2023), and the deep equity-and-cloud arrangement with OpenAI (2019 onwards) that has functionally made OpenAI a Microsoft business unit. ↩

  6. [Context] ASIC = Application-Specific Integrated Circuit. A chip designed for one workload (rather than a general-purpose GPU which can run many). In AI, the most prominent ASIC programmes are Google’s TPU, Amazon’s Trainium, Meta’s MTIA and Microsoft’s Maia. ASICs typically achieve better performance per watt on their target workload, in exchange for far less flexibility. ↩

  7. [Context] A kernel is a small piece of code that runs on the GPU and performs one specific computation, such as a matrix multiplication or an attention operation. Models are built by chaining thousands of these kernels. The same model running on the same hardware can be two to ten times faster or slower depending on how the kernels are written. ↩

  8. [Source] Cursor / Nvidia collaboration on Blackwell kernel optimisation (October 2025) reported a 38% geometric mean speedup across 235 production kernels, with 2x or more achieved on 19% of them. A separate vLLM / Nvidia collaboration delivered up to 4x throughput improvement at similar latency between Hopper and Blackwell. ↩

  9. [Reminder] The Reset Test is one of the three sanity checks proposed in our AI moats framework. It asks: if a frontier model launches tomorrow with 10x better capabilities, what protects you? Valid protections include proprietary data unavailable to new models, deep workflow lock-in, regulatory barriers, trust capital, and network effects. Invalid protections include “we are ahead today” or “we know the industry better”. ↩

  10. [Disclosure] I am an investor in OneBiosciences via Galion.exe. ↩

  11. [Disclosure] I am an investor in Raidium via Galion.exe. ↩

  12. [Disclosure] I am an investor in Living Models via Galion.exe. See also the AgTechNavigator coverage of Living Models’ BOTANIC foundation model. ↩

  13. [Disclosure] I am an investor in Harmattan AI via Galion.exe. ↩

  14. [Disclosure] I am an investor in PriorLabs via Galion.exe. ↩

  15. [Source] See PriorLabs’ technical reports, including the TabPFN-2.5 Model Report (November 2025), which substantially outperforms tuned tree-based models and matches the accuracy of AutoGluon 1.4 (a complex four-hour tuned ensemble) on industry-standard benchmarks with up to 50,000 data points and 2,000 features. The December 2025 Scaling Mode report extends the model to millions of rows. ↩

  16. [Caveat] One could object that contemporary China, despite its political controls, manages effective sectoral mobility through intense internal competition between state-affiliated actors. That is probably true. The difference with the USSR, which had heavy industrial capex without that internal competition, suggests that the critical condition is not the political system but the effective circulation of knowledge and competitive selection. This is a thread we explore further in a later entry of this series, dedicated specifically to the China question. ↩

  17. [Note] One might expect tech companies to have the structural advantage here, since software is precisely the leverage that translates institutional knowledge into compounding capability (codified workflows, reusable infrastructure, automated decisions). The fact that even tech-native incumbents systematically fail to innovate from within suggests that institutional knowledge alone, however technically sophisticated the carrier, is not enough. What seems to matter is the combination of institutional knowledge with a governance configuration that authorises sustained reallocation against current consensus, which we explore further below. ↩

  18. [Source] NVIDIA, Form 10-Q for the quarter ended July 26, 2026. Nvidia disclosed $279 billion of supply and capacity commitments, $29 billion of cloud service agreements, $25 billion of data-centre leases not yet commenced, $25 billion of equity-investment commitments and $8 billion of capital-expenditure commitments. These categories have different accounting treatment and cancellation rights. SEC filing. ↩

← Return to the series overview: Reading Nvidia: On strategy, defensibility, and the return of industrial AI