The Infrastructure Fallacy.

The most seductive investments in a platform shift are the picks-and-shovels plays. During the California Gold Rush, the merchants selling shovels and denim outperformed most miners. The analogy has been applied to every technology transition since. But it obscures a crucial distinction: the merchants who sold to miners captured a fraction of the value the transition created. The fortunes from the Gold Rush era were made in banking, railroads, and real estate. Not in the shovel business.

The current AI transition has produced a similar misreading. Approximately $700 billion in planned capital expenditure will be deployed by the four largest technology companies in 2026 alone into GPU clusters, data centers, and power infrastructure. For context, this exceeds global telecommunications capital expenditure by more than twofold. ^[1] The prevailing logic is straightforward: AI is transformative, models require massive compute, therefore compute providers win. This logic is not wrong about the technology. It is wrong about the economics.

The telecommunications industry built the infrastructure for the mobile internet. Between 2005 and 2015, global telcos spent approximately $2 trillion on network infrastructure that made smartphones possible. During that same period, global telco stock indices were essentially flat while the companies that built applications on that infrastructure (Apple, Google, Meta, Uber, Airbnb) created more than $5 trillion in market value. The people who funded the pipes got pipe economics. The people who built the experiences got software economics.

The question this paper addresses is not whether AI is transformative. It manifestly is. The real question is where in the value chain the returns will concentrate. The answer has been consistent across every platform shift in computing history: in the application layer, built on commoditizing infrastructure, solving specific problems for specific buyers who have exhausted the alternatives.

“Commodity infrastructure rarely captures value up the stack. Mobile networks are a trillion-dollar industry, but all the use-cases and value-capture are built by other people.” Benedict Evans, 2026

A platform shift occurs roughly every ten to fifteen years. Each one resets the technology industry, displaces previously dominant players, and creates a window in which new companies can be built without the incumbent advantages that ordinarily make competition prohibitively expensive. ^[1] We have seen this pattern clearly four times in the modern era: mainframes to personal computers, personal computers to the web, the web to smartphones, and now smartphones to generative AI.

The mechanism is consistent. In each transition, a new underlying capability becomes available that makes previously impossible things possible and previously expensive things cheap. Early capital flows into building the infrastructure required to deploy the capability at scale. The infrastructure providers appear, initially, to be the winners. They are growing fast, capturing headlines, commanding premium valuations. Then commoditization arrives. Competition, efficiency gains, and the commodifying logic of markets compress infrastructure margins. The companies that win the long arc are those that built durable businesses using the infrastructure, not those that built the infrastructure itself.

The pattern has one nuance that matters enormously for timing: each platform shift creates a window, and the window closes. The web created a period (roughly 1994 to 2000) in which almost any company that achieved real distribution could build a defensible position. After 2000, the window narrowed sharply. New entrants faced incumbents with established brands, network effects, and distribution advantages that made cold-start competition extremely expensive. The companies that built during the window (Amazon, Google, Salesforce) became the entrenched powers of the next decade. The companies that tried to build after the window closed became cautionary tales.

The AI window is open now. It will not remain open indefinitely.

Benedict Evans’ provisional thesis in “AI Eats the World” states it clearly: models appear to be commodities. They are capital-intensive to build, show no meaningful network effects, and as of mid-2026, leading frontier models are converging in performance on general benchmarks. The competitive differentiation that exists today is being competed away at a rate that should concern anyone betting on model-layer economics.

The inference efficiency curve alone tells most of the story. Compute efficiency for equivalent model capability is growing at 50 to 100 times per year, a rate that dwarfs Moore’s Law. The cost of running today’s frontier capability is declining sharply on a six-to-twelve-month horizon. Sam Altman’s articulation of intelligence as “like electricity or water” is accurate as a directional prediction, even if the timeline is uncertain. Utilities are necessary. They are not where fortunes are made.

The telco analogy is instructive here, not just as rhetoric but as a precise economic parallel. Bits on mobile networks are, from the consumer’s perspective, interchangeable. Telcos competed on coverage and price. Margins compressed. Capex-to-revenue ratios became punishing. The equity returns over the 2010–2025 period were, for telecom as a sector, unremarkable. Meanwhile, the companies that treated bits as free infrastructure and competed on the quality of the experience built on top of those bits captured extraordinary value.

Tokens are the new bits. The question is not who generates the tokens most efficiently. The question is who builds the experience, the workflow, and the outcome that tokens enable. The answer to that question is not the labs. The labs, by their own admission, cannot build all the applications. ^[1] They are building infrastructure. Someone else will build the applications. The entity that builds the right applications, with the right structural advantages, will capture the majority of the value this transition creates.

“Chat is a terrible UX. General use needs apps. Labs can’t build all the apps. Models will just be infra. Innovation will move up the stack.” Benedict Evans, provisional thesis, 2026

The application layer is where the economic value of this transition will concentrate. These are the products, workflows, and outcomes built on top of model infrastructure. This is not a controversial prediction. It follows directly from the pattern of every previous platform shift. What is less well understood is the specific mechanism through which application companies achieve durable defensibility in an environment where the underlying capability is rapidly commoditizing.

Three mechanisms produce defensibility at the application layer:

These mechanisms are not new observations. What is new in the AI transition is the speed at which they can be built and the scale at which they can operate. A team of ten, building on frontier model infrastructure, can now build a product that previously required a team of two hundred. Mark Zuckerberg has observed that “we’re seeing more and more examples where one or two people are building something in a week that would have previously taken dozens of people months.” ^[1]This compression of the relationship between team size and product capability is the structural advantage available to application builders in this transition that was not available in previous ones.

The implication is significant: the application layer in the AI era is simultaneously more accessible (lower barriers to entry) and more winner-take-most (the network and data effects that build in the first-mover compound rapidly). The companies that move quickly and with genuine insight into buyer problems will establish positions that become progressively harder to dislodge.

The dominant narrative about AI’s impact on industries is one of efficiency: AI will make existing processes faster and cheaper. This narrative is correct and largely irrelevant to the question of where the exceptional returns will be found. Efficiency improvements accrue to the buyer of the tool, not to the seller, unless the tool becomes the dominant operating system for an industry. And the industries where AI’s impact will be most profound are precisely those where the current operating model is not a matter of efficiency but of structure.

Health insurance, financial advice, legal services, corporate education, government technology procurement, clinical research: these industries are structured as they are not because their current form is efficient, but because they evolved in an environment of artificial scarcity: information scarcity, credentialing scarcity, distribution scarcity. The incumbent industry structure reflects the cost of information processing before AI made information processing effectively free.

Consider health insurance. The core economic function of an insurance pool is risk aggregation: spreading the unpredictable costs of individual health events across a population. This is a valuable function. The administrative superstructure around it (prior authorizations, claims adjudication, network management, formulary gatekeeping) is not a refinement of risk aggregation. It is the friction that accumulated when information about individual health status was expensive to collect, aggregate, and reason about. AI does not make claims processing faster. It eliminates the reason claims processing is the bottleneck. The industry that emerges on the other side of this transition will look structurally different from the industry that exists today.

The underlying capability transition makes this structural. We are building systems that can reason, and within this decade, those systems will be able to do the work of a qualified professional in a growing range of domains. The implication is not that professionals will be automated away. It is that the information asymmetry that made professionals indispensable intermediaries will erode. When everyone has access to a system that can reason about their health situation, legal options, or financial position with the competence of a skilled professional, the industries structured to manage that asymmetry will be forced to restructure.

This is Benedict Evans’ question framed precisely: “Which industries were protected by a cost base that AI can now automate to zero?” ^[1]The answer to that question identifies the industries where the replacement opportunity, not the efficiency opportunity, is real. The SerebrumAI thesis is grounded in a specific claim: the most attractive investments in the AI transition are not companies that make existing industries more efficient, but companies that eliminate the structural conditions that made those industries possible. We do not invest in industries. We rebuild them.

“Was the cost of that task your moat? The internet removed physical distribution costs that protected many industries from competition. Which industries now are protected by a cost base that AI can automate to zero?” Benedict Evans, 2026

The distinction between efficiency and replacement is not semantic. It determines the return profile. An efficiency improvement in an existing industry is captured by the buyer and competed away among sellers of efficiency tools. A replacement of the structural conditions of an industry creates a a genuinely new market, one that did not exist before, in which the company that defines the new model captures a disproportionate share. The difference in financial outcome between “we made the existing process 40% faster” and “we built the replacement for the existing process” is measured in orders of magnitude, not percentage points.

If the thesis is correct that application-layer companies will capture the majority of the value from the AI transition, a second question immediately follows: which applications? The YC batch composition data is instructive here: by 2025, more than half of all Y Combinator startup applications described themselves as AI companies. ^[1]A platform shift that generates a flood of new company formation does not, by itself, produce uniformly good returns. Most new companies formed during a platform shift fail. The returns are heavily concentrated in the small fraction that navigate selection correctly.

The venture return distribution is not normal. It is power-law distributed: a small number of investments generate the overwhelming majority of returns, and the median investment returns less than capital. This has been true across every technology cycle. In the AI transition, we believe the power-law concentration is more pronounced than in previous cycles, for a specific reason: the window is shorter. The inference efficiency curve, the rapid diffusion of open-source models, and the labs’ own aggressive expansion into adjacent application categories are all compressing the window in which a given category can be won by a specific entrant. AI-native tools commoditize faster than any prior platform technology.

In the mobile era, a company that built a compelling travel booking app in 2010 had several years before the category consolidated. In the AI era, the equivalent window may be twelve to eighteen months in consumer categories and somewhat longer in enterprise, where sales cycles and integration complexity provide natural buffering. This compression is not an argument against building. It is an argument for building with significantly more discipline about which category to enter and with what structural advantage.

Three variables determine whether an application-layer investment belongs in a portfolio at this moment in the transition:

Conviction alone is not a selection framework. Every founder has high conviction about their own idea. What discipline requires is a framework that separates the variables that matter, independent of how attractive the idea feels, and forces an honest assessment of each.

The SerebrumAI priority matrix uses two primary axes and one explicit tiebreaker:

The matrix is designed to reveal, not to decide. Its function is to surface the cases where an idea that feels attractive is positioned in a cell that, on honest assessment, does not belong in a portfolio. A venture with high conviction but a low outcome ceiling (<$50M) is a lifestyle business, not a portfolio company. A venture with a high outcome ceiling but a 36-month time to first revenue is a long-arc bet that requires the portfolio to have short-arc cashflow funding it. The matrix makes these relationships explicit.

What the matrix deliberately excludes is as important as what it includes. Capital intensity, regulatory complexity, and strategic fit with the AI-native thesis are not plotted on the matrix. Capital intensity informs the raise sequence and the timeline, but does not by itself disqualify an investment. Some of the highest-returning companies in venture history required substantial capital before achieving cash-flow positivity. Regulatory complexity is real in health and finance, but regulatory barriers that protect a market from competition are also structural moats; the question is whether the company is positioned inside or outside the barrier. And strategic fit with the AI-native thesis is a prerequisite for appearing on the matrix at all: every company in the portfolio must be able to articulate why it could not have been built before the current AI capabilities existed.

The selection framework describes what to build. The wedge structure describes how to build it in a way that maximizes the probability of survival through the period of market validation. A wedge is a pre-existing advantage that reduces the cost of acquiring the first buyers. In the context of the SerebrumAI portfolio, every speculative idea inherits the distribution of an existing venture.

This is not an accident. It is a deliberate structural choice grounded in a specific observation: in the AI transition, the cost of cold-start customer acquisition is rising, not falling. The explanation is counterintuitive but straightforward. As AI capabilities become widely available, more companies are attempting to sell AI-native solutions to the same buyers. The buyer’s attention and evaluation capacity is finite. The noise-to-signal ratio in their vendor conversations is increasing. The cost of breaking through that noise to a genuine evaluation is rising. A genuine evaluation is the point at which the buyer is seriously considering the product on its merits, not just tolerating a demo. A warm introduction from an existing trusted relationship compresses that cost dramatically.

The structural advantage is not capital. It is multiplicative distribution wedges: the accumulated buyer relationships, domain credibility, and pipeline access that each existing venture creates for adjacent ideas. An AI-native fund management platform (AlphaSigma) creates natural entry points into institutional finance for ventures that serve the same buyer population. An AI-native health and wellness platform (Vygor) creates natural entry points into consumer health relationships for adjacent nutrition, fitness, and wellness applications. An AI-native software lifecycle tool (TokenSource) creates entry points into engineering leadership and product management relationships for adjacent developer productivity tools.

The compounding logic here is not linear. Each venture that achieves market credibility compounds the distribution advantage of the ventures that follow it. At maturity, the portfolio is not simply a collection of independent companies. It is a network of mutually reinforcing market positions, each leveraging the relationships built by the others.

Building multiple ventures simultaneously creates a specific capital challenge that a single-company startup does not face: it must fund multiple bets simultaneously, some of which will take considerably longer than others to achieve cash-flow positivity. The standard response is to raise a large fund and deploy it across the portfolio. This works, but it creates a dependency on external capital that reduces optionality and increases dilution in the ventures that perform well.

The alternative is to sequence the portfolio’s revenue generation so that shorter-arc cashflow-positive ventures fund the operating costs while longer-arc ventures are built. This is the “cashflow choreography” approach: the portfolio is designed with explicit attention to the temporal relationship between revenue streams.

The critical design constraint is sequencing. Concentrating all capital in the long-arc bet before short-arc cashflow has proven out is taking on more risk than is necessary. Starving the long-arc bet in order to over-optimize the short-arc sacrifices the returns that make the portfolio worth building. The cashflow choreography framework holds both constraints simultaneously: prove the thesis at small scale first, then use that proof to unlock the capital and credibility required for the large-scale bets.

A thesis paper that does not engage seriously with the risks against its own argument is not a thesis paper. It is marketing. The following are the risks that the App Layer Thesis must navigate, and our current assessment of each:

The term “AI-native” has become so widely used that it risks becoming meaningless. Every company building on a language model API calls itself AI-native. A more precise definition is useful: an AI-native company is one whose core value proposition could not exist without AI capabilities, and whose operating model is designed around AI from its foundation, not retrofitted onto a prior architecture.

The distinction matters because the difference between an AI-augmented company and an AI-native company is not primarily a difference in technology. It is a difference in operating model. An AI-augmented company uses AI to make its existing processes faster or cheaper. An AI-native company has redesigned its processes from first principles around the assumption that AI capabilities are available, and has therefore structured its team, its data architecture, its buyer relationships, and its unit economics in ways that would not be possible without AI.

Mark Zuckerberg’s framing in “The Future is for Everyone” is relevant here, though in a different context: “Invention, not automation, will be the greatest contribution of superintelligence.” ^[3]The same distinction applies to company building. The application-layer companies that capture the most value will not be those that automated existing processes. They will be those that invented new operating models that only became possible when AI capabilities became available.

A health company that uses AI to automate prior authorizations faster is automating. A health company that redesigns the relationship between patient, provider, and payer around the assumption that continuous, longitudinal health intelligence is available for every individual is inventing. The financial return profile of these two approaches is not comparable.

One structural advantage of building multiple AI-native companies together is the ability to develop shared AI infrastructure that compounds across the portfolio. Individual companies building AI products in isolation each bear the full cost of developing the intelligence layer their product requires. Building multiple ventures in adjacent domains can amortize that cost across the portfolio, with the intelligence layer improving as it accumulates data from multiple applications.

In practice, this means that the data collected by a health outcomes application can inform the risk models used by a health finance application. The buyer intelligence accumulated through an enterprise software sales process can inform the go-to-market approach of adjacent enterprise applications. The AI research developed for one application becomes, with appropriate adaptation, infrastructure for the next.

This compounding of intelligence across the portfolio is not achievable by a standalone company and is not easily replicable by a traditional VC fund whose portfolio companies operate independently. It is one of the core structural advantages of the Serebrum approach in the AI transition: the portfolio compounds what it learns.

A thesis is not a portfolio. Getting the macro call right on a platform shift is necessary but not sufficient. The firms that generated the best returns from the PC, internet, and mobile transitions were not always the ones with the most accurate technology forecasts. They were the ones with the most disciplined capital allocation within a thesis. Having a strong view on which way the world is moving is table stakes. What separates returns is what you do with that view.

The first discipline is position sizing. Conviction should drive concentration. Equal-weighted portfolios encode a belief that every bet is equally likely to be right, which is a form of intellectual dishonesty. If the framework in Section VII genuinely distinguishes between a weight-4 build-now flagship and a weight-0 skip, the capital allocation should reflect that distinction. In practice this means the flagship venture draws meaningfully more time, operational support, and follow-on capital than the cash-flow plays at the mid tier. The matrix is not decorative; it is a capital allocation instruction.

Having a strong view on which way the world is moving is table stakes. What separates returns is what you do with that view.

The second discipline is the pre-mortem. Before committing to a venture, work forward to the most plausible failure scenario. Not the catastrophic ones, which are obvious, but the mundane ones: the wedge never converts to the core product, the regulatory pathway takes 36 months instead of 12, the model provider changes pricing and kills the unit economics. The pre-mortem is not pessimism. It is a forcing function for identifying which assumptions are load-bearing and what evidence would confirm or refute them within the first 90 days of operation. If no evidence is available in that window, the venture is operating on pure narrative, and narrative without evidence is not an investment thesis; it is a bet.

The third discipline is being non-consensus and correct. This phrase, from Howard Marks, captures something that most investment frameworks miss: a view that is widely shared cannot generate differentiated returns, even if it is right. The crowded trades in AI investing in 2025 and 2026 are infrastructure, model access, and horizontal AI tooling. All three are probably directionally correct. None of them offer the return potential they would have in 2022 because the consensus has already priced them. The application layer in specific verticals is less crowded precisely because it is harder to underwrite: the market structure is less clear, the moats are less obvious, and the path to $1M ARR is less predictable. That uncertainty is the source of the return premium.

The fourth discipline, specific to any multi-venture portfolio, is sequencing. The priority matrix defines a capital flow: the cash-flow plays at the $50M to $500M tier with sub-12-month time to $1M ARR are not just portfolio companies, they are the financial architecture that funds the long-arc bets at the top of the matrix. This is a coherent strategy only if the sequencing is explicit and maintained. Treating all ventures as independent entities misses the structural advantage of the portfolio: the ability to choreograph capital flows so that the portfolio as a whole can hold larger, longer positions than any individual venture could support.

Investment discipline is not a constraint on ambition. It is the mechanism by which ambition survives contact with reality long enough to compound.

AI doesn’t produce one kind of company. Every platform shift generates multiple structurally distinct modes of value creation: some ventures create categories that didn’t previously exist, some replace incumbents with architecturally superior models, some aggregate fragmented markets that coordination costs had made impossible to consolidate, and some rebuild an established category’s infrastructure from scratch. Identifying which mode applies to a given industry is not taxonomic housekeeping. It determines the competitive dynamics the venture will face, the capital profile it requires, the defensibility mechanism it needs to build, and the timeline on which returns are likely to arrive.

The four modes are not equally distributed across opportunities. Market structure, the nature of incumbent moats, and the specific ways AI changes the cost and capability profile of a given function determine which mode is available. Choosing the wrong mode is a structural error that no amount of execution can correct: building a Create play in a market where Re-platform is already available wastes the education premium; attempting to Consolidate a market where the fragmentation is regulatory rather than operational mistakes the source of the problem.

Some categories don’t exist until the enabling technology exists. The job of orchestrating a software development lifecycle as a continuous intelligence system — spanning planning, specification, coding, testing, and deployment — is not a job that predated AI. Coding assistants and automated review tools addressed fragments of the lifecycle. TokenSource addresses the lifecycle itself. The product accumulates context across the full development workflow: team velocity patterns, spec-to-code accuracy rates, recurring failure signatures, and the compounding history of what works for a given codebase and team. No point-solution competitor can replicate that accumulated context without operating at the lifecycle layer. The moat is not the product; it is the longitudinal intelligence generated by owning the layer.

Create plays carry patient capital and longer time to revenue. No prior market means no benchmark pricing and no established buyer journey. The ceiling compensates: a category creator that becomes the category name captures not just current market share but the definition of what the market is.

When an industry’s operating model is not just improvable but structurally obsolete, the opportunity is replacement. The distinction matters because incumbents optimizing their existing model cannot close the structural gap with a challenger that started from a blank sheet. Absolute return fund management has been practiced for decades. What changes with AI is not the function — identifying and capitalizing on market dislocations — but the architecture available to perform it. A traditional quant fund operates within the constraints of human review cycles, batch decisioning, and portfolio managers as single points of failure. AlphaSigma runs continuous AI-driven portfolio management with transparent, auditable decisioning that has no equivalent structural constraint. The function is identical; the operating model is not. Incumbents cannot adapt without rebuilding from scratch, which means admitting the existing model is wrong. Established funds almost never do this voluntarily. The moat compounds: every market regime AlphaSigma navigates adds to a track record and proprietary signal library that any new entrant would have to wait years to replicate.

Replace plays convert to revenue faster than Create plays because buyers understand the category — they are already spending on the function being replaced. The competitive obstacle is not market education but incumbent relationships and switching costs. The structural argument that the AI-native architecture is simply better, not just cheaper, is the wedge.

Health and wellness is a market fragmented by the absence of a unifying intelligence layer, not by competitive moats. A person managing their health today relies on a personal trainer for exercise programming, a nutritionist for dietary guidance, a separate app for macro tracking, another for meal planning, and some combination of habit tools to hold it together. Each point solution is optimized for its slice. None of them reason across the full picture. The result is an expensive, high-friction experience that demands significant self-coordination — and produces inconsistent outcomes because no single intelligence is working across nutrition, movement, and recovery simultaneously.

What AI changes is the cost of unification. Vygor builds the intelligence layer that connects nutrition, movement, and recovery into a continuous, adaptive system: meal planning and macro tracking that adjusts to the user’s actual training load that week; exercise programming informed by dietary state and recovery signals; longitudinal behavioral intelligence that identifies what is actually working for a given individual across months of data. This is not a bundled app. It is a reasoning system — one that gets meaningfully smarter the longer it runs on a specific user’s data, and that no point solution can replicate without access to the full behavioral picture.

The rollup strategy compounds the platform. Vygor consolidates health and wellness businesses — personal trainers, nutritionists, fitness studios, wellness coaches — into an AI-powered network where each provider is enhanced by the platform’s intelligence rather than operating in isolation. Each acquisition extends the data surface: more dietary adherence patterns, more exercise response data, more individual behavioral signals across a wider population. Three years of continuous behavioral data — macro-to-outcome correlations, exercise response curves, adherence patterns by lifestyle segment — is not reproducible by any general-purpose model regardless of its underlying capability. The longitudinal data asset is the moat.

Consolidate plays have more predictable capital profiles than Create or Replace. The revenue base exists in a growing subscriber base and contracted providers. Technology risk is lower than in Create plays. The execution challenge is integrating acquired businesses without degrading their quality, and maintaining personalization at scale as the platform grows. AI is the structural differentiator; operations are the work.

A fourth mode sits between Create and Replace: taking a validated category and rebuilding it from scratch on AI-native infrastructure, at a fraction of the cost and with substantially higher intelligence. The function, the buyer, and the budget exist. What doesn’t exist is a product built for the AI era rather than retrofitted to it. FedStat operates in this mode within federal technology portfolio management — budgeting, application rationalization, technology business management, capex planning, and goal tracking. The category is established; products like Apptio and ServiceNow ITBM have demonstrated sustained enterprise willingness to pay for this function. What they have not demonstrated is the ability to reason across the data they contain, respond to natural language queries over complex budget and application portfolios, or generate rationalization recommendations that would previously have required weeks of analyst time. FedStat rebuilds this function AI-natively. The competitive advantage is not category creation — buyers already understand what they need. It is that the AI-native architecture is structurally cheaper and meaningfully smarter than what incumbents built before AI existed.

Re-platform plays typically reach revenue faster than Create plays because the buyer journey is established. The obstacle is inertia, not education. Inertia yields to a sufficiently large cost-and-capability advantage — which the gap between a legacy architecture and an AI-native one provides, and which widens every quarter that the underlying models improve.

The four models are not fixed positions. A successful Consolidate play accumulates the longitudinal data that may make a future Replace move viable within the same market. A Re-platform play that has captured an installed base can extend into adjacent categories that only become possible once the underlying data is centralized. The model identifies the mode of entry. Sequencing across modes is the longer game.

The AI transition is the most significant platform shift since the emergence of the mobile internet. It will create companies worth more than any created in previous platform shifts, because it is the first platform shift in which the underlying capability is not just a distribution channel but an intelligence substrate: it makes things possible that were not previously possible, not just things cheaper that were previously expensive.

The returns from this transition will not accrue to the infrastructure layer. They will accrue to the application layer: specifically, to the companies that build AI-native operating models that replace, not improve, the industries structured around artificial scarcity, and that do so with the structural advantages necessary to move through market validation before the window closes.

The discipline is the edge. In an environment where AI capabilities are broadly available, where most new companies describe themselves as AI companies, and where the labs themselves are expanding their application surface, the companies that will win are not those with the most general capability but those that have made the most precise choices: the right industry, the right buyer, the right structural advantage, in the right sequence, at the right moment in the transition.

The window is open now. How long it remains open is uncertain. What is certain is that the companies being built today, with genuine insight into where the structural replacement opportunities are, and genuine discipline about how to build them, will define the next decade of the industry.