Funding your own replacement

Satya Nadella, Microsoft's chief executive, published an essay this month with the shape of a warning. Every firm, he argued, must now build two kinds of capital: human capital, the judgment and relationships of its people, and the "token" kind, the AI capability a company owns and controls. The task is to wire them into a learning loop that compounds, one that lets a firm swap a general-purpose model in and out without surrendering the accumulated, company-specific expertise stacked on top. His sharpest line is also his simplest: "you can never offload your learning." It is good advice, and it describes something already going wrong that Mr Nadella is careful not to name.

What the loop defends against is an extraction operation well underway, and it announced itself, in capital terms, a few weeks before the essay. In May, two of the world's leading AI labs spent a combined $5.5 billion to become consulting firms. OpenAI stood up a deployment company valued at around $10 billion, anchored by roughly $4 billion in committed capital and seeded with about 150 engineers it acquired from Tomoro, an Edinburgh consultancy. Anthropic launched a competing services venture worth $1.5 billion, backed by Blackstone, Goldman Sachs, and Hellman & Friedman. Both copied the same template, pioneered by Palantir: embed full-stack engineers inside a customer's walls and build production systems on top of their own models.

For an ordinary consultancy, the fee is the business; for an AI lab, it is almost beside the point. What the arrangement actually buys is proximity: a seat inside a Fortune 500 company, watching how the work really gets done.

Consider what a forward-deployed engineer actually delivers. Inside the client, the engineer builds the evaluation suites that catch a model's mistakes, the retrieval pipelines that feed it a company's internal data, and the agents that wire it into legacy software. The premium such work commands is rational: MIT's NANDA initiative, after studying 300 enterprise AI projects, found that 95% produced no measurable effect on profit, and the model itself was rarely the reason. Getting it to function inside compliance regimes and workflows never designed for it was the hard part, and competent implementation remains scarce.

But the deliverable has a second life. The eval suite that flags a bank's hallucinations and the workflow that reshapes an insurer's claims process are also the precise raw material a lab needs to make its next model better at banking, or at claims. Enterprise contracts increasingly forbid training on client data, and that restraint is real, yet it misses what is actually being extracted. Even when the data never leaves the building, the lab departs having learned the contours of the domain: which tasks carry weight, where the model breaks, what good looks like in a business it had never before seen from the inside. That tacit knowledge is the scarce input to post-training, which means the labs have arranged to be paid a premium to acquire the very thing they would otherwise have to buy.

Embedded interest

None of this is sentimental: OpenAI's share of the enterprise model-API market fell from roughly half in late 2023 to a quarter by mid-2025, by Menlo Ventures' estimate, while Anthropic climbed to 32% and pulled well ahead in code generation. The model has stopped being the differentiator; what now separates the winners is whoever can move the thing into production, work that draws, by an old enterprise rule of thumb, six dollars of services spending for every dollar of software. Losing the API race, OpenAI responded by building a consulting company.

Which sets up the choice every large enterprise now faces. It can keep buying, and grow dependent on a vendor whose models improve partly by absorbing its business; or it can begin to in-house the loop itself — the models, the evals, the reinforcement-learning environments, the workflows. The eval and the RL environment are collapsing into a single artifact, and that artifact, rather than any individual model, is the asset worth owning.

That this is feasible is no longer theoretical. When Cursor, the AI coding startup, shipped its Composer model this spring as proprietary, a developer reading its API traffic found something underneath: Moonshot's Kimi K2.5, a Chinese open-weight model, carrying Cursor's own training on top. Cursor later confirmed it, noting that the dominant share of the compute, roughly 85% for the latest version, went into post-training rather than the base. The result reaches close to the coding performance of frontier proprietary models at about a tenth of the per-token cost. Holding the open base fixed for sixty days, Cursor's own training added between six and eleven points across coding benchmarks. The base model was the commodity. The loop stacked on top was the moat, and the product it ships inside is the lock-in.

The catch is talent. Cursor employs machine-learning researchers most Fortune 500 companies cannot hire, which is the precise gap the platform layer now sells into. It is also the logic beneath Europe's "sovereign AI" ambitions, which on inspection amount to the same maneuver at national scale: post-train an open base, then run it on domestic GPUs.

Which returns us to Mr Nadella, and to the part of his essay that goes unspoken. The remedy he prescribes is correct, and it runs, conveniently, through his own infrastructure. Microsoft amended its OpenAI partnership in April, unveiled seven in-house MAI models at its Build conference on June 2nd, and now runs OpenAI's, Anthropic's, and its own models side by side in Foundry. The forward-deployed arrangement is precisely the offloading of learning he warns against; the loop he urges firms to own is one he would prefer they rent from Azure. While the labs sell the extraction, Microsoft sells the antidote.

The decision facing the firm, then, is the oldest one in corporate finance, make or buy, with a wrinkle no spreadsheet captures. Buying rents a capability and, with it, hands the firm's learning curve to a counterparty who will sell it onward. The curve may depreciate regardless — every model generation resets the baseline, and a loop tuned to this year's frontier may be worth little against next year's. Perhaps. But if competent implementation is a genuine edge, the firm that rents it is paying a premium to school its own competition. Mr Nadella is right that you can never offload your learning; the going rate for trying, this year, was $5.5 billion.