The chief executive of the company pouring tens of billions of dollars into building frontier AI models has a counter-intuitive message for everyone else: stop obsessing over which model is best. In a new post on X, Microsoft CEO Satya Nadella laid out a theor
y of how companies will survive — and compete — in an economy where AI can absorb and resell human expertise. His central claim is that picking the smartest model is the wrong race. The real prize, he argues, is the learning loop a company builds on top of whatever model it happens to use, and crucially, the one it owns.
To frame the stakes, Nadella introduces a new pairing. Every firm, he says, will have to build two kinds of capital. Human capital is the familiar one: the knowledge, judgment, relationships, ingenuity and pattern recognition of its people. Token capital — the coinage that is now circulating well beyond Microsoft — is the AI capability a company builds and owns. (The “token” is the basic unit AI models consume and produce; here it stands in for an organisation’s homegrown AI muscle.)
The provocative part is what he says about the relationship between the two. Human capital, in his words, does not shrink as AI grows — it becomes more valuable. People set the ambitious goals, connect dots across domains and spot the patterns that matter; the machine supplies scale. “Without human direction,” he writes, “you have compute running in circles.”
The architecture Nadella sketches suggests firms stop treating AI as a single, swappable foundation model and start building agentic systems — AI that can take actions and run workflows, and, importantly, that improves with every use.
The test of whether a company actually controls its AI, he says, is simple: it should be able to swap out a “generalist” model for a newer one without losing the “company veteran” expertise baked into its system. In other words, the model is rentable and replaceable; the accumulated institutional know-how layered on top is the asset you cannot afford to lose.
Three pieces make that loop work, and each is worth translating:
Private evals — a company’s own scorecard that measures whether a model is getting better at the outcomes the business cares about, rather than against generic public benchmarks. Private reinforcement learning environments — sandboxes where models are trained and sharpened on the firm’s real internal data and decision traces. A queryable knowledge base — institutional memory made searchable, which also makes the AI’s token use more efficient.
Nadella calls the result a “hill climbing machine”: unlike most corporate assets, it compounds. Every improved workflow produces better training signal, which deepens the firm’s tacit knowledge, which improves the next workflow. Companies that start early, he argues, build a lead that is hard for rivals to copy — and that holds even when the next big model arrives.
Underneath the management advice is a sharper political argument. Nadella warns against a future in which a handful of AI models “eat everything they see” and capture all the economic returns, leaving entire industries to watch their expertise commoditised out from under them.
He reaches for a historical parallel: the first wave of globalisation, when outsourcing hollowed out industrial economies. The headline GDP figures looked healthy, he notes, but the displacement was real and is still being felt. His blunt verdict on a repeat in the AI era: “There is no societal permission for an AI future that hollows out entire industries.”
The alternative he proposes is what he calls a “frontier ecosystem, not just a frontier model” — value spread across companies, industries and countries, with every organisation owning the learning loop that encodes its own knowledge. He frames it as an extension of the classic platform bargain: platforms thrive when they let others create more value on top than the platform itself captures.
The timing is not incidental. Nadella’s argument lands fresh off Microsoft’s Build 2026 developer conference and echoes themes he aired on Reid Hoffman’s Possible podcast, where he made a similar case for human and token capital compounding together. It also arrives against a fraught backdrop: the world’s largest tech companies have announced more than $700 billion in AI capital expenditure for 2026, fuelling investor anxiety about over-building — and pointed questions about who, exactly, will capture the returns.