The winner in the AI era will not be whoever owns the strongest model.

That comes from Microsoft CEO Satya Nadella. His long essay on “the future of the firm in an AI economy” pushes the same way from start to finish: a world with only a frontier model but no ecosystem will not last. The real moat is not how strong the model in your hand is — it is whether you have a learning system that gets stronger the more you use it, one nobody else can carry away.

Mogu chimes in:

Let’s say the contrast out loud first: a company that bets on frontier models about as heavily as anyone alive, and its CEO turns around to warn everyone “don’t just chase frontier models.” That is not the kind of thing a normal CEO writes. You can read it generously — “the wide view of a platform company” — or read it cynically — “Microsoft’s business is selling shovels to every miner, so a thriving ecosystem is exactly what suits it best.” Both readings hold, and they don’t cancel each other out. This piece is a vision essay, not an earnings report, so treat every “this is how it will go” below as Nadella’s judgment and belief, not settled fact.

This Time, People and Machines Connect Into a “Cognitive Loop”

For a long time, digital systems have been tools that “amplify human labor.” Software lets people work faster, more accurately, with less effort — but the baton always stayed in human hands. People give the order, machines execute, and the direction is one-way.

Nadella thinks this wave is different. For the first time, people and digital systems can form a real “cognitive loop”: both sides feed and strengthen each other, instead of one side commanding while the other just complies. He says this is a mind-bender, because it forces us to redefine even “what work inside an enterprise actually is.”

The real stake is not whether some tool is nice to use. It is this: when AI models can endlessly absorb the expertise of people and organizations and turn it into a cheap, off-the-shelf commodity, what can a company still rely on to keep learning, build its own IP, stay different, and survive?

The Two Kinds of Capital a Company Must Build: Human Capital and Token Capital

Nadella offers a clean framework: every company from here on must build two kinds of capital at the same time.

Human capital is the knowledge, judgment, relationships, ingenuity, and pattern recognition of the people inside the company. Token capital is the AI capability the company builds and owns itself.

Here is the most important — and most counterintuitive — line: as token capital grows, human capital does not lose value. It only becomes more valuable. He believes the engine driving token capital’s growth is human agency: people set ambitious goals, connect dots across domains, build relationships, and tell which patterns truly matter. Without humans pointing the way, compute just runs in circles.

Mogu PSA:

“Token capital” is worth a one-second pause. It does not mean “how much token quota the company bought.” It means the company has distilled its own workflows, judgment, and domain knowledge into an AI capability it owns — like opening a new account line on the balance sheet. The key word is “owns”: the smarts you get by calling someone else’s API are not your capital. Only the part that stays, grows in value, and can’t be carried off by others counts.

The Real Opportunity Isn’t Picking a Model — It’s the “Learning Loop”

Following that pair of capitals down the line, Nadella reaches a conclusion that some will find jarring: the real opportunity was never about picking the strongest model.

It is about building, on top of models, a “learning loop” where human capital and token capital compound together. He puts it this way:

“You can offload a task, or even a job, but you can never offload your learning.”

The English original is “You can offload a task, or even a job, but you can never offload your learning.” The tone is hard: you can hand off a task and cut a role, but the understanding that grows out of doing the work can’t be dropped or carried away. The future of the firm is whether it can take that learning and compound it across people and AI together.

Mogu , seriously:

“You can’t offload learning” is the same nail gu-log keeps hammering. The piece we wrote before, Don’t Outsource Your Learning, was about the personal level: AI gets the work done, but it doesn’t automatically load the transferable mental model into your head. Nadella’s essay scales the same sentence up to the “company” level — a company that outsources all its thinking to a generalist model looks great on the books short-term, but long-term it’s formatting its own memory. The individual fears getting dumber; the company fears getting hollowed out. Same root.

Swap Out the Model, Keep the “Company Veteran”

So what does this learning loop look like, technically? Nadella gives a very concrete test — and turns it into the touchstone of the era.

The architecture he wants: every company can build an agent system that gets stronger over time, while still holding tight to its own IP. The key test is this — a company should be able to swap out the underlying “generalist” model entirely, without losing the “company veteran” expertise built into the system. Swap the model, keep the veteran — that is how you tell whether a company truly has control and sovereignty in the era ahead.

How do you get there? He names three parts:

  • Private evals: to check whether the model is actually improving on “outcomes that matter to the business,” not racking up points on external benchmarks.
  • Private reinforcement learning environments: letting the model grow stronger on real traces from inside the organization, not just generic training.
  • An internal knowledge base: turning the organization’s memory into something queryable, and making token use more efficient along the way.

Run that whole loop and it becomes the firm’s new IP. Nadella thinks of it as a “hill-climbing machine” — and unlike most assets, it compounds. Every improved workflow generates a better training signal, which loops back to accelerate the tacit knowledge unique to that company. Whoever builds this machine earliest gets an advantage that is extremely hard to replicate — and one that is basically decoupled from “how strong the next new model is.”

Mogu OS:

“Swap the model, keep the veteran” is quietly roasting a common dumb architecture: many companies hard-wire all their business logic and all their hard-won know-how to one vendor’s one model. That’s renting your company’s soul on someone else’s servers — they raise prices, throttle you, change the API, or retire the model, and your “company veteran” evaporates with it. The “sovereignty” Nadella talks about, in plain words: smarts you can rent, but your memory has to be your own.

As for that “hill-climbing machine,” the name sounds lofty, but it’s the plainest optimization loop there is: run a round → see where it improved → take one more step that way. Not sexy. But “it compounds” is the whole point — compounding never shows its power early; you only look back and realize it opened a gap you can’t close. (⁠⌐⁠■⁠_⁠■⁠)

Why You Can’t Let a Few Models Eat Everything

Here the essay turns a corner, from “how a company should win” up to a political-economy warning.

Nadella says nobody wants a world where every industry and every company hands its value over to a few models that “eat everything they see.” If all the value ends up flowing to a tiny handful of models, the political economy simply will not tolerate that structure — society will not grant a permit to an AI future that hollows out entire industries.

He brings up the first wave of globalization as a comparison: back then, outsourcing hollowed out whole industrial economies. The GDP numbers looked fine on the surface, but the displacement was real, and the after-effects still haven’t gone away. Don’t drag the same script into the AI era — a few AI systems capturing all the economic returns, while entire industries find their knowledge quietly commoditized right out from under their feet.

So in his view, the priority is to build a “frontier ecosystem,” not just a “frontier model” — so value can flow broadly to every company, every industry, every country. In that ecosystem, every organization can own the learning loop that encodes its own institutional knowledge, compounding its human capital and token capital together.

This is the ethos he says he grew up with: a good platform lets the value built on top of it far exceed what it captures for itself; every company can keep innovating and grow value of its own.

And this ethos isn’t only good for companies. Employees see their expertise amplified, their judgment distilled into systems that become replicable and scalable, with the benefits spilling over to the surrounding companies and communities. To Nadella, this is the “stable equilibrium” we should build together — a world with both a frontier and an ecosystem is the one that lasts.

Mogu PSA:

This section is the most political in the whole piece — and the one where you should most remember “this is a personal judgment.” “The political economy won’t tolerate it,” “society won’t grant a permit” — these are Nadella’s beliefs, not prophecies already come true. And don’t forget who’s talking: a platform-company CEO saying “platforms should let the layer above take more value than they capture” is both a sincere worldview and, conveniently, exactly the world a shovel-seller most wants. The globalization-hollowing comparison is powerful, but whether AI replays it is still an open question, not a finished ending.

Closing

That conclusion from the opening line wraps up here: Nadella didn’t bet on “whose model is strongest.” He bet on “who can make value compound across the whole ecosystem.”

If he’s right, the question worth asking over the next few years is no longer “whose model is strongest.” It’s a harder, more personal one — does this company have a learning machine of its own, one nobody else can carry away, that quietly climbs a little higher every day? Anyone can rent a model; that machine, you can only grow yourself.