What truly matters are the assets built on top of the models.

Over the past year, plenty of AI founders have hit a point where they declare, "We're AI-native." But what actually counts as an AI-native workflow?

It could mean a company that has genuinely fed its workflows into a model, growing smarter with use. Or it could mean a company that's merely wrapped someone else's large model — one that would have to be torn down and rebuilt if the vendor changed tomorrow. In a demo, the two are often indistinguishable.

Satya Nadella, CEO of Microsoft, published a piece last week titled A frontier without an ecosystem is not stable — one of the more interesting reads to come out recently.

Not because it introduced many new concepts. Most of its key points have been circulating for the better part of a year. What makes it notable is who said it: someone personally running massive AI infrastructure. And Nadella, in plain language, clarified two things that matter enormously: What does it actually mean for a company to possess AI capability? And where will the value from this cycle ultimately flow?

Satya's post drew significant attention on X.

Below are the four most discussion-worthy points from the article, followed by a full translation at the end for those who want to read it themselves.

01. The Dual-Capital Model: Human Capital + Token Capital

A core argument in Nadella's piece is that every company will need to build two forms of capital going forward: the human kind (human capital), and the AI capability that a company constructs and truly owns (token capital).

But before getting to this dual-capital model, there's something more fundamental worth mentioning first: the human capacity to bear responsibility. It is the bedrock for everything else.

This year everyone has been arguing over whether AI will replace humans. But the more pressing question is this: Is there something that AI is structurally incapable of taking over? The answer is "the capacity to bear responsibility." An AI can be infinitely capable, yet when it screws up, the blame ultimately lands on a person.

For any system to function, there must be a subject that can be held accountable: someone whose assets can be seized, who can face legal and economic consequences. AI is impervious to punishment. And it is precisely "being punishable" that enables a person to truly bear responsibility. This sits at the forefront because it undergirds everything that follows.

With that in mind, returning to Nadella's framework: First, this wave differs fundamentally from any previous technology cycle. PC, internet, mobile, cloud — all were essentially about "augmenting" humans, giving people faster tools, but the human remained the sole agent doing the work. This time, for the first time, human and system have formed a genuine cognitive loop. The system has shifted from a role of handing tools from the sidelines to a participant that works alongside the human, even proactively suggesting what to do next.

Because of this, Nadella says every company must build two forms of capital going forward: human capital and token capital. Human capital is the human element: knowledge, judgment, relationships, creativity, pattern recognition. Token capital is the AI capability that a company builds and truly owns. Notably, token capital is an asset that simply did not exist on any company's balance sheet five years ago, yet is now becoming one of the most important assets for a modern company.

As for the relationship between these two capitals, Nadella argues: Human capital does not depreciate as token capital grows. On the contrary, it only becomes more valuable. He goes so far as to say that human agency is the engine of token capital growth — because it is humans who set ambitious goals, connect dots across domains, build relationships, and identify which patterns actually matter. Without human direction, the compute power in your hands is just spinning in place.

But Nadella anchors human value in "agency" and "judgment," and both of these are likely merely阶段性优势 (phase-specific advantages), not the fundamental boundary between humans and AI. Agency, at its core, stems from Maslow's hierarchy of needs — yet this mechanism of "wanting, then actively reaching" is being replicated in code. Proactive agents, auto-research tools — these are already digging up their own needs, proposing their own goals, as long as you bake it into the system. Judgment, too: human judgment is largely trained out by data, not innate, and "learning judgment from data" is precisely what models are getting better at.

Of course, for the present moment, judgment remains scarce and valuable — when AI compresses "execution" into a near-zero-cost commodity, knowing what is worth doing, knowing when an output counts as "right" (even if technically flawless), knowing where the high-leverage decision points sit in a complex system. These still lack formulas, and they are indeed sorting people into two camps: those who lived on execution speed are being squeezed, those who lived on judgment are being amplified. But this is no long-term moat. What remains truly long-term, and what cannot yet be offloaded to AI, returns to that earlier point: who bears responsibility, who can be punished.

At the company level, this becomes a new high-leverage form: a small cluster of people with strong enough judgment who can also own the results, plus sufficient token capital, producing what a far larger team once could. This is precisely what truly AI-native companies look like now — per-capita output compounding upward, headcount staying flat or shrinking.

Ultimately, the future company form places scarce judgment (and that final accountability) at the center, then scales execution behind it with as much token capital as possible.

02. The Learning Loop: A Moat That Grows Smarter With Use

So how exactly is token capital "built" and "owned"?

Nadella believes the real opportunity lies not in picking the "strongest model," but in building a learning loop on top of the model — turning a company's workflows, domain knowledge, and accumulated judgment into an AI system that grows smarter with use.

Specifically, three things: 1) Private evals, measuring whether the model improves on the business outcomes that actually matter, not just external leaderboards; 2) A private reinforcement learning environment, where the model continuously strengthens based on real internal interaction trajectories; 3) Plus a knowledge base that makes institutional memory instantly accessible.

He calls this mechanism a "hill-climbing machine" that keeps ascending toward better solutions, or simply: "it compounds." Unlike most depreciating assets, every improved workflow generates better training signals, which accelerate the沉淀 (precipitation/accumulation) of tacit knowledge unique to that company, which in turn improves the next workflow.

To illustrate: two companies, same model, same business, same industry — yet five years later they could be worlds apart, with the only difference being that one built up this learning layer and the other didn't.

The model itself is a replaceable "input"; the accumulated knowledge is the uncopyable asset that retains value even if the model is swapped. Nadella puts it well: you can offload a task, even an entire job, but you can never offload your own "learning and growth."

Going one level deeper: this so-called "learning loop" is essentially an upgrade from the prompt. The most primitive prompt is a sentence you throw at a model. The most complete prompt is an entire environment — one that contains clear evaluation criteria (what counts as "good"), readable and writable state and memory, and the real trajectories continuously generated within it.

The model is more like an "amplifier": give it a sentence, it squeezes that sentence to its extreme (you give it a horse, it assembles a car; you give it a city, it can折腾 (work/disturb) an empire). So the real craft lies not in which model you pick, but in how complete an environment you can feed it.

The frontier beyond "building a good environment" is how that environment evolves itself. The most朴素 (simple/naive) way is linear iteration: run once, observe results, revise. But others are already pursuing different paths — approaches like AlphaEvolve, closer to DNA: fission into many variants at once, throw them into the environment to collide, keep the "survivors," then fission and mutate again; Go-style self-play is another update mechanism. "How to update" itself is becoming a craft with real nuance.

So the question remains open: is a "complete environment" the end of context? Or is there something larger, still dimly seen, that wraps around it?

03. If You Swapped the Model, What Would You Still Have?

As noted, the moat is the learning loop, not the model itself.

Going further, Nadella proposes a test: Can a company swap out a "general-purpose" model without losing the veteran-like private expertise that has already settled into its learning system?

In other words: If you changed models, would the moat still stand?

This test works well when applied to specific companies: For a self-proclaimed AI-native company, if the underlying model were swapped from one provider to another, or even to a future model that doesn't yet exist, would it maintain its operational advantage? Would the accumulated "veteran experience" in the system remain intact? Would the encoded workflows, context, and operational judgment still function as before?

With Anthropic restricting Mythos access to a small number of institutions and Fable 5 being banned, "model availability" has suddenly become a major headache for countless CEOs and CTOs, highlighting how fragile it is to bet your lifeblood on a single model.

If swapping a general-purpose model leaves that painstakingly accumulated "veteran experience" intact, you've genuinely built token capital — the model is merely a replaceable component for you. If the swap breaks you, then your so-called moat was rented from the model vendor — and if they raise prices, change behavior, or shut off API access one day, your barrier evaporates overnight.

Companies that truly pass this test tend to share several traits: infrastructure that remains model-neutral, internal evaluations that track business outcomes rather than any particular model's strengths, and accumulated interaction trajectories that can be replayed on any model.

04. Don't Let a Handful of Models Eat Everyone

Here Nadella pulls the lens back from individual companies to the entire economy — and this is precisely the most important question: Where will value ultimately flow?

Some context is needed first. On the surface, Microsoft is OpenAI's largest external investor, so one might think it "should most want value to pool at the model layer." For Nadella to come out arguing "value should flow above the model" is easily read as noble selflessness against self-interest.

But in reality, Microsoft is simultaneously a shareholder in model companies, a cloud provider, and an entry point for model usage. Pretty much whatever "fission" happens in this ecosystem falls within its coverage; it's hard for something to emerge that could disrupt it entirely.

In other words, he wins however this plays out: at minimum, he captures value wherever it lands; offensively, his most comfortable scenario is precisely model companies not dominating alone, with plenty of agent companies and model companies coexisting, none able to swallow the others — which happens to be exactly the "ecosystem" Nadella's article advocates for. So rather than nobility, this is someone who wins regardless, staking out the arrangement most favorable to himself.

A side note: Elon Musk's entire response to this piece was a single word — "Interesting." Someone else seated at this same table, leaving such a cryptic "Interesting," invites multiple readings.

Last August, when Nadella announced full GPT-5 integration into Copilot, GitHub, and Azure, Musk had thrown down: "OpenAI will eat Microsoft alive." Nadella's piece today — "don't let a few models eat everyone" — is in some sense a response to, even an implicit validation of, Musk's remark from back then. So Musk's "interesting" also carries the light prick of a bystander who sees through the game.

What Nadella now fears is a world where every company across every industry hands its value over to a few models that "eat everything they see."

What gets eaten is nothing less than the professional knowledge, judgment, and IP assets that industries have painstakingly accumulated. Once these are commoditized, slowly siphoned out from underfoot, companies are left with nothing to differentiate themselves from competitors.

In his view, if all value ultimately pools into the hands of a tiny handful of model companies, the political economy simply won't tolerate it. An AI future that hollows out entire industries will not obtain societal permission.

He deliberately invokes the first phase of globalization as analogy: industrial economies one after another were hollowed out by outsourcing. GDP numbers looked fine on the surface, but the dislocation was real, its consequences still being paid for today. That script should not be replayed in the AI era.

And history's lesson is that such a reckoning comes sooner or later — the only question is whether the industry adjusts proactively while still profitable, or waits to be forced by crisis.

Ultimately, two outcomes lie before us: either value concentrates highly in a tiny number of models, or it distributes broadly across many layers — cloud providers, platform builders, vertical AI companies, open-source ecosystems, infrastructure, and a whole crop of new companies growing atop this foundation, even society itself. Only then can value flow to every company, every industry, every nation.

Behind this lies a rather朴素 (simple) "platform value": a platform's health is measured by whether the value created atop it far exceeds what it extracts for itself.

This is precisely what Nadella's title means: A frontier without an ecosystem is not stable.

---

Full Translation: A frontier without an ecosystem is not stable

By Satya Nadella (Chairman and CEO, Microsoft)

@satyanadella

I've been thinking lately about what the future of the firm looks like in an AI-driven economy.

This shift is different from any platform change that came before it. In the past, we used digital systems to augment human capital; this time, for the first time, we can create a cognitive loop between people and digital systems. This is paradigm-shifting because it changes how we think about work itself inside an organization.

What really matters here is not how any particular digital tool or system is used, but this: in a world where AI models can continuously absorb and commoditize the specialized knowledge of people and organizations, how does an organization keep learning, building IP, differentiating, and thriving?

Every company will have to build two things, which I call human capital and token capital. Human capital is the knowledge, judgment, relationships, ingenuity, and pattern recognition of employees. Token capital is the AI capability that a company builds and owns itself.

What matters is that human capital does not depreciate as token capital grows. On the contrary, it only becomes more valuable! I believe human agency will be the driver of token capital growth. It is people who set ambitious goals, connect dots across domains, build relationships, and identify the patterns that matter most. Without human direction, compute is just spinning in place.

This means the real opportunity is not in picking the "best model," but in building a learning loop on top of the model, where human capital and token capital compound together. You can offload a task, even an entire job, but you can never offload your own learning. The future of the firm depends on making this learning compound continuously between people and AI.

To do this requires a new architectural approach: enabling every company to build agentic systems that improve themselves over time, while retaining control over their own IP. A company should be able to swap out a "general-purpose" model without losing the "company veteran" expertise that has already settled into its learning system. This is the litmus test for whether you will have control and sovereignty in the future.

Companies need to turn their workflows, domain knowledge, and accumulated judgment into AI systems that grow smarter with use. Private evals should measure whether a model improves on the business outcomes that actually matter, not just external leaderboards. A private reinforcement learning environment should make the model stronger based on real interaction trajectories inside the organization. Its knowledge base should make institutional memory instantly accessible and use tokens more efficiently.

This loop becomes the company's new IP. I think of it as a hill-climbing machine: continuously ascending toward better solutions. And unlike most depreciating assets, it compounds. Every improved workflow generates better training signals, which accelerate the沉淀 (precipitation) of tacit knowledge unique to that company, which in turn improves the next workflow. The earlier a company builds this mechanism, the harder its advantage is to replicate — and this advantage persists regardless of how far any individual model's capabilities leap ahead.

None of us want to see a world where every industry and every company hands its value over to a few models that eat everything they see. If all value ultimately pools into the hands of a tiny handful of models, the political economy simply won't tolerate it. An AI future that hollows out entire industries will not obtain societal permission.

Think about the first phase of globalization: industrial economies were hollowed out one after another by outsourcing. GDP numbers looked fine on the surface, but the dislocation was real, and its consequences are still being borne today. Don't let this replay in the AI era: a few AI systems capturing all economic returns, while entire industries watch their knowledge be commoditized and slowly siphoned out from under them.

In my view, our top priority should be building a frontier ecosystem, not just a frontier model, so that value can flow broadly to every company, every industry, and every nation. In such an ecosystem, every organization can own its own learning loop, encode its institutional knowledge, and let human capital and token capital compound together.

This is what I have always believed: a good platform is one where the value created by others on top of it far exceeds what it takes for itself; on such a platform, every company can keep innovating and grow its own value.

When this happens, companies will create value for themselves and for the economy around them. Employees will see their expertise amplified, their judgment entering systems that replicate and scale it; and these benefits will flow to the companies and communities around them.

This is how a company creates value for itself and for the larger economy. This is the stable equilibrium we should build together.

---

If Satya's perspective interests you, or if there's anything you'd like to discuss with us, feel free to reach out — we're always here: