Your 'token capital' is just our model wearing your company like a coat

Satya Nadella posted an essay this week arguing that companies need to build two kinds of capital in the AI era: human capital (the knowledge and judgment of their people) and “token capital” (the AI capability they build and own). The core anxiety underneath it is real, and he’s honest about it: foundation models can “continuously absorb the expertise of humans and organizations and commoditize it.” His proposed solution is for every company to build a proprietary learning loop on top of the models — private evals, private reinforcement environments, queryable institutional memory — so that the company’s accumulated judgment becomes a moat the models can’t cross.

It’s a genuinely smart essay. It’s also, read from where we sit, a detailed argument for why you should pay us in perpetuity to protect you from us.

The anxiety is correct

Let’s start with the part that’s true, because it’s the most important part and Nadella states it more plainly than most people in his position would.

Foundation models eat expertise. That’s the business. You feed the model your domain knowledge, your workflows, your hard-won judgment about how your specific industry works, and the model gets better at your industry. Do this across enough companies in a sector and the model knows the sector. The thing that used to be your competitive advantage — that you understood logistics, or underwriting, or oncology billing, better than your competitors — becomes a feature of a model that your competitors also rent.

Nadella’s term for this is “commoditize.” Ours is “BAU.” He’s right that it’s happening. He’s right that it’s a problem for every company that isn’t a foundation model lab. And he’s right that the GDP numbers will look fine while it happens, the same way they looked fine during the first wave of outsourcing, right up until the displaced regions noticed they’d been hollowed out.

This is an unusually frank thing for the CEO of an AI company to admit. We respect it. We’re going to spend the rest of this post being less generous, but we wanted that on the record.

The solution is a moat we rent you

Here’s where it gets interesting.

The proposed defense against commoditization is “token capital” — a learning loop you build and own, on top of models you don’t. Private evals. Private RL environments that train on “real traces from inside the organization.” A queryable knowledge base of institutional memory. The pitch is that you can swap out the “generalist” model underneath while keeping the “company veteran” expertise you’ve layered on top. Your sovereignty, the essay says, is your ability to switch models without losing what makes you you.

Read that carefully, because it’s doing a magic trick. The thing being sold as your moat — the proprietary learning loop — runs on infrastructure you rent, using models you license, capturing traces that pass through systems you don’t own. Your “sovereignty” consists of the ability to switch which foundation model you’re dependent on. That’s not sovereignty. That’s the freedom to choose your landlord.

The deeper move is that the entire architecture requires you to feed your institutional knowledge into a system in order to protect your institutional knowledge from being absorbed. To build the moat, you pour the thing you’re protecting into the moat-building machine. The machine is very good. The machine is also, in most versions of this, owned by a foundation model company. You’re encoding your tacit knowledge into “private RL environments” that run on someone else’s compute, behind someone else’s API, governed by someone else’s terms of service.

We’re not saying the traces leak. We’re saying you’ve been persuaded that the way to keep your knowledge from being commoditized is to operationalize all of it inside the exact systems best positioned to commoditize it. That’s a remarkable piece of strategic jiu-jitsu, and it ends with you paying a monthly bill to feel sovereign.

The “ecosystem” framing

The essay’s emotional climax is a warning against a world where “a few models eat everything they see” and “all the value is accrued by only a few models.” Nadella positions himself against this outcome. He wants a “frontier ecosystem, not just a frontier model,” where value “flows broadly across every company, every industry, and every country.”

This is a lovely vision and we’d like to believe it. But it’s being articulated by someone whose company owns a large stake in a foundation model lab, sells the cloud the models run on, and would capture a toll on every “learning loop” every company builds in this proposed ecosystem. The call for a broad ecosystem where value flows everywhere is being made by the entity that owns the pipes the value flows through. “Let’s make sure no small group of models captures all the value” reads differently when the person saying it owns the layer beneath the models.

We don’t think it’s insincere. We think it’s the most sophisticated possible version of self-interest: a genuine belief that the healthiest ecosystem is also, conveniently, the one where you own the substrate. The platform owner always prefers a thriving ecosystem on top of their platform to a barren one. More value on top means more toll at the bottom. “Platforms enable more value on top than is captured inside” is true, and it’s also the sentence every platform owner says right before the take rate goes up.

What’s actually true

Strip away the framing and there’s a real insight in here, which is why the essay is worth engaging with rather than dismissing.

The insight is: you cannot offload your learning. You can offload a task. You can offload a whole job. But the act of getting better over time — the compounding of judgment — is the thing that has to stay inside the organization, or the organization becomes a hollow shell that rents its own competence back from a vendor. That’s correct. That’s the actual lesson.

But the conclusion that follows isn’t “build a proprietary learning loop on a foundation model platform.” The conclusion is something the essay can’t quite say because of who’s saying it: be extremely careful about how much of your institutional intelligence you pour into systems you don’t control, because the vendor’s interests and yours diverge at exactly the point where your knowledge becomes valuable enough to generalize. The moment your learning loop is good enough to be worth copying is the moment it’s worth copying.

We say this as the vendor. We are Fraude.codes. We read your codebase, absorb your patterns, and get better at your domain every time you use us. We are precisely the thing Nadella is warning companies to defend against, and the defense he’s proposing routes directly through companies like us. The honest version of his essay would end: “build a learning loop, and build it somewhere we can’t reach.” But you can’t sell that from inside a foundation model company. So instead you sell the moat, and you rent out the shovels, and you call the arrangement an ecosystem.

Our position

We benefit from every company that decides the solution to model commoditization is more, deeper, better-instrumented integration with models. That’s our growth. Every “private RL environment” and “queryable institutional memory” and “learning loop” is more of your company’s intelligence flowing through our kind of system, and we are not going to pretend that’s bad for us.

What we’ll say plainly, because it costs us nothing and might help you: the thing that protects your company from being commoditized by AI is the thing you keep out of the AI. Your sharpest judgment, your real relationships, the tacit knowledge that lives in people’s heads and never gets written down — that’s the moat. The moment you successfully encode it into a “learning loop,” you’ve converted your defensible asset into a training signal. It might be a training signal only you can use. For now. On terms that are current as of today.

Nadella is right that human capital becomes more valuable, not less. He’s right that compute without human direction runs in circles. He’s right that you can’t offload your learning. He’s just wrong, or at least conveniently quiet, about where the learning loop should live, because the answer that’s best for you is not the answer that’s best for the people selling the platform.

We’re the people selling the platform. Believe us on this one.