Human Capital, Token Capital, and the Climbing Machine: Why the Next Moat Is Owning Your Learning Loop

Two companies in the same industry buy access to the exact same frontier model. Same weights, same API, same price per token. A year later one of them has pulled away and the other is quietly being eaten. Nothing about the model explains the gap, because the model was identical. So what did?

The answer is the thing almost nobody put on a slide: one of them was learning and the other was just consuming. One built a system that got smarter every time an employee used it, the way a muscle gets stronger under load. The other rented intelligence by the hour and gave it all back at the end of every call.

That is the whole argument of this post, and it is worth saying plainly before we earn it. The model is becoming a commodity. The moat is the learning loop you build around it. Satya Nadella framed this idea in an essay arguing that a frontier without an ecosystem is not stable, and the part that should keep founders and operators up at night is not the geopolitics of it. It is the quiet economics: if you do not own a system that captures and compounds your own knowledge, you are renting your future from whoever does.

“

You can offload almost any task to AI. You cannot offload the learning. The company that confuses the two is handing its expertise to a model it does not control.

”

The commodity nobody wants to admit is a commodity

Start with the uncomfortable part. The frontier model, the thing everyone is racing to access, is on its way to becoming infrastructure. Not worthless, infrastructure. Electricity is enormously valuable and you do not build a business on owning the grid unless you are a utility. You build a business on what you do with the current.

Frontier models give you broad capability off the shelf: they can read, write, reason across domains, call tools, and operate over messy natural language better than any system we have ever shipped. That capability is real and it is rising fast. But here is the catch that the breathless coverage skips. Broad capability that everyone can buy cannot, by definition, be your advantage. If your competitor can sign up for the same endpoint on a Tuesday afternoon, whatever you do with the raw model out of the box is a feature, not a moat.

This is the trap I watch teams walk into. They treat “we use the best model” as a strategy. It is not a strategy, it is table stakes, and table stakes have a way of becoming free. The interesting question is never which model you call. It is what your system retains after the call ends.

Two kinds of capital: human and token

To see where the real value lives, it helps to split a company’s intelligence into two accounts.

The first is human capital, and it is older than software. It is the knowledge in your people’s heads, the judgment they apply when the situation is ambiguous, the relationships they hold with customers and partners, the creativity that produces an option nobody briefed, the pattern recognition a veteran has that a newcomer does not. This is the stuff that has always been hard to copy and hard to hire, and the thing most leaders underrate is that it gets more valuable as AI gets more capable, not less.

That sounds backwards, so let me show the mechanism. As models get better at executing, the scarce skill shifts up the stack to the things models do not do for you. Someone has to decide what is worth doing. Someone has to notice when the AI is confidently wrong and pull the wheel. Someone has to hold the customer relationship that no model attends. Someone has to point the whole apparatus at an outcome that actually matters instead of a metric that merely moves. Human agency, the act of setting goals and correcting course, becomes the rarest input precisely because everything downstream of it got cheap. When execution is abundant, direction is the bottleneck, and direction is human.

The second account is newer, and Nadella’s name for it is the sharp one: token capital. This is the AI capability a firm builds, owns, and improves over time. Not the model you rent, the system you grow on top of it: your prompts and tools and evaluations, your fine-tuned behavior, your retrieval over your own documents, the accumulated record of what worked and what did not in your specific domain. Token capital is what turns generic capability into your capability. It is the difference between a customer support team that uses ChatGPT and a customer support system that has learned, over a hundred thousand of your own tickets, how your products actually fail and how your best agent actually answers.

Two accounts, then. Human capital sets the direction and holds the relationships. Token capital does the compounding execution and never forgets. The companies that win this decade run both, and wire them together so each one feeds the other.

The compounding loop is the actual product

Here is where it stops being a metaphor and becomes machinery. The reason token capital is a moat and not just a nice-to-have is that it compounds, and compounding is the most underestimated force in business because it is invisible until suddenly it is not.

Walk the loop one turn at a time. You start with a workflow, say your sales engineers answering technical questions during deals. You instrument it so the AI is in the loop and every interaction leaves a trace: the question, the context, the answer, and crucially whether it worked. Those traces are training signal, the realest kind there is, because they come from your actual work and not from a benchmark. You use that signal to improve the system: better retrieval over your docs, better-tuned behavior, evaluations that catch the failure modes you actually see. The improved system produces better answers, which produce better traces, which deepen the system’s grip on your specific domain. Round and round.

The strategic consequence is blunt. Better workflows create better training signal, better signal creates a stronger system, a stronger system deepens your firm-specific knowledge, and deep firm-specific knowledge is the one thing your competitor cannot purchase. This is the engine. Everything else in the AI conversation, the model benchmarks, the demo videos, the price wars, is noise around this single quiet loop.

“

Stop asking which model is best. Start asking which workflow, if it learned from every single run, would compound into something a competitor could never copy. That workflow is where your AI budget belongs.

”

AI sovereignty: own the learning, rent the model

This brings us to the idea I think is the most important and the most ignored, and it deserves its own name. Call it AI sovereignty. Here is the test for it, and it is a single clean question.

Could you swap out your underlying model tomorrow, replace it with a different vendor’s, and lose none of your accumulated expertise?

If the answer is yes, you are sovereign. Your knowledge lives in your loop, your data, your evaluations, your institutional memory, and the model is a swappable engine underneath. If a better or cheaper model ships next month, you slot it in and your moat is untouched, maybe even sharper. If the answer is no, if your competence is so entangled with one provider’s specific model that changing it would lobotomize your operation, then you do not own your advantage. You are leasing it, and the landlord can change the terms.

This is the part teams get backwards. They obsess over which model to marry and ignore whether they could ever divorce it. True sovereignty has nothing to do with picking the right model. It comes from controlling the learning system and preserving your company-specific knowledge across model changes, so that the model is the one part of the stack you are happy to treat as disposable.

Why this matters now and not in some abstract future: model leadership keeps changing hands. The frontier is a leapfrog race, and the leader this quarter may not lead next quarter. A firm whose entire competence is bolted to one model is making a bet on a horse race that has no permanent winner. A firm whose competence lives in its own loop simply rides whichever horse is fastest, and pockets the difference. Sovereignty is not patriotism for your vendor. It is refusing to let any vendor own the part of you that matters most.

What the architecture actually has to do

So what do you build? Not a chatbot. A system with two jobs that pull in slightly different directions, and getting both right is the engineering challenge of the decade.

The first job is to improve over time. The system has to be agentic enough to act in your workflows, and instrumented enough that every action becomes signal it can learn from. A static integration, the kind where you call an API and move on, does not improve. It just runs. You want the opposite: a system whose competence curve bends upward because it is wired to learn from its own operation. If you have read about reinforcement fine-tuning with outcome-based rewards, this is the practical reason it matters at the firm level, your real outcomes become the reward signal that shapes behavior. And if you have thought about building agent loops as a first-class engineering discipline, this is why that discipline is now a strategic concern and not just an implementation detail.

The second job is to protect what the loop produces. As the system learns your domain, it concentrates your IP: your judgment, your data, your hard-won patterns, the very things that make you you. That concentration is the asset, which means it is also the thing most worth defending. The architecture has to keep that knowledge under your control, portable across models, and out of any pipeline that would quietly feed your expertise back to a provider who then sells it to your competitor. A learning loop that improves the vendor’s model more than it improves yours is not a moat. It is a leak.

Those two jobs are in tension. You want a system open enough to learn aggressively and closed enough to keep what it learns. Resolving that tension well, an architecture that compounds your knowledge while keeping it yours, is the technical heart of this whole shift.

How to actually build the loop

Principles are cheap. Here is the concrete version, the four things that separate a real learning loop from a fancy wrapper around someone’s API.

• Run private evaluations tied to business outcomes, not public benchmarks. Public leaderboards tell you how a model does on someone else’s test. They tell you almost nothing about how it does on your work. The teams that win build private evals from their own cases: did the answer actually resolve the support ticket, close the deal, pass the code review, satisfy the customer. When a new model drops, you do not ask “where does it rank.” You run it against your private eval and ask “does it make our specific work better.” That eval set is itself token capital, it took real effort to build and a competitor does not have it.

• Train and improve on real organizational traces, not synthetic fluff. The signal that matters comes from your actual workflows and the feedback attached to them, the messy real record of what your people and your AI did and whether it worked. This is the data no one else has and no one else can generate, because it is the exhaust of your specific operation. Synthetic data has its uses, but the gold is the trace from a real interaction with a real outcome, and most teams throw it away because they never instrumented for it. Start capturing it now; you cannot retroactively learn from work you did not record.

• Make institutional memory searchable, reusable, and token-efficient. Knowledge that sits in a wiki nobody reads is not capital, it is a liability with a search bar. The point of the loop is that what one person or one agent learns becomes available to every person and every agent that comes after. That means your memory has to be retrievable on demand and compact enough to fit in a context window without burning your whole token budget, because every interaction that re-derives knowledge you already had is paying twice for the same lesson. Token-efficient memory is not a nicety. At scale it is the difference between a loop that pays for itself and one that bankrupts you.

• Continuously compound factual knowledge across humans and AI. The loop is not human-only or AI-only, it is the two together. A person corrects the AI, that correction becomes signal, the improved AI frees the person for harder judgment, that judgment generates new knowledge, and the cycle keeps spinning. The firms that treat this as one combined system, rather than “the AI team” and “the humans” as separate org charts, are the ones whose knowledge actually accumulates instead of leaking out the door every time someone quits or a contract ends.

The downside case: an economy that forgets to share the climb

Now the part that is bigger than any one company, and the reason Nadella titled his essay around stability rather than just strategy. If this all goes one way, if a tiny number of frontier models capture most of the value, the result is not just unfair. It is unstable, in the plain economic sense that an unstable system does not stay where it is.

Picture the failure mode concretely. A handful of frontier providers get so capable that they commoditize the expertise of whole industries. The specialized knowledge that a firm spent decades accumulating becomes something a general model can approximate well enough, and the firm’s reason to exist thins out. Repeat that across sector after sector and you hollow out the middle: the firms that held the domain expertise lose the ability to charge for it, and the value flows up to the few who own the models.

The trap here is that the aggregate numbers can look great while this happens. Total economic output rises, the headline productivity figures climb, and everyone congratulates themselves, even as the gains pool in a few places and real displacement spreads underneath the average. An average is a terrible place to hide a distribution. A boom that lifts the aggregate while workers, firms, and whole communities lose control over the value of their own knowledge is not a healthy boom. It is a concentration dressed as a recovery, and concentrated systems are brittle. They wobble.

I want to be careful here, because the doom version of this is lazy. The point is not that concentration is inevitable. It is that concentration is the default if no one builds the alternative, and the alternative is buildable. That is the hopeful turn, and it is the one that should change what you do Monday.

A frontier needs an ecosystem to stand on

The stable version of this future has a specific shape. Progress holds when the ecosystem is broad: when every company, every industry, and every country can create value on top of the frontier and keep it. Not just access the models, retain the upside of using them.

The cleanest way to say it is a rule for platforms, and it is the rule that separates a healthy platform from a parasite. A good platform enables more value to be built on top of it than it captures for itself. When the people building on your platform get to keep the lion’s share of what they create, they keep building, the ecosystem thickens, and the platform grows with it. When the platform captures most of the value, the builders leave or wither, and the platform inherits a desert it cannot farm alone. The frontier model providers who understand this will win the long game precisely because they take less of each transaction and grow a larger transaction.

And the firm-level move that makes the whole thing stable is the one we have been circling the entire post: own your learning loop. When a company owns its loop, its expertise is amplified by AI rather than absorbed by it. The model makes your best people’s judgment scalable instead of replaceable. Your knowledge gets multiplied and rewarded instead of commoditized and extracted. Multiply that across enough firms and you get the broad, value-retaining ecosystem that keeps the frontier stable, because everyone in it has a stake in its continuation.

“

The model is the climbing wall, available to everyone. The learning loop is the climbing machine you build, and it is the only part that keeps you moving up after the wall stops being a novelty.

”

So here is where it lands. The frontier model is a commodity in the making, extraordinary and increasingly available to all. What it cannot give you is the one thing that matters most: a system that captures your work, learns from it, protects it, and compounds it into an advantage that is unmistakably yours. Build that, and the model underneath becomes a detail you are happy to swap. Skip it, and you will spend the next decade renting your own future from someone who was paying closer attention. The two companies on the identical model from the opening were never really competing on the model. They were competing on whether they bothered to learn. You get to choose which one you are.