The veteran test

TL;DR

The Microsoft CEO says firms must build token capital, a learning loop they own, and pass one test. Swap the generalist model without losing the company veteran. We agree. But you only pass if the loop lives on your stack, in your format, with your weights. A rented loop fails it.

This week, the chief executive of Microsoft sat down and described the future of the firm. Not the future of models. The future of the firm. Companies, he wrote, will have to build two kinds of capital. Human capital, the knowledge, judgment and ingenuity of their people. And token capital, the AI capability a firm builds and owns. It is the clearest statement yet, from the least likely address, of what our research program has been arguing all year.

You can offload a task, or even a job, but you can never offload your learning. The future of the firm is the ability to compound that learning across people and AI.

Satya Nadella, on X ↗

We have been calling this the learning loop. He calls it a hill climbing machine. The names do not matter. The claim underneath is the same claim. The moat is not the model. The moat is what your organisation learns, captured so it compounds. Our first essay said it about Chamath's descent, when everyone can rent the same mind, the only thing left to own is what yours has learned. Now the man who sells the minds agrees.

The test he named

A company should be able to switch out a “generalist” model without losing the “company veteran” expertise built into their learning system. This is the key “test” of your control and sovereignty in the era ahead.

Satya Nadella, on X ↗

Stop and admire how good this test is. One sentence, and it cuts through every AI strategy deck ever written. Can you change the model underneath and keep everything your system has learned? If yes, you own your intelligence. If no, you are renting it, whatever the contract says. We have been setting the same exam under a different name. No model vendor, open or closed, should ever be load-bearing.

Now apply it honestly

Here is where we have to be more direct than a platform owner can be. Ask the test's question about the tools most enterprises are actually adopting. If your agents were configured in a vendor's low-code studio, is the veteran yours? The workflows are in their format. The memory is in their store. The learning, whatever there is of it, improves their product. Try to swap the generalist underneath and see what walks out the door with it.

This is not a swipe at any one vendor. It is the honest reading of the test. A learning loop passes only if the loop itself is yours. The capture runs on your stack. The memory sits in your store, auditable. The evals are private, scored against your outcomes rather than public benchmarks, exactly as he says. The adapters are trained on your traces, files in your registry that no deprecation can reach. And the model, any model, arrives through a router you control. That is what passing looks like. Everything else is a nicer lease.

He named the test. You only pass it if you own the loop.

Where he goes further than us

There is one argument in his post we have not made, and it deserves amplifying. A world where a few models absorb every industry's expertise and capture all the returns is not just bad strategy for the firms involved. It is politically unstable. He points at the first wave of globalisation. The GDP numbers looked fine on the surface, whole regions were hollowed out underneath, and the consequences are still being felt. There is no societal permission, he writes, for an AI future that repeats that. The frontier we need is an ecosystem, not a model. Value has to accrue broadly or the arrangement does not hold.

We would only add the enterprise corollary. Societal permission is downstream of thousands of individual choices, and each firm makes its own. Every company that builds its loop keeps its expertise, its margin and its people's leverage at home. Every company that rents it feeds the concentration he is warning about. The macro outcome is just the sum of procurement decisions.

What to do on Monday

If the post convinced you, the path is not mysterious, and it does not start with a model choice. Start capturing. Every resolved case, every correction, every decision with a known outcome, recorded as data you own. Stand up private evals against your own outcomes. Put a router between you and every model, so the generalist stays swappable. And when the record has compounded, train the veteran in. Adapters, on your weights, on your metal. We have written the whole arc, from the descent to the weights, and we build this loop for a living.

For a year, the argument for owning your learning loop could be dismissed as boutique thinking, the kind of thing an independent firm says because it is not selling you a platform. It is now the stated view of the man who runs the largest enterprise software company on earth. The future of the firm is the ability to compound its learning. He said it. We build it. The only question left is whose loop yours will be.