The Part of Nadella's "Frontier Ecosystem" You Can Build Now

Satya Nadella published an essay this week. I'd hand it to every CEO I work with. One line carries it: "You can offload a task, or even a job, but you can never offload your learning."

His argument is that every company now runs on two kinds of capital. Human capital, the judgment, relationships, and pattern recognition of your people. And token capital, the AI capability you build and own. The risk he names is the one I've been raising with clients all year: rent a model, stop there, and you watch your own expertise get commoditized out from under you.

I made the front half of this case a couple of weeks ago, when Cloudflare cut the people it called "measurers". The knowledge worth keeping is the kind no model was trained on: the operator who knows which customer gets the exception, and which step exists because something broke years ago. Keeping it is step one. Nadella's essay is step two, what you build once you do.

He's right. And he gives a clean test for whether you own it: can you swap the model underneath and keep the "company veteran" expertise your system has built? If a better model ships next quarter and adopting it resets you to zero, you were renting, not building.

Here is where I'd extend it. The mechanics Nadella lists sound like work for a company with a research lab. Private evals. Reinforcement learning environments. A knowledge base you can query. Most firms don't have that. It is what to build toward. You don't wait for it to start.

The same property is within reach now with lighter parts. Take the workflows your best people already run. Encode their judgment into systems that get better each time they're used. Point them at your own data. Keep the model underneath swappable. That is the loop Nadella is describing, and it compounds: every improved workflow makes the next one sharper. It also stays yours, whoever you end up buying your models from.

Nadella named the thing. The companies that win with it won't be the ones running the biggest model. They'll be the ones who started encoding what they know first.