Rent the model. Own the learning.

Here is a prediction you can hold me to: the companies that win the next five years of AI won't be the ones running the best model. They'll be the ones that own what their model learned. I didn't expect the CEO of Microsoft — the company with more riding on frontier models than almost anyone — to make the same bet in public. He just did.

Satya Nadella published an essay titled "A frontier without an ecosystem is not stable." It's been viewed tens of millions of times, and most of the discussion fixed on the ecosystem politics. The line that should actually change how you buy AI is quieter, near the middle: "You can offload a task, or even a job, but you can never offload your learning."

That's the whole thing in one sentence. You can hand the work to an agent. You can't hand over the learning that comes from doing it. And if you're not careful about how you buy, you'll do the first while quietly giving away the second.

The test hiding inside the essay

Nadella's sharpest move is what he frames as a test of your control and sovereignty. A company, he argues, should be able to "switch out a 'generalist' model without losing the 'company veteran' expertise" its system has built up over time.

Read that as a buying criterion, because that's what it is. The model is the part you swap. Models get cheaper, better, and more interchangeable on a schedule measured in months — by the time you've standardized on one, there's a stronger one. So the model can't be the thing you're protecting. The thing worth protecting is everything the system has learned about your customers, your offers, what you've already tried, and what didn't work. If swapping the model also erases that learning, you never owned it. You were renting it.

The one question to ask any vendor

So here is the question worth asking any AI tool, platform, or agency you're about to pay — including us:

If we swap out the model underneath, do we keep what the system learned?

The point isn't that you have to own everything. Plenty of the learning in your stack is commodity — a transcription model, a generic deflection bot that was never going to be your edge — and renting that, letting it live in someone else's platform, is a perfectly good trade. Not everything is worth the cost of owning. The mistake is narrower and more expensive: letting the learning that is your edge — what works on your customers, your offers, your channels, the dead ends you already paid to find — pile up somewhere you don't control. Own the part that compounds into how you compete. Rent the rest on purpose, not by accident.

So the honest version of the question is sharper: for the learning that actually matters to us, if we swap the model — or the vendor — do we keep it? If yes, the accumulated judgment lives in files you hold, and you're building an asset that's yours. If it's a shrug, or "it's all handled inside our platform," then for the part that matters you're renting the compounding — and the day the contract ends, you start that part over from zero.

Owning it is what bends the curve

Here's why the learning that matters is worth keeping: it compounds, and compounding is the rare thing that bends a growth curve over time instead of just lifting it once. Rented work resets every cycle — you pay again to relearn what you already knew. Owned learning stacks: each cycle keeps what worked and what didn't, so the next unit of growth costs a little less to produce than the last. Nadella calls the loop "a hill climbing machine" and notes that, "unlike most assets, it compounds." Over a year, that's the difference between a growth line that needs constant new spend to hold its slope and one that gets cheaper and faster on its own.

We've been making this case from inside the work

The strange part of watching Nadella publish this to the whole market is how familiar it reads. We've been arguing it from the practical side for months — four posts, one spine:

• The gap is not your model. When a rollout stalls, teams reach for a better model. The real gap is that the work was never written down clearly enough for an agent to run it — codify the function and an ordinary model runs it well. The written-down rubric is the asset, not the prompt.
• Is your growth compounding or rented? Two brands can post the same revenue curve — one built a flywheel, one rented a treadmill. The tell is the fully-loaded cost to produce a unit of growth, watched as a trend: falling means the system kept what it learned; flat under rising spend means you pay to relearn it every quarter.
• The CRO tool that "runs itself." Buy the platform, let it optimize, switch vendors eighteen months later — and find the learnings lived inside the vendor's model. The next tool starts from zero, on your budget. Exactly the lock-in Nadella names.
• The real test of an AI engagement. It isn't the demo. It's what's still running in a repository you own six months after the agency stops billing. Installed, not rented.

We didn't have a Microsoft-CEO essay to point at when we wrote those. Now we do — and the argument we'd been making quietly, inside client repos, is suddenly the one the whole market is having out loud. The advantage was never the shiny model. It's the boring, owned learning loop — and that just stopped being the contrarian view.

The part we'll cover in the next post

There's one piece of Nadella's essay we've deliberately left alone, because it deserves its own post. He splits a company's value into two kinds of capital that grow together: human capital — the judgment, relationships, and pattern recognition of your people — and what he calls token capital, the AI capability you build and own. And his claim cuts against the reflex that AI makes people matter less: as the AI side grows, the human side becomes more valuable, not less, because "without human direction, you have compute running in circles."

That reframes what your team is even for in an AI-heavy company. It's the next thing we're writing.

For now, the takeaway fits on a sticky note. Keep the learning that's your edge where you control it, and the cost to produce a unit of growth falls over time instead of holding flat while you spend more. The next time someone sells you on how smart their model is, ask the one question that matters: when we swap it out, do we keep what it learned?