A Warning Dressed as a Thought Piece
Satya Nadella posted something on Sunday that took me a few reads to fully land. It’s dense in the way CEO writing often is, philosophical enough to sound like a thought piece but specific enough to be a warning.
Translated out of CEO-speak, the claim is blunt: AI is on track to turn your company’s expertise into a commodity. Not your tools. Your knowledge.
Software Used to Amplify People. Now It Absorbs Them.
For a long time, companies bought software to make their people more productive. The tools amplified the work, but the people created the value; software was a multiplier, not a replacement.
What Nadella is pointing at is different. AI models can absorb what your organisation knows, learn from your data and your workflows, and then make that knowledge available to anyone who pays for access.
The thing you spent years building quietly becomes a standardised service. Your competitors don’t need to develop the expertise themselves. They can just buy a version of yours.
Why the Model Isn’t the Moat
Nadella’s sharpest line is that the model you chose is not your competitive advantage. And if it’s the only thing you’ve built, you’re already losing.
Every competitor has access to roughly the same frontier models you do. If your entire AI strategy is “we picked a good one and plugged it in”, there’s nothing stopping a rival making the identical choice next quarter. The model can’t be your moat because it isn’t yours. It’s a subscription.
What’s actually defensible is everything you build around it. The data you feed it that nobody else has. The workflows you’ve tuned over hundreds of real cases. The feedback loops that make it slightly better at your specific problems every week.
That layer is yours. The model underneath it is rented. And rented things don’t compound in your favour.
The Globalisation Parallel
Nadella draws a comparison that I think is genuinely useful. Think about the first wave of globalisation. Entire industrial economies outsourced their way to cheaper unit costs. The GDP numbers looked fine.
But the underlying capability, the knowledge of how to actually make things, left with the jobs. By the time anyone noticed the structural damage, it was already done.
He’s warning that the same pattern could play out with AI. Companies that fail to build their own systems will find their expertise absorbed and commoditised by the models themselves. The economics look clean right now. The hollowing out happens quietly.
Two Kinds of Capital
The useful framework he offers is the split between human capital and token capital.
Human capital is what your people carry. Judgment, relationships, the ability to spot what matters in a messy situation. Token capital is the AI capability your organisation actually owns and controls: not the model you subscribe to, but the systems you’ve built, the data you’ve shaped, the learning loops you’ve designed.
One doesn’t replace the other. Human expertise becomes more valuable as AI grows, not less, because people set the goals and connect ideas across domains.
The organisations that get this wrong treat AI adoption as a procurement decision: find the best model, buy the seats, deploy.
The problem is that approach builds nothing you own. You get efficiency but no compounding. The model gets smarter in general. You get no smarter in particular.
The Loop Is the Thing
The actual work isn’t choosing a model. It’s building the loop where your people make your AI more useful, and your AI makes your people more capable.
Each iteration compounds. Each project leaves the organisation slightly smarter than it was before.
Nadella proposes a good test for whether you’ve built something real. Could you switch to a different foundation model tomorrow without losing the institutional intelligence you’ve built up? If the answer is no, you’ve built something that matters. If the answer is yes, or if the question doesn’t quite make sense yet, that’s worth paying attention to.
The Question Worth Asking
Most of the AI adoption conversation focuses on speed and cost. Those things are real. But they’re also the things your competitors are getting too.
The better question is whether the work your team is doing with AI is making your organisation distinctively smarter, or just faster at generic tasks.
There’s a version of AI adoption that looks like progress on every dashboard and leaves you competitively no better off in three years. Possibly worse, if the expertise your people used to hold has slowly been ceded to a model you don’t control.
What I Take From This
I’m still working this through, and I don’t think there’s a clean answer yet. But the distinction at the heart of it feels right.
You can hire someone else to do a task. You can even replace a whole role. But you can’t let someone else do your learning for you.
The organisations that understand that early are building something that compounds. Everyone else is renting. And one day they’ll look up and find a model has quietly absorbed the thing they assumed nobody could touch.
Read Nadella’s full post: A frontier without an ecosystem is not stable