Satya Nadella has a line that should make every founder uncomfortable: "AI is not a technology. It's the future of the firm." Most CEOs are still treating it like a technology - they do a press release with a tech vendor, point to eight agents they built, claim an AI strategy, and move on. Nadella thinks that's a category error on the scale of saying you don't really know what the human capital in your company does.
His sharper framing is the one worth stealing. Every firm, he says, now runs on two kinds of capital: human capital and "token capital." Human capital you understand. Token capital is new - it's the knowledge your business compounds when humans and AI work together, captured as context, skills, or model weights that you own and control. And here's the test he'd put to any CEO: can you concretely say what token capital you built last night? Across all the work your company did, what got turned into something you own, can point to, and have set up to compound? If you can't answer that, you don't have an AI strategy. You have a vendor.
The reason this matters so much comes down to a warning most leaders haven't internalized: your tacit knowledge is leaking, and it's a one-way door. Every company's real edge is its tacit knowledge - the taste, the judgment, the unwritten ways work actually gets done. Nobody has a balance-sheet line called "tacit knowledge," but we all assume we own it because it lives in our people. Nadella's point is that AI can now extract that tacit knowledge from the trajectories of how work happens and encode it into model weights. If those trajectories flow out to someone else's foundation model, the edge goes with them. His tell: watch how the model companies actually learn - they hire people who used to work at your company and set up reward gyms to reconstruct what those people knew. Once it leaks, you're done.
So the move he's pushing is an inversion. Don't send your data and your work out to the frontier model. Bring the model in. Let it "hill climb" inside a machine you control - your data is its context, you collect the traces of how humans and agents get work done, and you keep that loop as IP instead of leaking it. The frontier model is a guest in your house, not your landlord.
That reframes what the new IP even is. It is not the frontier model - everyone will rent those. The IP is your evals and your objectives: knowing exactly what outcome you want and how to score it. "Everything else is pretty mechanical," Nadella says. Knowing what data to train on and how to reward it is where the next layer of value gets created. For an operator, that's the whole game in one sentence - the durable asset is your definition of "good," not the engine that chases it.
Which leads to his most quotable operating rule: don't use frontier models for non-frontier problems. If you're processing retailer trade-promotion claims, that's a repeatable, deterministic workflow - not a hunt for new material science. Throwing a frontier model at it is the AI equivalent of a thousand-dollar Pentagon toilet seat. Take a small model, fine-tune it on your own traces, and it can outperform a prompted frontier model at a fraction of the cost. In an age of token abundance, the person who figures out how to use tokens efficiently for outcomes that matter gets ahead by definition. Capital allocation didn't go away; it moved to tokens.
He sees the shape of the future in how coding already changed: autocomplete, then chat, then agent mode for small tasks, then full fire-and-forget autonomy where you hand over a high-level intent and accept a finished pull request. Knowledge work is next. The funny tell is that engineers now run a hundred agent sessions at once and drown in the cognitive load - so we're rebuilding the IDE, except now it's an inbox of agents you steer. Twenty employees, twenty million agents, all in a loop. The job becomes macro-delegation plus micro-steering.
Two ideas round it out. First, the human skill that becomes scarce isn't expertise - expertise is now abundant and cheap. It's "cognitive coverage": your ability to actually understand and verify what the agent did. One of his colleagues built a tool that quizzes you on the agent's work, the way test coverage checks code. That's the literacy of the next decade. Second, the same logic scales to countries: real sovereignty isn't firewalls and data residency, it's keeping your companies thriving in a token economy. Off the frontier, you fall behind; dependent on a single frontier model, you aren't sovereign. The answer is the same as for firms - hill-climb your own, one company at a time.
The throughline, and the part Nadella is most insistent on, is that none of this earns itself the right to happen. The industry has spent two years telling people their jobs are gone while asking them to be excited about the technology erasing them - and people have stopped believing us, "rightfully so." Permission gets earned with tangible, local, boring proof: a data center that genuinely helps the town's tax base and schools, concrete new jobs at nameable wages, every firm actually participating in the frontier instead of being a feeder of data to someone else's model. Lose that, and the best technology of the century gets regulated into a corner.
For anyone building a company right now, the assignment is simple to state and hard to do: stop buying AI, and start compounding token capital you own.
No comments:
Post a Comment