Compute is a production input

You do not own the substrate, and you cannot win the scale race. What is left is management.

There are two ways to get the compute layer wrong, and most firms manage both at once. The first is to treat compute as a cost line, a number in the budget, watched by finance and trimmed when times are hard. The second is to treat it as a faith bet, where you raise more, spend more, back the right provider, and trust that scale will carry you. One view is too small and the other is too credulous. The last month has shown why both fail.

The substrate you do not own

On 12 June, at 5:21pm Eastern, Anthropic received a letter from the US government. By that evening it had disabled its two most capable models, Fable 5 and Mythos 5, for every customer in the world, three days after launching them. Not an outage. An export control directive that barred foreign nationals from using the models, including the company’s own foreign-born staff, and since there is no clean way to verify a passport at the API layer, the only compliant move was to switch them off for everyone.

You can have the safety argument elsewhere, and people are having it loudly. The operator’s lesson is simpler and colder. The compute layer is not only expensive. It is contingent. A model you have wired into a live workflow can be withdrawn overnight, by a government on another continent, for a reason you will never be shown. Your deployment was sound. Your vendor did nothing wrong. The substrate moved, and the work stopped.

This is the failure mode the production system calls lock-in, and most teams discover their exposure to it the way you discover any single point of failure, which is all at once, on a Friday evening. The instinct is to fix it by procurement, to diversify providers and sleep a little better. That helps, but it answers the wrong question. The question is not which horse to back. It is what happens to the work, and who carries it, when the substrate you were standing on is gone.

Managing that is unglamorous work. It means knowing which parts of the operation rest on a single model, keeping a tested fallback for the ones that matter, and building the abstraction that lets you swap what is underneath without rewriting the work above. None of it is exciting. All of it is the difference between a supplier problem and a stopped line.

Scale is the entry fee

Now look at the money, because the money tells you how the rest of the market is answering that question, which is by spending. In late May, Anthropic raised 65 billion dollars in a single round, at a 965 billion dollar valuation, briefly worth more than OpenAI. Fifty billion is going into a US data centre build-out. Gigawatts of power are under contract. One reported compute arrangement runs to 1.25 billion dollars a month, to a single supplier. These are real numbers, and they describe a real race.

They do not describe a strategy. At this layer, scale is the entry fee, not the edge. Everyone with capital can buy more compute, and everyone is. The advantage does not sit in the size of the bet.

I keep coming back to one line. The winners of the next decade will look more like Toyota in 1985 than a model lab in 2023. Toyota did not win by building the biggest plant. It won by knowing, to the unit, what each car cost to make, and by removing everything on the line that did not earn its place. The prize at the compute layer is the same, and it cuts against the instinct of the moment. It is not lower cost. It is legible cost.

Legibility is harder than it sounds, and it arrives late. Stable production is the only phase in which you can finally measure what good enough actually costs, and then right-size to it on purpose. Until then you are guessing, and the guessing has a trap in it. Efficiency per token keeps improving while total spend keeps climbing, because cheaper compute invites more of it. That is Jevons, and it is not a paradox to fear. It is a reason to know your own numbers cold, because the bill will rise even as the unit price falls, and only a firm that can read its own consumption will know whether that rise is growth or waste.

Compute as a production input

Put the two halves together and the discipline becomes obvious, and unfashionable. You do not own the substrate, so you cannot fully trust it. You cannot win the scale race, so you should stop trying to. What is left is management. Treat compute the way a manufacturer treats a critical raw material. Know your dependency and your continuity exposure, so that the loss of one supplier is a bad day rather than a stopped line. Know your unit cost, so that you can right-size deliberately rather than spend to keep pace. Run leaner models yourself, at the edge, where the work allows it, not as an act of thrift but as an act of control.

The firms that come through the next decade will not be the ones with the cleverest model or the largest build-out. They will be the ones who can answer two dull questions on an ordinary Tuesday morning. If this model is gone before lunch, what happens to the work? And what does a single unit of that work actually cost us to produce? Most firms cannot answer either. That, and not the next raise, is where the next decade will be decided.