The Compute Settlement

Every era has a piece of infrastructure that quietly decides who gets to participate in the economy and on what terms. In the nineteenth century it was rail. In the twentieth it was the electrical grid and, later, the network. In ours it is increasingly large-scale compute — the data centres, power contracts, and supply of advanced chips that determine who can train and run the most capable systems.

The pattern with this kind of infrastructure is consistent. It is built fast and privately during a land-grab phase, it becomes a bottleneck, and then society spends decades and enormous political energy trying to retrofit public accountability onto something that was never designed to accept it. We argue the compute build-out is in its land-grab phase now, and that the window to set its terms is short.

This brief does not call for nationalising anything. It makes a narrower claim: that the questions usually asked only of regulated utilities — who gets access, at what price, under what reliability and transparency obligations — are the right questions to be asking of frontier compute today, while the answers are still cheap to change.

The default — doing nothing deliberate — is itself a choice, and usually the most expensive one.

We sketch three governance postures a state could take — laissez-faire, light-touch standards, and explicit public-works treatment — and trace the likely consequences of each over a fifteen-year horizon. None is free. The point of laying them side by side is that the default, doing nothing deliberate, is itself a choice with consequences, and usually the most expensive one.

The settlement we favour is unglamorous: a small set of access, transparency, and reliability standards, set early, that keep the market competitive without pretending the market will handle distribution on its own. The history of every prior infrastructure suggests that the cost of acting early is a fraction of the cost of acting late.