💶 Money’s here.

At Monday’s Choose France summit, the French government’s annual pitch to foreign investors, SoftBank pledged up to €75bn for data centres across the country. One site, at Bosquel in northern France, will reach a gigawatt, the output of a nuclear reactor. SoftBank brings the money; the operation falls to Sesterce, a small French firm that rents compute to businesses. The site is the very plot Fluidstack, the British operator, walked away from a year earlier, having chased bigger American contracts.

Attracting this capital is welcome. But money was never the scarce ingredient. At last year’s Paris AI summit, French AI had already drawn €109bn in pledges. The bottleneck that matters is turning those cheques into gigawatts that actually run.

Connecting a gigawatt to the grid takes years of queuing, a problem as familiar in Britain and Germany as in France. Feeding it clean power means signing fifteen-year contracts, scarce when everyone is building at once. Cooling servers at that density while meeting the service levels you have promised is engineering few have mastered. These obstacles are measured in years of planning and procurement. The money is already in the room.

The standard objection is that the value sits in the models or the chips, not the server sheds. Partly true, and worth conceding. But models are commoditising: labs multiply, open weights eat the margins, and Mistral is fighting in a field where any lead melts within six months. The leading-edge chips stay American. Europe holds one link upstream, ASML’s lithography machines, and little else. Running infrastructure at this scale is the step nobody replicates quickly. Investors can be replaced; the operator who can keep a gigawatt humming, far less so.

The money comes from Tokyo, the best customers sit in the United States, and Fluidstack left precisely to serve them. Holding the operating layer is not the same as holding the whole chain. Europe supplies the site and runs the machines, while a good share of the value is booked elsewhere. That part the story does not resolve, and the optimists tend to skip it.

Bosquel is about more than servers. The race for models remains largely American, carried by speed and capital. But AI has bolted on a heavy physical layer: concrete, megawatts, contracts that outlast three product cycles. On that ground Europe turns up with two things it is rarely credited for: decarbonised power the Americans lack at scale, and the industrial patience it is usually mocked for, now an asset rather than a handicap. The advantage goes to whoever can build in the real world. A genuine position, not a consolation prize.