Middle powers AI race: leverage is a delivery date

aieuropegeopoliticsoperations

Foreign Policy has a piece on the leverage middle powers have in the AI race. The setup: the White House reportedly asked Anthropic and OpenAI to withhold early access to their latest models from the UK's AI Security Institute, so no ally can count on US AI forever. Their proposal is a "grand compute bargain". The US struggles to build data centers at home, allies offer sites with reliable power and predictable governance, and in exchange they get guaranteed access to frontier models.

I like it because it is concrete. Most AI geopolitics writing stays at the level of values and summits. This one says: you have land and power, they need those, so trade.

But read what the authors actually ask allies to offer. Not subsidies. Predictability. Clear criteria, and around two years from project approval to first operation, which they call fast by the standards of most advanced democracies. So the bargaining chip is not the decommissioned industrial site in Germany. It is the ability to say yes and then deliver in two years.

That changes the question. In the middle powers AI race, you don't have leverage because you own a nice piece of land. You have it if you can turn that land into a working data center before the other side finds someone quicker. Leverage you can't deliver on schedule is just a nice slide.

A couple of weeks ago I wrote about European competitiveness and the boring stuff that slows things down: paperwork that stacks, 27 rulebooks, procurement scared of anything new. Here it is again, at a much bigger scale.

Since 2023 my job has been launching new ventures from inside an Italian company, on a strict budget. We launched two in six months, and the lesson I keep coming back to is that speed is not a personality trait. It is a set of decisions about who can approve what, and how many steps sit between an idea and production. Countries are not startups, obviously. But being slow works the same way.

There is a second problem the article doesn't really touch. Say the deal works and Europe gets frontier models at the same time as the US. Then what? If a public administration needs a year to buy a tool, and a mid-size company has nobody who can integrate it, guaranteed access just means guaranteed shelfware. Model access is the input. Teams that ship are the output.

So if I had to write a European AI strategy on a napkin, I would put the operational stuff first:

  • Regulation with clear criteria, so a project that meets them gets approved without a three-year negotiation.
  • Procurement that can actually buy from new suppliers, including young ones.
  • Making it easy to hire and keep engineers who can build on whatever models we get.

None of that sounds like European tech sovereignty in a speech. But I think it is closer to sovereignty than a sovereign model nobody uses. The authors admit their deal means accepting the US stays ahead on frontier AI. I find that uncomfortable, but for the next few years it sounds realistic.