The wild world of putting numbers on the AI build-up craze

The EU and national governments are dead-set on improving Europe’s place in the global AI race – but it’s not really clear what they’re going up against.

The Commission has pledged billions of EU funding for its planned AI gigafactories – dedicated data centres packed with AI chips for training large AI models – while smaller AI factories, which draw on the existing EU supercomputing network, are already opening across the continent.

Private companies are also pumping money into building out AI infrastructure in Europe: Only on Tuesday, several German companies announced a plan to spend €1 billion on Nvidia chips for an AI data centre in Munich, meant to make AI available to industrial players.

But where will all these investments and plans put Europe on the AI capacity map? Frankly, it’s hard to say exactly.

A big reason for this fuzziness is that this wave of AI investment is being driven by private companies – which don’t need to disclose how powerful their AI infrastructure is. That doesn’t stop company CEOs from hyping up plans for another ginormous AI data centre, of course, but whether these compute hubs really materialise as promised is another matter.

There’s a bona fide business to be had here. US research firm Semianalysis builds and sells a widely used AI power dataset by compiling information on more than 5,000 AI projects around the world. US thinktank Epoch AI also released a similar dataset on Tuesday – though, for now, it’s focused on the US.

When it comes to tracking AI capacity, a key data source both firms use is, curiously enough, satellite imagery.

“The resolution is high enough that we can see cooling infrastructure on the outside of the buildings,” explained Ben Cottier, a researcher at Epoch. He measures the diameter of these fans, counts them up, and uses that information to estimate how much power the data centre consumes – since powering the facilities generates waste heat that needs to be cooled.

How the AI hubs generate all this power is another way to measure capacity, according to Semianalysis researcher Jeremie Eliahou Ontiveros. “You can look at the substation, you can look at the transformer, you can look at diesel generators,” he said.

Meanwhile, US companies continue cranking up their AI build-ups. Several data centres with a capacity that exceeds one gigawatt are due to come online over the pond in 2026. Epoch points out that’s comparable to the electricity consumption of entire countries – such as Rwanda or Barbados.

The EU’s four or five AI Gigafactories, meanwhile, still only exist on the drafting board, with the next step – selecting actual sites – expected by the end of the year.

While these planned EU AI training hubs do – currently – place somewhat highly (just outside the top 30) on Epoch’s dataset of the most powerful AI centres globally, the think tank estimates they would have only around a tenth of the power of the colossal AI hubs that US hyperscalers are likely to bring online by 2026.

(nl, aw)