France and the EU announced big AI investments
How relevant are they to training frontier AI?
At the French AI Action Summit, both the French national government and the European Commission announced big AI investments, with a mix of public and private money.
As someone who focuses on the governance of the most advanced AI systems, I was interested in the extent to which these funds are relevant to developing frontier models.
(By “frontier models” I mean the most capable AI models at a given point in time. These tend to be general-purpose and very large. Correspondingly, they are trained with a large amount of computational resources, i.e. “compute”. Current examples of frontier models include Claude 3.5 Sonnet and GPT-4o.)
Of course, a wide range of investments could in some ways support frontier model development. For example, even initiatives to implement existing small models into specific applications could conceivably contribute to frontier model development by raising public excitement about AI.
What I assess here is the extent to which the investments are for infrastructure that would be relevant for frontier models in particular, such as very large datacenters and energy infrastructure to provide the huge amounts of power that these datacenters require.
Here are my findings in brief:
France announced €109 billion of private money. Around €65 billion of this seems to me to be relevant to frontier model training, though it depends a lot on definitions.
The European Commission announced €200 billion of public and private money. Around €20 billion of this seems relevant to frontier model training.
These investments seem like important evidence that France, and to a lesser degree the EU, will have the necessary compute for frontier model training.
France
President Emmanuel Macron announced €109 billion of investments by the private sector in French AI. The official list of projects is here. Three projects, listed below, seem particularly relevant to training frontier models.
1. Up to €50 billion to create a 1 gigawatt campus “dedicated to AI in France.” This is from a Franco-Emirati consortium led by MGX.1
“Campus” is a term for a group of datacenters located on the same site. Datacenters in the same location can relatively easily be networked together to be used for the same training run. If the investors do this, the 1GW campus would be of a scale required for frontier model training. For example, Fist and Datta write:
The largest AI clusters being built today have around 100,000 accelerators, consuming tens of megawatts of power. Based on the current plans of U.S. firms (and provided sufficient power is available), by 2030, the largest clusters will require closer to one million accelerators, consuming around five gigawatts (GW) of power.
2. €10 billion to create a 1 gigawatt “supercomputer for AI.” This is from Fluidstack, a British cloud provider.
Supercomputers are not always designed specifically for AI, but this one fairly clearly is. The press release from Fluidstack says that there will be 500,000 “next-generation AI chips”. It also describes the facility as “advancing AGI” and “purpose-built to train advanced AI.” When comparing the project to the Fist and Datta numbers above, we see the investment is indeed of the scale required for training frontier models.2
One uncertainty I have: The two projects apparently aim to create similarly sized AI compute clusters but come with very different price-tags. The Emirati project is expected to cost up to five times as much. My guess is that this is mainly because the Emirati figure is a high estimate whereas the Fluidstack figure is a low one. The French government elsewhere described the Emirati investment as costing between €30 and €50 billion. In contrast, the Fluidstack press release seems a little vague about what is being funded by the “initial” €10 billion investment, versus what will subsequently be funded.
3. €5 billion for AI-focused energy infrastructure. This is from Apollo, an American investment firm.
Datacenters for training frontier models are very energy-intensive, as discussed above. This means they are more likely than other datacenters to require significant new energy infrastructure. That said, I haven’t seen confirmation that this is the purpose of this specific investment.
Bonus: ~€25 billion for various projects to create hundreds of megawatts of geographically distributed AI datacenter capacity. (This is a ballpark figure because the document is sometimes ambiguous about the amount of new funding for these projects.)
Several projects in the French list have these two characteristics:3
The project would create >100 megawatts’ of AI datacenter capacity, and;
It is stated, or I expect, that the capacity will be spread across different parts of France rather than in the same location.
Looking at the Fist and Datta numbers, hundred-megawatt projects are at a relevant scale. However, training across multiple locations is much less efficient. This makes it seem less likely that these datacenters are intended for frontier model training.
(It might be feasible to train frontier models across multiple geographically distributed locations, if the developer creates significant bandwidth between the various locations. The short descriptions in the projects lists don’t mention plans to create this bandwidth, though.4)
Overall view
These are big investments in the ability to train frontier models in France. The total “headline figure” of the three investments that I highlighted above is €65 billion. As a more conservative estimate, the total cost could be around €40 billion. This uses the lower estimate of how much the Emirati campus could cost, and excludes the Apollo energy infrastructure, which isn’t necessarily for frontier model training.
European Commission
European Commission President Ursula von der Leyen announced €200 billion for an “InvestAI” initiative. 50 billion of this comes from redirecting existing European Commission R&D funding, with the remaining 150 billion from “mobilizing” private sector capital. Her speech is here and the accompanying press release is here; my information primarily comes from these sources.
In her speech, von der Leyen said that the focus of the money would be on AI applications:
AI has only just begun to be adopted in the key sectors of our economy, and for the key challenges of our times. This should be Europe's focus. Bringing AI to industry-specific applications and harnessing its power for productivity and people. This is where Europe can truly lead the race.
That said, the announcement does include €20 billion for building four “gigafactories.” These are datacenters that seem intended for frontier models.
The EU's InvestAI fund will finance four future AI gigafactories across the EU. The new AI gigafactories will be specialised in training the most complex, very large, AI models. Such next-generation models require extensive computing infrastructure for breakthroughs in specific domains such as medicine or science. The gigafactories will have around 100,0005 last-generation6 AI chips, around four times more than the AI factories being set up right now.
100,000 cutting-edge chips in a cluster is appropriate for training the kind of models that the press release describes. As mentioned above, Fist and Datta write that the largest clusters being built today have around one hundred thousand GPUs.”
The sources also mention ongoing EU efforts to support “AI factories”, i.e. smaller AI datacenters, with €10 billion. Smaller datacenters are less likely to be relevant for frontier training, unless (as discussed above) they are well-networked together.
The numbers in context
If France and the EU successfully turn their plans into datacenters and energy infrastructure, how relevant would this all be for frontier model training?
The investments would be sufficient for AI models roughly at the current frontier. Fist and Datta write that GPT-4 required around 25,000 GPUs and consumed around 30 MW of power. The two big French datacenter projects, as well as the EU’s gigafactories, are all bigger than this.
What about in the future?
Epoch researchers estimate that training runs are becoming around four times bigger per year, and that this trend can likely continue (at least) until 2030. They estimate that such training runs would require infrastructure costing hundreds of billions of dollars.
Along these lines, OpenAI may be working on a datacenter that would cost up to $100 billion, with millions of chips, and energy requirements of up to 5GW. The Information reported in 2024 on the plans for this datacenter, which was called Stargate and to be built with Microsoft. (“Stargate” now refers to a larger—$500 billion—set of investments in AI infrastructure more broadly, though with Microsoft less centrally in the picture. Reporting seems unclear on whether the ‘new’ Stargate still includes a $100 billion datacenter.)
The EU Commission investments—€20 billion for the largest datacenters, split between four sites—are much smaller than the massively scaled datacenters envisioned just above.
On the other hand, the French investments are broadly comparable (though still smaller). In comparison to OpenAI’s reported $100 billion and 5GW, the French-Emirati project is €30-50 billion and 1GW.
Macron has compared the French investments to Stargate. This was not bluster.
The press release for this specific investment also highlighted the possibility of additional joint investments, such as in talent development.
“AI chip” and “accelerator” have similar meanings—see footnote three in Fist and Datta.
These are the projects from Brookfield (€15 billion), Digital Realty (>€5 billion), Eclairion (unclear amount), Evroc (up to €4 billion), Iliad (unclear amount), Prologis (€3.5 billion), and Sesterce (€400 million).
I would be interested to see someone go through more detailed descriptions of the individual investments to see whether they include efforts to create this bandwidth.
The phrasing seems a little ambiguous about whether it is 100,000 chips each or in total. From context, I assume “each.” If I am wrong, the gigafactories seem less relevant to frontier model training.
“Last-generation” is a confusing word choice; it could plausibly refer either to the “latest” or to the “previous” generation of chips. I’m fairly confident the meaning is “latest,” partly because that makes more sense in context, and partly because the German version of the press release uses the term “neuste” (“newest”).