Europe jumps into ‘incredibly costly’ AI supercomputing race
The EU will fund AI-optimized supercomputers across the bloc to help startups compete with U.S. rivals.
BRUSSELS — The European Union has committed hundreds of millions of euros to help startups catch up in the global artificial intelligence race.
But critics wonder whether the bloc knows what it’s getting into.
The European Commission has pledged €750 million for seven sites across Europe to establish and maintain AI-optimized supercomputers accessible to startups to train their AI models, it announced Tuesday. The sum is half of a total investment of €1.5 billion; EU member countries are on the hook for the rest.
It’s a first step toward the EU’s goal of becoming an “AI continent” — a term coined by the EU’s new tech chief, Finland’s Henna Virkkunen.
With these supercomputers, the bloc wants to give European rivals of ChatGPT-maker OpenAI a shot at developing their own chatbots. Training AI models requires vast amounts of computing power. The EU wants to build out such capacity to avoid new reliance on United States Big Tech firms — and to avoid losing another tech race.
Earlier this year, French AI hope Mistral struck a deal with Microsoft to access the U.S. tech company’s supercomputing capacity. Some read it as a sign that Europe was already lagging in the AI race.
The European Commission wants to correct that with a plan that’s off to an ambitious start.
It had previously said that it could spend up to €800 million under the current budget, which runs until 2027. Virkkunen had told lawmakers she wanted to launch five so-called AI Factories in the first 100 days of her term.
Luckily, the bloc doesn’t have to start from scratch.
Three of the chosen AI proposals are in locations where some of Europe’s existing powerful supercomputers are already based: MareNostrum in Barcelona, Spain; LUMI in Kajaani, Finland; and Leonardo in Bologna, Italy.
Cost explosion
However, experts warn that keeping the infrastructure up to date after launch will be key, which can become costly. Meanwhile, the budget until 2027 is already almost entirely already allocated.
Building out AI infrastructure is “incredibly costly,” said Bertin Martens, a senior fellow at think tank Bruegel. Over the last couple of years, there has been “a cost explosion” in building AI cloud and supercomputing capacity.
Those investments also don’t last long.
“You build a new infrastructure, eight months later, [there are] the next-generation models and chips,” Martens warned.
The history of the Barcelona-based supercomputer MareNostrum 5 is an example of how supercomputers must be updated regularly: MareNostrum 3 was installed in 2012, and version 4 came to life around 2017.
Over the years, the supercomputer shifted from being driven solely by more “traditional” CPUs (central processing units) to being partly driven by GPUs, the graphical processors that allow for AI optimization.
“Our current system, the MareNostrum 5, is already a little bit optimized for AI,” said Josep Maria Martorell, associate director at the Barcelona Supercomputing Center.
The Commission selected the project and it’s now bound to get a further upgrade. But that upgrade will have to be more than just a technical one, Martorell said, given that the EU wants the new AI supercomputers to be accessible to startups.
“The need [of] a startup is very different than the need [of] a scientist,” Martorell said. Startups don’t just need computing capacity; they also need data storage, which will have to be added to the supercomputing facility.
Over time, European AI startups will have to generate revenue to justify the Commission’s initial investment.
It’s hard to say whether they’ll get there, especially since it’s already a challenge for leading, established AI companies.
Martens said it could be hard to avoid U.S. Big Tech companies in that regard after all.
“Many AI startups, including European startups, collaborate with Big Tech companies in the [United States] … Microsoft, Google and so on, Meta … because those companies have the global business models that can generate the revenue that you need to repay those investment costs. In Europe, we don’t have those companies.”
What's Your Reaction?