October 17, 2023

AI startups need more data centres. France wants to build them

Private cloud providers in France are investing in compute resources for European AI startups. But will it be enough to compete against US tech giants?

Lexie Yu

AI startups may be gobbling up capital in 2023 but they are even more hungry for compute power and data centres. 

Increased competition has made these key elements even harder to come by. Even more so for European companies since the majority of this key infrastructure — like graphics processing units (GPUs), the chips needed by AI companies — is based in the US. 

That’s a big liability for Europe’s startups. In France, where a growing number of promising AI startups are beginning to emerge, French investors and policymakers are going above and beyond to secure this infrastructure, giving homegrown startups a chance at taking on their American rivals. 


Doing that will take far more than simply buying access to supercomputers. It will also involve encouraging European companies to start building the same kind of cloud capabilities as Amazon, Microsoft and Google, the world’s biggest tech companies. 

“It’s important for sovereignty reasons that GPUs are on French territory because otherwise access could be cut at any time,” says Stanislas Polu, cofounder of Paris-based AI startup Dust. “If these are products that are necessary to power society, it creates a dependency.”

French ambition 

The most significant contribution to this push so far has come from French billionaire Xavier Niel, who last month pledged another €200m investment in AI via his telecoms company iliad Group. 

This included the purchase of NVIDIA’s latest-generation supercomputer, specially designed to train large language models (LLMs) such as the one powering ChatGPT. Startups will be able to access the supercomputer through iliad’s sister company and cloud provider Scaleway. 

“To be relevant in the AI market, we need compute power. To have compute power, we need supercomputers,” said Niel at the time. 

Supercomputers are key to AI companies developing LLMs because they house GPUs, the “compute” used to train AI models. Supercomputers themselves are hosted in data centres tailored for AI applications. 

While companies may choose to purchase a supercomputer and build their own data centre, startups often opt for cloud-based access to GPUs that are located in cloud providers’ buildings. And most of that is from US cloud providers: 70% of the global cloud market is held by US players Amazon, Google and Microsoft.

Take Paris-based Mistral AI — founded by Meta and DeepMind alumni — which raised €105m in June to build models that could rival OpenAI. The company trains a large part of its model in the US.

Not only are companies largely dependent on the hardware provided by foreign players, but they also have to go with their rules — which can become more constraining as access to GPUs becomes increasingly competitive and waiting lists for key components grow. 

Sending training data to countries outside of Europe also poses the question of privacy in countries where laws provide a lesser degree of privacy than the GDPR. 


“To gain value from AI applications, you have to use data that sits at the heart of enterprises and has the most value,” says Aude Durand, deputy CEO of telecoms company iliad Holding. “That’s becoming a reason to worry.”

The French government steps up 

The French government has also pledged to provide more access to computational resources for AI startups. Last June, for example, it announced a €40m investment to add more GPUs to the Jean Zay supercomputer, which sits in the ministry of education.

In Europe, there are also initiatives to increase supercomputing power, with eight supercomputers currently located across the continent for businesses and academia. 

But these devices are not necessarily optimised for AI research. In addition, access to publicly procured supercomputers like Jean Zay are limited to certain allocated hours, which is unlikely to respond to the needs of businesses building very large models. Mistral’s first model, for example, required 200k of GPU hours.

“It’s not enough to have computing infrastructure in one place. Players in the ecosystem have to easily access that infrastructure,” says Guillaume Avrin, the French government’s national coordinator for AI. 

“This is where the private sector has a big role to play — in providing supercomputing-as-a-service for the training of AI models.” 

Creating an ecosystem

The key to encouraging more European private-sector investment in cloud capabilities is if European companies think they can find enough demand. Demand can be hard to predict when competing against giants like Amazon Web Services (AWS) and Google Cloud Platform (GCP).

Niel’s strategy to encourage more private-sector investment seems to be to generate both supply and demand. Alongside purchasing a supercomputer from NVIDIA, the billionaire is also the lead investor in AI startup Poolside, which is already working with Scaleway to secure access to the computational resources it needs to train its models. Poolside AI, which is building an LLM tool to write code, relocated to Paris from the US earlier this year.  

Although on a much smaller scale, the initiative resembles models like the Andromeda cluster in the US — a powerful supercomputer that is financed and run by investors Nat Friedman and Daniel Gross to be used by the startups they back. 

The question is whether Niel’s investments will be enough to start a brand-new ecosystem in France and across Europe that could meet the needs of current and future customers. “Right now, we are pretty confident, but I couldn’t say what the timeline is in terms of potential clients,” says iliad Holding's Durand.

Scaleway is already working with a number of startups in addition to Poolside. And while no partnership has been announced yet, Mistral seems like the logical next customer. 

“If there are cloud providers in Europe that had the intention of building data centres, we would be ready to secure partnerships from an early stage and even to share the investments,” says Mistral CEO Arthur Mensch.

France’s edge

It’s not just about building up cloud capacity. Historically, US cloud giants have had much better access to innovation, such as latest-generation GPUs. “Europe never has access to technology at the same time as the US,” says Yaniv Fdida, head of infrastructure at French cloud provider OVHcloud.

Fdida says that the gap with the US is slowly closing. Just a couple of days after Niel unveiled his investment in AI, OVHcloud announced that it was adding NVIDIA’s latest GPUs to its offering, specifically intended for customers developing machine-learning models and LLMs. 

Another major factor could play to France’s advantage. “We’re lucky in France that we have one of the most carbon-free energies in the world,” says the government’s national coordinator for AI, Guillaume Avrin.

France ranks third in Europe, behind Sweden and Finland, for carbon-free electricity. As AI models grow larger and require more power-hungry data centres, this is likely to become a strong differentiating factor for companies looking to secure their green credentials.

Positioning itself as a leading provider of AI compute power for European startups is an opportunity that France is keen to take, therefore — and with commitments from two private players, hopes are high that the country is in the right place at the right time.

“I think we are in the right timing,” says iliad’s Durand. “We absolutely have to continue in this rhythm and not lose momentum with the creation of startups and with bringing supercomputers out of the ground — to train the European ecosystem and remain in the race.”

Daphné Leprince-Ringuet

Daphné Leprince-Ringuet is a reporter for Sifted based in Paris and covering French tech. You can find her on X and LinkedIn