Rare, expensive and essential, graphics cards, the cornerstones of the development of artificial intelligence (AI), are the subject of severe competition between European cloud operators, who rely on significant computing power to rent to their customers, in a context of supply tension.
Eight days apart at the end of January, two French companies, OVHcloud and Scaleway (a subsidiary of the Iliad group, owned by businessman Xavier Niel), announced that they were making new graphics processors (GPUs in English) available. calculation units much faster and more powerful than a classic microprocessor) from the American brand Nvidia.
Do you know the DATAROOM of FRENCHWEB.FR our database of French startups and innovative companies: key information, funds raised, turnover, organization chart, areas of development. Access the information we have collected on more than 1000 companies
Once acquired by these companies, these GPUs can be rented remotely by customers wishing to develop AI models, with hourly or even minute rates.
The principle is widely used in Europe: the German giant SAP, like the small British start-up Ori, are trying to exist against the American heavyweights Google, Amazon and Microsoft.
Faced with competition from these behemoths which, according to the French Competition Authority, represented “80% of the growth in spending on public cloud service infrastructure and applications in France” in 2021, Europeans are playing the sovereignty card.
If the French start-up Mistral AI, which wants to establish itself as a European alternative to the American leader OpenAI, has notably chosen to train its Mixtral model at Scaleway, “it is also a question of image”, notes Hanan Ouazan, partner at Artefact, a data and AI consulting company. “The sovereignty argument would fall apart if all the data ended up on American servers.”
Appearing in 1999 and first used in video games, graphics cards are now experiencing a second wind. “There is no alternative to doing generative AI other than using GPUs,” explains Hanan Ouazan.
But their cost is high: if the price of the star H100 model, marketed by Nvidia, is not communicated, it would be sold for around $40,000, according to the most common estimate on the market. However, a single GPU is far from being enough to develop sophisticated AI, and their maintenance is also a delicate matter.
– Opacity –
The cloud therefore becomes an interesting solution for AI developers, including the largest ones, such as Mistral AI.
However, cloud platforms are not immune to recurring supply tensions in the market for semiconductor materials, on which GPUs are based. Each operator therefore tries to do their part to benefit from deliveries of the latest graphics cards.
Essential, Nvidia represented 82% of GPUs delivered worldwide at the end of 2022, according to the Californian firm Jon Peddie Research. Contacted by AFP, the company did not confirm but displays several hundred “partners” in Europe on its site, a status which seems far from guaranteeing a privileged link.
“What makes the difference is technical expertise: faced with GPUs that are extremely rare, the worst thing that can happen to Nvidia is someone who buys GPUs and doesn’t know how to put them available to people who need it,” argues Damien Lucas, general manager of Scaleway.
David Chassan, director of strategy at Outscale, a Dassault subsidiary specializing in the cloud, praises a “very close relationship with Nvidia”, which allows the company to anticipate needs and availability over a year.
The platforms also maintain a certain opacity, refusing to reveal the number of GPUs, and H100 models, that they own. “In this environment, to exist, you have to count in thousands,” nevertheless concedes the general director of Scaleway, who specifies that his company made an investment of 100 million euros in AI in 2023, mainly directed towards the acquisition of GPUs.
Despite this accumulation, and the flexibility promised by the clouds, the scarcity of the resource is still felt by developers. “We have customers who are investing directly in GPUs, because they can’t have usage quotas reserved unless they commit to a cloud provider to use GPUs for a fairly high amount,” notes Hanan Ouazan.
The latest articles by THE EDITORIAL OF FRENCHWEB.FR (see everything)