But Europeans are not sovereign on hardware because of Nvidia’s ‘international monopoly,’ France’s military AI chief Bertrand Rondepierre tells POLITICO.
PARIS — Even if France cannot master all the tech needed to develop artificial intelligence systems, it must create its own algorithms, France’s military AI chief told POLITICO.
« We have to do everything related to algorithms, training and models. It’s the very basics. Otherwise, there is no point in setting up an agency, » Bertrand Rondepierre, the director of the French ministerial agency for military AI, said in an interview.
The capability to design algorithms is « non-negotiable, » he added, referring to the instructions or rules that enable computers to recognize patterns, understand natural language, problem-solve and make decisions.
Last month, Armed Forces Minister Sébastien Lecornu said that France needs to become a global leader in military AI as of this year.
The French AI agency, known as AMIAD, was created in May last year. It has an annual budget of €300 million and employs more than 100 people. Engineers work on issues ranging from anti-drone warfare to the development of large language models to summarize hundreds of pages of documents and help with military planning.
Training AI models requires vast quantities of data and huge computing power. In September, AMIAD will receive a supercomputer, described by Lecornu as the most powerful classified one in Europe.
The supercomputer was purchased from U.S. company Hewlett-Packard and France’s Orange; the decision to buy from a foreign firm created quite a stir. However, Lecornu dismissed concerns, arguing that a competing offer from Atos wasn’t fully sovereign either because the French firm also buys its chips from the U.S.
Although the French AI agency can fairly easily develop algorithms on its own, there are bigger problems with Europe’s ability to make the hardware needed to actually train the models.
U.S. company Nvidia’s graphics processing units and its CUDA software — which are crucial to enable high computing performance — currently dominate the Western market. « On this layer, there’s a question of international monopoly, which makes it complicated, » Rondepierre said, referring specifically to the American firm.
His agency is currently looking at how it can « go down into the layers » to try and master more of the technological bricks: « Not the cloud part, because it’s not our role to store data, but for everything that’s related to software, in the layer above the hardware. »
AMIAD is also working on a partnership with French AI startup Mistral.
« I need access to world-class models. The whole logic of partnership isn’t just to give them money, it’s about building things together, » Rondepierre explained. Mistral gives AMIAD access to technology, and the agency gives Mistral access to the defense ministry, which is a « win-win situation, » he said.
That’s the sign of a wider trend where tech companies are getting cozier with defense ministries. Google this week dropped a promise not to deploy AI for weapons.
« We no longer have this separation between tech and defense, whereas before, putting [them] in the same sentence was a bit of a heresy, » said Rondepierre, who spent five years at Google’s AI research laboratory DeepMind before joining AMIAD.
That’s not only because of the war in Ukraine, he said, but in the wider context of increased competition between states. « It has become part of the engineers’ mindset, who are asking themselves more and more questions about sovereignty and the usefulness of their mission. »
However, Rondepierre conceded that the military isn’t leading on AI innovation.
« It’s important to have the humility to realize that it’s not usually the military that’s driving the civil sector, » he said. « The military has to keep up with civilian technology, and eventually surpass it. »
Emile Marzolf and Laurens Cerulus contributed to this report.
