French start-up also concerned about lack of diversity in language training models.
Europe lacks the data centres to train artificial intelligence models that match the current demand, an official from French AI company Mistral AI warned on Friday.
“We are reaching the capacity; we need to build datacentres and ensure there is enough electricity for the scale of AI development today. There is a huge amount of work to be done,” Audrey Herblin-Stoop, Head of Public Affairs at Mistral AI said at a conference organised by the European Court of Auditors (ECA).
Mistral AI, which was founded a year ago, used supercomputer facilities opened by the European Commission to train its models.
“The access to infrastructure is important, if you don’t have this, you cannot build large language models. We could train the first ones on [EU supercomputer] Leonardo. We were lucky enough not being blocked access to computing, but this will come for other companies. The volumes that are needed for AI are incredible,” she added.
The company announced a new large language model set to rival OpenAI’s ChatGPT, last February.
The model, called Mistral Large, is fluent in French, English, German, Spanish and Italian. However, its strategic partnership with Microsoft, that is looking to invest €15 million in Mistral AI, faces scrutiny by the European Commission.
Different scales
Mistral’s remarks come as the ECA warned in a report published last month that the European Commission needs to invest more in artificial intelligence if it wants to achieve its ambitions and be on a par with the US and China.
Reacting to the report, Killian Gross, responsible for the development and coordination of AI policy at the Commission, said that EU companies lack “the financial fire power” that US giants such as Microsoft have.
“We are in a different order of magnitude. We cannot catch up that quickly, but this is not a reason to be frustrated – it’s a reason to strengthen our efforts,” Gross said, speaking at the same conference.
MEP Dragoș Tudorache, the lawmaker who drafted the European Parliament’s position on the AI Act, warned that despite the potential for harmonisation that the incoming legal framework offers, approaches of member states vary a great deal on issues such as skills.
“Enabling the take up of AI is something that member states have to put in place. But national strategies see a very scattered landscape. And the contents of the strategies are not all the same,” Tudorache said at the conference.
Access to large data swaths of data needed
Access to data is another important factor in ensuring that AI companies in Europe can grow and continue to train their models, Mistral AI’s Herblin-Stoop said.
Mistral AI trains its models in Italian, French and Spanish but it needs access to large volumes of data in those languages to ensure its systems are relevant.
“We need to find a way to have access to data in a better way to have other cultures represented. The content will come from the US and China in the future. But as a global company we do care about building models that are strong in languages other than English,” she added.
The AI Act, stringent EU rules to regulate high-risk AI systems, was signed off this week and will appear in the Official Journal of the EU in 20 days.
The general-purpose AI rules will apply one year after entry into force, in May 2025, and the obligations for high-risk systems in three years.