PARIS, June 17 - OVHcloud, Europe’s largest cloud infrastructure provider, said it intends to train frontier AI models - the large-scale, ground-up systems that require extensive data and compute - a strategic shift the company says could make it a second major European developer in the large language model space after Mistral.
Speaking at the VivaTech conference, OVHcloud CEO Octave Klaba framed the decision as essential to the company’s long-term prospects. "It became quite clear to us that if we don’t master this technology, we can’t guarantee our future," he said.
Klaba pointed to recent reductions in the cost of developing such systems, attributing the change to advances in chips, training techniques and the use of synthetic data. He said a project that might once have required about 1 billion euros could now potentially be undertaken for roughly 150 million to 200 million euros.
Describing the sector as entering a "second wave," Klaba said new entrants are now able to build on the technical foundations established by early players such as OpenAI, Anthropic and Mistral. He also stressed that OVHcloud will not use customer data to train its models.
Rather than attempting a single, all-purpose system, OVHcloud plans to develop a family of models tailored to specific uses. "We can clearly see that the major players release multiple models, because each model is built for something specific," Klaba said. "There’s no one model that does all the magic alone."
The company pointed to DragonLLM, a recently acquired start-up, noting that pre-training has been completed on a model using Jupiter, which was described as Europe’s fastest supercomputer. At the same time, Klaba cautioned that OVHcloud is not yet prepared to make detailed claims about performance.
OVHcloud also said it intends to open-source its models once they achieve sufficient performance standards. "We’ll see when we’re good enough to open source them, but that is indeed the goal," Klaba said.
The announcement comes amid increased interest from governments and firms in alternatives to U.S. and Chinese AI systems, a concern that has been heightened by the recent abrupt switch-off of Anthropic’s top-tier models. The company reiterated that mastering frontier AI is central to its future strategy.
Key points
- OVHcloud will train frontier AI models from scratch, aiming to be a European challenger to Mistral.
- CEO Octave Klaba cited lower development costs driven by chip advances, improved training methods and synthetic data, estimating potential project costs at 150 million to 200 million euros versus prior figures near 1 billion euros.
- The company plans a family of specialized models, will not use client data for training, and intends to open-source models once they reach sufficient performance.
Risks and uncertainties
- Performance uncertainty - OVHcloud has completed pre-training on a model but said it is not yet ready to make detailed performance claims; outcomes remain to be demonstrated. This affects AI and cloud markets.
- Market and operational urgency - governments and companies seeking alternatives to U.S. and Chinese systems have been prompted by the abrupt shutdown of Anthropic’s top-tier models, highlighting continuity and trust risks in the AI supply chain.
- Cost and execution risk - while development costs are reported to have fallen, building frontier models still requires substantial investment and execution capacity, affecting capital allocation in cloud and semiconductor sectors.
Note: Currency conversion referenced in remarks was $1 = 0.8627 euros.