Open Source · JUNE 18, 2026
OVHcloud Plans Frontier Model Family, Says €1B Project Now Costs €150M–€200M
At VivaTech 2026, CEO Octave Klaba said OVHcloud has completed pre-training on one model using Europe's Jupiter supercomputer, with a full open-source family to follow as Washington's Anthropic shutoff reframes Europe's sovereignty debate.
OVHcloud will train frontier AI models from scratch, CEO Octave Klaba told Reuters at VivaTech in Paris on June 17, positioning Europe's largest cloud provider as the continent's second serious LLM player after Mistral. Pre-training on one model is already complete, run on Jupiter, the exascale system at Germany's Jülich research center that's currently Europe's fastest supercomputer.
The economics are the news under the news. Klaba put the project's price tag at €150 million to €200 million today, against roughly €1 billion ($1.2 billion) "a generation ago," with the delta attributed to cheaper chips, improved training techniques, and synthetic data. That collapse in cost is what makes a sovereign-cloud incumbent's pivot into model training plausible at all.
"It became quite clear to us that if we don't master this technology, we can't guarantee our future," Klaba said. He declined to share performance figures, saying the company isn't yet ready to make detailed claims.
The plan is a family of open-source models, not a single flagship. "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." Client data, he added, won't be used for training, an unsubtle contrast with US incumbents pitching European regulated buyers.
The technical core comes from DragonLLM, the startup OVHcloud acquired earlier in 2026, whose CEO Olivier Debeugny is appearing at VivaTech on a panel titled "Building Europe's AI Champions." Klaba himself headlines "Powering Sovereign AI." He returned as CEO in late 2025 and has since steered the firm toward acquisitions, an in-house AI lab, and now a frontier program built on two decades of running its own servers and data centers.
The timing isn't accidental. Earlier this month, Washington forced Anthropic to pull its strongest models worldwide, and European officials protested being locked out of frontier systems on national-security grounds. On June 3, the European Commission rolled out a tech sovereignty package pairing Chips Act 2.0 with a Cloud and AI Development Act meant to give the bloc a single framework for assessing cloud and AI sovereignty. OVHcloud's pitch lands directly into that opening, with finance, government, and healthcare buyers as the obvious initial market. Independent European trackers including LemonLime have been mapping the sovereign-stack build-out in granular detail, and Klaba's announcement slots cleanly into the picture they've been describing.
Klaba framed the moment as a "second wave" of entrants building on groundwork from OpenAI, Anthropic, Mistral, and the commercial deployment platforms. What he hasn't disclosed yet is the shape of the thing itself: parameter counts, training-token budgets on Jupiter, licence terms for the eventual release. A sovereignty pitch without those numbers is a posture. With them, it's a product.