OVHcloud Plans to Train Frontier AI Models as Europe Races for AI Sovereignty

At VivaTech 2026 in Paris, OVHcloud CEO Octave Klaba made a declaration that would have been unthinkable a year earlier: Europe's largest cloud provider plans to train its own frontier AI models—the most advanced, large-scale systems built from vast data and computing power. The announcement positions the French firm as a direct challenger to Mistral AI and marks the most dramatic escalation yet in Europe's race for technological sovereignty .

"It became quite clear to us that if we don't master the technology, we will be dependent on others," Klaba told Reuters at the conference. He said OVHcloud would launch a family of models rather than a single system and pledged not to use client data for training .

The timing was no coincidence. Five days earlier, the Trump administration had demonstrated exactly what that dependency could cost.

How the Anthropic export ban catalyzed Europe's AI panic

On June 12, 2026, the U.S. Commerce Department issued an export control order barring Anthropic from distributing its latest AI models to any foreign national—including those physically inside the United States. The targeted models were Mythos 5, renowned for cybersecurity vulnerability detection, and Fable 5, its consumer counterpart. Anthropic, unable to verify nationality across its cloud infrastructure, disabled both models for all users worldwide within 90 minutes of receiving the order .

Commerce Secretary Howard Lutnick said the action was taken because officials feared the models could be deployed by military intelligence users in China, Russia, or other countries of concern . This was the second time in 2026 that the Trump administration had scrutinized Anthropic under national security export controls .

The order covered not just users outside the U.S., but foreign nationals inside the country—including Anthropic's own non-citizen employees. "In a statement, Anthropic said it had no alternative, as US authorities had ordered it to cut off access for 'any foreign national, whether inside or outside the United States,'" Le Monde reported, declaring that "the AI war has begun" .

Europe's reaction

The European Commission immediately condemned the move, warning that the restrictions "should not be discriminatory" and assessing the implications for EU technology access . The Register reported that the clampdown "sends EU sovereignty surge into overdrive," with the Commission citing it as proof that the EU must achieve "technological autonomy" .

VivaTech founder Maurice Lévy told Bloomberg that U.S. export controls had "not been well received" in Europe, describing them as a wake-up call that directly reinforced the urgency of OVHcloud's announcements at the same event .

The cost engine: record margins, not a price cut

The frontier AI plan is not funded by a breakthrough in training costs or a sudden drop in GPU prices. Instead, OVHcloud engineered the investment capacity through aggressive financial discipline:

Front-loaded capital expenditure. In H1 FY2026, OVHcloud deliberately pushed CapEx to 43% of revenue (€238 million) to secure memory, disk, and AI component supply and limit exposure to soaring component inflation. The CEO justified this as a strategic choice to lock in supply before prices rose further .

Record operational margins. The company achieved an adjusted EBITDA margin of 40.9% in H1 FY2026—its highest since the IPO—driven by operational leverage and AI integration into internal processes. This freed capital for ambitious AI investments while keeping the core business healthy .

Controlled customer price increases. Rather than fully passing on RAM costs projected to rise 250% to 300% by end of 2026 versus September 2025, OVHcloud capped average increases at 9–11% for cloud deployments between 2026 and 2028. Legacy deployments before 2025 saw more moderate 2–6% increases. The company absorbed part of the component inflation to stay competitive while still generating the cash needed for its AI push .

The model strategy: open, sovereign, and built for regulated industries

OVHcloud's AI model strategy rests on several distinct pillars:

Sovereign infrastructure from the ground up. All AI services run on OVHcloud's own European cloud infrastructure, governed entirely by GDPR. Customers can deploy and fine-tune models without data leaving European jurisdiction—a direct contrast to the extraterritorial reach of U.S. export controls .

An open, agentic platform. Alongside the frontier model announcement, OVH Labs unveiled OVHai Workspace in preview—an open, collaborative agentic AI platform. The company is also investing in open-source LLMs and demonstrated new inference-at-scale infrastructure built with SambaNova's RDU accelerators .

Targeting regulated industries. Through its March 2026 acquisition of Dragon LLM, OVHcloud gained specialized fine-tuning expertise for sectors like banking, legal, healthcare, and defense. These industries require data residency, audit trails, and regulatory compliance that U.S. hyperscalers often struggle to guarantee under evolving export control regimes .

From Dragon LLM to frontier models: the escalator strategy

On March 25, 2026, OVHcloud announced the acquisition of Dragon LLM (originally founded as Lingua Custodia), a Paris-based startup that had developed its own AI architecture trained on European supercomputers. Dragon LLM was a winner of the European Commission's Large AI Grand Challenge in 2024 and had built specialized generative AI models deployable on local infrastructure for regulated sectors .

Alongside the acquisition, OVHcloud launched its AI Lab to develop new LLM-based services for its 1.7 million customers. The Dragon LLM deal gave the company immediate credibility in fine-tuning and regulated-industry AI .

But the VivaTech announcement represents a step-change in ambition. Where Dragon LLM specialized in adapting existing models, OVHcloud is now committing to train frontier foundation models from scratch—moving from fine-tuning someone else's technology to building its own. As Klaba put it, the economics of developing these models "have improved significantly" .

What this means for European AI

The Anthropic export ban did not merely inconvenience European AI users—it demonstrated a structural vulnerability. If the U.S. government can order an American company to disable its most advanced models for all non-Americans with 90 minutes' notice, no European business or government can genuinely rely on U.S.-based frontier AI for critical operations.

OVHcloud is betting that this realization will drive demand for genuinely sovereign alternatives. By combining its own cloud infrastructure, Dragon LLM's fine-tuning expertise, a commitment to open-source development, and a new push into frontier model training, the company is positioning itself as Europe's full-stack answer to AI dependency.

The challenge is formidable: training frontier models requires enormous compute, talent, and sustained investment. But five days before VivaTech, Washington made OVHcloud's case more effectively than any pitch deck ever could.