OVHcloud bets on frontier AI as European alternative

What happened?

OVHcloud, one of Europe's leading cloud providers, has announced its entry into frontier artificial intelligence model development. The French company plans to train a family of models from scratch and release them under an open-source license once they reach certain performance thresholds, according to CEO Octave Klaba's statements to Reuters. The first model has already completed its pre-training on the Jupiter supercomputer, part of the EuroHPC network in Germany, considered the fastest in Europe and the first to reach exascale scale. This supercomputer, located at the Jülich Supercomputing Centre, has a computing capacity of 1 exaFLOP, placing it among the most powerful in the world, though still behind systems like Frontier (US) or the Chinese Sunway TaihuLight. OVHcloud has not disclosed detailed performance metrics of the pre-trained model, creating uncertainty about its actual quality compared to competitors like GPT-4o or Claude 3.5.

Why is it important?

This move positions OVHcloud as a direct competitor to Mistral AI, the Parisian startup that has so far been the European benchmark against US (OpenAI, Anthropic, Google) and Chinese labs. Mistral AI has raised over €1 billion and launched models like Mistral Large, competing in benchmarks with GPT-4. However, OVHcloud, with an annual revenue of €897 million in 2024 and over 1,600 employees, has a consolidated enterprise customer base, which could facilitate the adoption of its models. The bet gains particular relevance after the US government's export control directive that forced Anthropic to suspend access to its Fable 5 and Mythos 5 models for foreign citizens, even outside the country. According to the January 2025 executive order, the US Department of Commerce expanded restrictions on AI models considered dual-use, affecting non-US entities. This episode has intensified the debate on technological dependence and the need for sovereign alternatives in Europe, especially after the approval of the EU AI Act, which requires transparency and governance for high-risk models.

Klaba argues that the economics of training advanced models have changed dramatically: improvements in chips (such as NVIDIA's H100 GPUs, now renting for less than $2 per hour), training methods (like distillation and reinforcement learning), and the use of synthetic data have reduced the cost of a project from approximately $1.15 billion (the estimated cost of training GPT-4) to less than $230 million. However, analysts like Neil Shah, vice president at Counterpoint Research, warn that this figure only covers initial training. Recurring costs for fine-tuning, sovereign infrastructure, storage, security, distribution, and enterprise support can multiply the required investment. Shah notes that 'the model is seen as a depreciating asset if it is not consistently trained and kept fresh with the data.' To maintain a competitive model, OVHcloud might need to invest an additional $50 to $100 million annually in updates and maintenance.

Consequences and challenges

OVHcloud's success depends not only on technical capability but also on economic viability and political support. If the company achieves competitive models, it could incentivize European businesses and institutions to migrate workloads from US providers, reducing data leakage and strengthening the continent's digital autonomy. Conversely, if the models fail to meet expected performance, the market might distrust European alternatives, consolidating the dominance of current players. A precedent is the failure of the European 'GAIA-X' initiative, which aimed to create a sovereign cloud but failed to achieve expected adoption due to a lack of competitive services.

Moreover, the open-source business model means OVHcloud must find ways to monetize beyond model sales, likely through cloud services, consulting, or enterprise versions with additional guarantees. Competition with giants like Google and Anthropic, which offer proprietary models with integrated ecosystems (Google Workspace, Vertex AI), will be fierce. OVHcloud will also have to compete with other open-source initiatives like Meta's Llama 3, which already has an active community and optimized versions for different use cases. The company plans to release its first open-source model in the second quarter of 2026, according to sources close to the matter, though it could be delayed if performance thresholds are not met.

Political support is crucial: the European Commission has allocated €1.5 billion for AI infrastructure through the EuroHPC program, and the French government has announced an additional €500 million for sustainable data centers. However, bureaucracy and compliance requirements of the AI Act could slow development. On the other hand, dependence on US chips (NVIDIA) remains a weak point, though OVHcloud is exploring alternatives with European chips like those from SiPearl or Graphcore.

What readers should know

• Real costs: Initial training is just the first step. AI models require constant updates with fresh data to avoid depreciation, implying ongoing investments. OVHcloud will need to allocate at least 20% of its annual R&D budget (around €40 million) to model maintenance.
• Data sovereignty: OVHcloud's decision is framed in a context where continuity of access and data governance weigh as much as technical performance. The EU AI Act requires high-risk models to be auditable, which could be an advantage for OVHcloud over US competitors.
• European ecosystem: The initiative could catalyze a more diverse AI ecosystem, but it needs institutional support and business adoption to be sustainable. Companies like Airbus, Siemens, and Deutsche Telekom have already shown interest in sovereign alternatives.
• Pending benchmarks: OVHcloud has not yet published detailed performance metrics, so it is too early to assess the actual quality of its models. They are expected to compare with models like Llama 3 70B or Mistral Large in reasoning, coding, and language understanding tasks.
• Risk of chip dependence: Although OVHcloud uses EuroHPC supercomputers, they rely heavily on NVIDIA GPUs. The company is evaluating European chips (like Esperanto Technologies' RISC-V) to reduce that dependence, but no firm agreements are in place yet.

'Model is seen as a depreciating asset if it is not consistently trained and kept fresh with the data,' notes Neil Shah, reminding that the real challenge is not just creating a model but keeping it relevant. OVHcloud must demonstrate it can sustain a cycle of continuous improvement to compete in the long term.