OVHcloud is considering participating in the European AI gigafactory program but sees major problems with the current setup. According to founder and CEO Octave Klaba, both the business model and the choice of a national, centrally built infrastructure are difficult to reconcile with reality.
The EuroHPC program is seeking consortia to build and operate up to seven AI gigafactories. These large computing facilities are intended to train, fine-tune, and run AI models.
At first glance, this aligns with OVHcloud’s own strategy. The company says it has 250 MW of data center capacity it can deploy directly for GPUs.
OVHcloud is also developing an AI portfolio that ranges from GPU infrastructure and GPU orchestration to Tokens as a Service, AI PaaS, and AI SaaS. According to Klaba, the scale of the planned gigafactories and the required technical infrastructure align with the roadmap OVHcloud has mapped out through 2030.
Nevertheless, he sees three major concerns.
National scale too small
The first problem is the national approach. OVHcloud operates in several European countries and believes that AI infrastructure should also be organized on a European scale. According to Klaba, individual national markets are too small to make investments of this magnitude profitable.
This also plays a role in the question of whether OVHcloud will join a French consortium. Klaba favors a new partnership among major GPU users, including OVHcloud, OVHai, Shadow, Qwant, Gladia, and Dragon LLM.
One large location
Klaba also has technical criticisms. According to him, the current setup relies on a single location, a single power supply, and a single building. He calls this an architecture from the 2000s. Modern infrastructure, by contrast, is distributed across multiple locations to mitigate risks and increase availability.
For pre-training large models, it may be necessary to group GPUs together. According to Klaba, this is not the case for post-training and inference. He sees particular advantages in having multiple independent infrastructures, especially for inference.
He adds that very large data center projects in Europe face social and environmental objections. A large, centralized project therefore also carries planning risks, especially when deadlines and penalties are involved.
Public procurement accounts for a small portion of revenue
Klaba raises the biggest question mark regarding the business model. The French government and the European Union are expected to purchase 200 million euros worth of AI services over the next five years. That amounts to 40 million euros in annual revenue.
At the same time, under the terms of the agreement, an AI Gigafactory must invest at least 600 million euros in GPUs. On top of that come investments in data centers, networks, and software, as well as operating costs for energy, personnel, and other expenses.
OVHcloud estimates that a gigafactory needs annual revenue of 300 to 400 million euros to break even. According to Klaba, the guaranteed public purchase therefore represents only 10 to 15 percent of the required revenue.
The operator must therefore sell 85 to 90 percent of its capacity elsewhere. According to Klaba, this is precisely where the combination of conditions poses a problem: the infrastructure is located in a single site, focuses largely on a national market, and consists of GPUs that become obsolete relatively quickly.
OVHcloud has not ruled out participation yet. The company is investigating whether the conditions align with its European strategy. Even without participating in an AI Gigafactory, OVHcloud says it will continue to invest in its own European AI infrastructure.