I am often asked the question: will OVHcloud back one of the AI Gigafactory projects and join one of the consortia now forming in France?
The only way for us to take part in the AI Gigafactory project is to build a new consortium, bringing together the major GPU users such as OVHcloud, OVHai, Shadow, Qwant, Gladia, Dragon LLM, and other start-ups and companies from our ecosystem.
Our AI trajectory is part of the Step Ahead strategic plan, which I presented to our customers at Summit 2025 and have been rolling out since I returned to the helm of OVHcloud 11 months ago. We are executing this strategy with its artificial intelligence component, which rests on five layers:
- GPUs in our datacentres;
- GPU orchestration;
- Tokens aaS;
- AI PaaS;
- AI SaaS.
Today we have 250 MW available in our datacentres, ready to be used immediately to deploy GPUs. That is enough to start fast and then scale up to execute the plan. We have already developed, and now operate, the systems that drive these GPU layers, for training as well as inference. On Tokens aaS, we continue to expand the catalogue, building on open-source models, and with Dragon LLM and Gladia we have started training our own models to offer the broadest and richest range in Europe.
AI has swept through the software development process. We need new tools, new ways of developing, new products. We are already building these blocks to offer our customers this new generation of AI PaaS. Finally, with OVHai Workspace, which we unveiled at VivaTech 2026, we are offering a new approach to productivity, digital presence and enterprise application development, where artificial intelligence is the main means of interaction.
On many points, EuroHPC’s AI Gigafactories call for proposals is compatible with our Step Ahead strategy. The GPU volumes required, the expected revenue, the technical stacks and the services to be offered are all on our strategic trajectory to 2030. The question we are asking ourselves is this: can the AI Gigafactory help us go faster and further, with more means to succeed at what we have already set in motion?
Three elements are not aligned with our strategy.
A purely national logic
We are a European company, present in France, Ireland, Spain, Portugal, Italy, Germany, Poland and the United Kingdom, and we operate datacentres in several of these countries. So we do not think nationally, we think Europe, because it is the only scale at which there is a critical mass of customers who will use AI.
Any national strategy is doomed to fail: there are simply not enough customers to make investments of this size pay off. In our view, national demand is too low and we need to operate at European scale.
A technical architecture from the 2000s
The project requires everything to happen in a single place: one address, one power feed, one building. Technically, this design is out of date. Today, risks and constraints are spread across several sites to gain redundancy and high availability and reduce every type of risk.
Even though all the GPUs do need to sit in the same place for pre-training, those constraints do not apply to post-training or inference. On the contrary, for inference it is preferable to have several autonomous, independent infrastructures. For us, this design point is a real mismatch with the reality of what is needed.
Furthermore, the social and environmental acceptability of giant datacentre projects in Europe has to be taken into account. There are real execution risks in a giant project, with imposed deadlines and penalties for delays.
A symbolic public procurement
The announcement mentioned 30% of CAPEX. That figure is still true, but the devil is in the detail.
The French State and the European Union commit to buying €200 million worth of AI over five years. And these €200 million must represent at most 30% of the CAPEX that the AI Gigafactory has to invest. In other words, the AI Gigafactory has to invest at least €600 million in GPUs alone.
The €200 million is paid out as revenue over five years, i.e. €40 million a year. If that revenue were to represent 30% of total revenue, it means the AI Gigafactory would make €120 million in revenue a year. That revenue has to be set against the €600 million of investment in GPUs alone, to which must be added the CAPEX for datacentres, network and software, and then all the OPEX, including teams and energy.
It goes without saying that the numbers do not add up, and we estimate that between €300 and €400 million a year is needed just to reach 0% margin, especially as the rapid obsolescence of GPUs also has to be factored in.
So the €40 million a year does not represent 30% of the amounts at stake, but only 10% to 15%. The remaining 85% to 90% of capacity has to be sold to other customers, knowing that the offer is not that attractive: a single site for all the GPUs, and that site only in France.
Everyone knows the State does not have to fund everything. It is up to companies to invest, to go and find customers, to carry the risk. That is not the point. The point is that four constraints have to be added together, one of them cutting across the others. A national project. A single site. A customer that accounts for only 10% to 15% of revenue. And the rapid obsolescence of GPUs.
This is precisely the knot we are working on today: is it possible to reconcile these four constraints of the AI Gigafactory project with our Step Ahead strategy? It is up to us to work out the right equation between ambition, use cases, customers, architecture and business model, and then decide whether the AI Gigafactory can be an accelerator for us. This is why we welcome this political initiative and Europe’s ambition to accelerate on AI.
In any case, Europe needs sovereign AI and we will be there to provide it, with the five AI layers we are building, whatever our decision on the AI Gigafactory.