OVHcloud will not join AION, nor the Æther consortium, the two French groups vying for the EU’s AI Gigafactory program.
The message from Octave Klaba, filtered through a blog post and then clarified with Reuters, leaves little room for doubt.
In a piece with the provocative title “AI Gigafactory: 404 Business Model Not Found (yet)?”, the head of the Roubaix-based group explains it with a series of arguments that target less the European ambition of the project than the way France is proceeding to make it a reality.
A French Positioning Seen as Out of Scale
The first sticking point lies in the geography of the market.
AION, led by Iliad and Scaleway with about thirty partners, and Æther, built by 2CRSi and its allies around two sites in Strasbourg, remain two projects designed primarily to meet French needs.
Yet OVHcloud already runs data centers in Ireland, Spain, Portugal, Italy, Germany, Poland, and the United Kingdom. From Roubaix, the demand of a single country simply does not carry enough weight to amortize an infrastructure of several hundred megawatts. “Any national strategy is doomed to fail,” says Octave Klaba.
The Second Argument Relates to the Wallet
Octave Klaba starts from the assumption that the public procurement available for a French gigafactory would hover around €40 million per year, i.e., €200 million over five years in total. A figure not to be confused with the government’s official commitment, which announced at the end of July €100 million of computing capacity purchases starting in 2027. The €200 million thus reflects the leader’s estimate, not an appropriation approved by the treasury.
Or, according to OVHcloud’s founder, the procurement needed for GPUs alone would require at least €600 million, not counting data centers, networks, software, energy, and staffing.
To reach financial equilibrium, a gigafactory would need to generate €300–€400 million in annual revenue. The public procurement mentioned covers only 10–15% of that amount, meaning most of the funding would have to come from private customers yet to be found — even as chips depreciate rapidly from one generation to the next.
For Octave Klaba, the risk is twofold: financial, if solvency demand does not materialize, and structural, if the infrastructure is built before its market is secured.
A Unique Risk-Taking
The third reservation concerns the architecture itself.
The head of OVHcloud considers the idea of a single giant site that concentrates all execution risks — land, construction timelines, and dependency on a single power supply — to be outdated.
Because while this configuration makes sense for the initial training of models that require highly interconnected GPUs, it makes far less sense for post-training and inference.
And these are the two uses that, he argues, benefit from being distributed across several autonomous infrastructures closer to users. A way to spread both the load and the risk in case of a failure or incident.
This skepticism also stems from an industrial reality: OVHcloud does not need to start from scratch.
The group already claims 250 MW of available capacity in its data centers capable of housing GPUs as well as the systems needed for training and inference.
Its strategy, dubbed “Step Ahead,” revolves around five stacked layers: the GPUs themselves, their orchestration, a “Tokens as a Service” layer, a Platform-as-a-Service dedicated to AI, then SaaS applications integrating artificial intelligence. Rather than funding a new heavy, uncertain brick, Octave Klaba prefers to consolidate an already-built edifice.
Infrastructure Already in Place
Pragmatic by nature, he prefers to build around recognized GPU users who are also his own companies. That is to say OVHcloud, OVH AI, Shadow, Qwant, Gladia, and Dragon LLM.
Each plays a role: OVHcloud for infrastructure and cloud, OVH AI for usage and computation, Shadow for GPU-intensive workloads, Qwant as a driving use case for search, Gladia for voice AI, Dragon LLM and Gladia again to form a portfolio of in-house models.
This vertical integration is meant to ensure a utilization rate well above that of an infrastructure that goes in search of its customers afterward.
Yet Octave Klaba does not close the door on the EuroHPC program. The real question, he writes, is whether a gigafactory would allow OVHcloud to go “faster and farther” along a trajectory already set in motion.
