Europe’s sovereign AI race just got a third contender, and it is the one that owns the data centres. OVHcloud, Europe’s largest independent cloud operator, expects to launch its first proprietary AI model in early 2027, CEO Octave Klaba told Reuters on Tuesday, October 20, after the company published its annual earnings report.
The announcement places a hosting company squarely in the lane of Mistral, the French model lab, and alongside the European Union’s planned AI gigafactories, the state-backed computing programme Klaba has spent months saying is built wrong. If the early-2027 timeline holds, the company that sells the shovels will be selling the model too.
A mixture-of-experts model trained on Jupiter
Klaba said the model will use a mixture-of-experts design. In a conventional large language model, the entire network activates for every request, which makes each answer expensive to produce. A mixture-of-experts architecture routes each task to specialised sub-models instead, so only the relevant experts fire. The result is a model that can feel very large in capability while costing far less to run at inference time. It is the same design philosophy behind efficient open models: the comparison game in 2026 has moved from raw parameter counts to performance per dollar of compute, as seen in recent DeepSeek coding benchmark results.
The company is not starting from scratch. Klaba told Reuters in June that OVH had already begun training its own AI models on Jupiter, Europe’s fastest supercomputer. Four months of training runway sits behind this announcement, which makes the early-2027 launch window look like a timetable rather than an aspiration.
There is also a software front. Klaba said OVHai Workspace, a productivity platform tuned on OVH’s internal data and running workplace applications such as email, would release its first pilot shortly, with the first AI model following next year. The workspace is the customer-facing proof of concept: a European-built alternative stack, from infrastructure up through the application layer, before the model itself ships.
110 megawatts of spare capacity: OVH’s real advantage
The most striking numbers Klaba shared were not about parameters. OVHcloud has 250 megawatts of capacity connected, with 140 MW in operation and 110 MW ready to take on new clients. His argument is blunt: OVH can onboard AI customers without spending on new construction, while rivals must invest in additional capacity to do the same.
This is the core of the original angle, and it is worth spelling out. Training a proprietary model is usually framed as a research story. For a cloud operator, it is a utilisation story. Every megawatt sitting idle is stranded capital; training an in-house model converts that spare capacity into a product that can be sold at a margin, under European data-sovereignty terms, on infrastructure the company already paid for. A model lab has to buy or rent compute to train and serve. A cloud operator that already has 110 MW of headroom starts the race from a different place.
That framing also explains why the sovereign-AI argument lands differently coming from Klaba than from a startup founder. He is not asking a government to buy him compute. He already has it.
The Mistral-sized hole in the room
The obvious comparison is Mistral. The French lab remains Europe’s best-known independent model builder, and this site covered its trillion-parameter Mistral Large 4 “Le Chonk” open model earlier this month. Mistral’s strength is the model itself and its developer ecosystem. Its weakness, relative to OVH, is that it does not own the underlying infrastructure at comparable scale.
Klaba is positioning OVHcloud as a potential European challenger to Mistral, and the contest is now legible as a split: Mistral owns the model layer and rents or partners for compute; OVHcloud owns the compute layer and is building the model on top of it. Whichever stack customers trust with sensitive workloads, under EU jurisdiction, gets the sovereign-AI contract.
That trust angle matters beyond Europe’s internal politics. AI sovereignty is also about who builds the tools governments and enterprises are willing to depend on, a point underscored by the US government’s purge of DeepSeek from federal agencies. When states start choosing AI on trust and jurisdiction, an EU-based operator with its own model has a card to play that a foreign model simply cannot.
Klaba vs the EU gigafactory plan
The third side of the triangle is the EU’s own programme. Europe’s planned AI gigafactories, the state-backed computing initiatives meant to give the continent its own training muscle, came in for familiar criticism from Klaba. He reiterated that their financing, split half from the EU and half from member states, naturally pushes projects toward national initiatives, producing a race of national data centre projects, none backed by a customer base to match.
The critique is sharper than a policy disagreement. Last month Klaba called national strategies for the gigafactories “doomed to failure” and said a proposed 200 million euro public commitment to buy AI services over five years was too small to underpin the required private investment. His logic: a data centre without committed demand is a building, and 200 million euros over five years is not enough committed demand to justify what the private sector would need to spend alongside it.
Put the three pieces together and Europe’s sovereign-AI contest looks like this. The EU wants to fund capacity from the top down. Mistral builds models from the research side up. OVHcloud sits in the middle with paying customers, connected power, and now a training programme. Klaba’s bet is that infrastructure plus utilisation beats policy plus ambition.
What to watch
Three things will decide whether this announcement becomes a real product. First, the OVHai Workspace pilot. If a workplace platform tuned on OVH’s own data lands with enterprise customers soon, it validates the whole stack story ahead of the model. A pilot that stalls or stays internal is a warning sign.
Second, the early-2027 date. Four months of training on Jupiter is a head start, but mixture-of-experts models at competitive scale need extended training, evaluation, and safety work before launch. Watch for interim updates: benchmark claims, a named model, or an API preview would signal the timeline is real. Silence through mid-2026 would suggest otherwise.
Third, pricing, when it comes. The whole argument for a cloud operator training its own model is margin and sovereignty, and both live or die on price per token against Mistral and the US frontier labs. OVH’s cost base is its own idle megawatts. If that translates into aggressive inference pricing on EU soil, the gigafactory debate gets a lot harder to have on the old terms.
Reporting by Leo Marchandon in Gdansk for Reuters underpins the announcement; an AI news digest (u-s-news.com) has summarised the same earnings-day disclosure. The numbers that matter are all on the record: 250 MW connected, 140 MW operating, 110 MW spare, training started in June, model expected early 2027. Everything else is execution.
