OVHcloud unveiled 55 new products at its AI Summit, held on 19 November 2026 at the Carrousel du Louvre in Paris, 17 of them dedicated to artificial intelligence. The group is launching its own family of language models, Olala, an AI workspace available as a free trial and new NVIDIA computing capacity. The bet is clear: to sell the running of models (inference) within a European framework, rather than compete on giant-scale training.
An expanded catalogue, from networking to agents
Of the 55 announcements, 38 strengthen the cloud foundation and 17 target AI, from access to graphics processors through to applications, sums up the analysis by EuroLand Corporate published by Boursorama. The official OVHcloud blog details these launches by family: compute and dedicated servers, storage, security, networking, private cloud, tools, AI services and Olala models. Some of the products are still in alpha or beta, and some only have a “coming soon” page so far.
On the AI side, the AI Endpoints service, which gives access to models through a simple programming interface, is adding around a hundred more models, notably for voice and image, according to Solutions Magazine. A partnership with the company Gladia brings transcription and speech synthesis in more than a hundred languages. The new AI Agents service deploys software agents connected to business tools through nearly 480 connectors and MCP servers; according to Solutions Magazine, it is billed only when the agent is working. A router promises a single access point compatible with OpenAI’s interface, capable of automatically switching to another model in the event of failure. These two services are listed on OVHcloud Labs, the preview channel where the group tests its alphas and betas.
Olala, in-house models designed to run on little hardware
The most symbolic novelty is Olala 1.0, available in five variants: mini 7A1B, Translate, Summarise, Base and SFT. According to OVHcloud, the mini version has 7 billion parameters, of which around 1 billion are used for each request, allowing it to run on a graphics card with 8 GB of memory, or even on a standard processor. It is published under the Apache 2.0 licence on Hugging Face, with the documentation required by the European AI regulation. The models were trained on three European supercomputers, Jupiter, MeluXina and Jean Zay, and version 2.0 is due to exceed 100 billion parameters.
The company highlights a first customer: the search engine Qwant, where Olala produces the “flash” answer displayed at the top of the results. OVHcloud claims 3 million daily requests, served with three times fewer resources and responses 20% faster than with the previous model. These figures come from the vendor and have not been independently verified.
The same model powers OVHai Workspace, a collaborative workspace launched on the day of the summit as a free alpha, initially in France. Each user has 5 GB of storage; the data is hosted in Europe and is not used to train the models, the group promises.
NVIDIA GPUs to keep up with demand
OVHcloud is adding RTX 6000 Pro and Blackwell B300 cards to its offering and, later, Rubin. The B300s are arriving on the public cloud from November in more than eight regions, according to the blog, and a complete configuration of eight GPUs is offered as a dedicated server. Solutions Magazine mentions a capacity of up to 60,000 B300s or 35,000 Rubin R200 chips. The group’s announcement does not mention any pricing for these new products.
Yaniv Fdida, chief product and technology officer, justifies this integrated offering with an observation reported by Solutions Magazine: “In Europe, most AI projects never get beyond the prototype stage”, because going into production requires combining several products and suppliers, with their pricing grids and data transfers.
An inference strategy with revenues still some way off
For EuroLand Corporate, the architecture rests on three building blocks, language models, voice and document search, which give meaning to the acquisition of Dragon LLM in March. The software layer would be offered with free access and paid options, then monetised through infrastructure consumption. The research firm considers the outlook favourable, while warning that the additional revenues should not make a difference before 2028. The real test, according to the analyst, will be whether OVHcloud’s 1.7 million customers adopt these tools.
Two limitations are worth bearing in mind. First, the SecNumCloud qualification of the new dedicated platform is presented by OVHcloud’s blog as a process under way, not as a done deal. Second, compared with the American hyperscalers, a 7-billion-parameter model aims for efficiency and sovereignty more than raw performance: the announced variants (translation, summarisation) point to targeted tasks, with the step up in power deferred to version 2.0.
For French and European companies subject to data localisation constraints, the appeal is real: a single provider now covers compute, models, agents and the collaborative work tool, which simplifies contracts and compliance audits.
Frequently asked questions
Is Olala an open source model?
The mini 7A1B version is distributed under the Apache 2.0 licence on Hugging Face. OVHcloud describes the other variants as open-weight models, without specifying their licence in its announcement.
Is OVHai Workspace paid?
It has been offered as a free alpha since 19 November, in France first, with 5 GB of storage per user. The final commercial terms have not been announced.
When will these announcements have an impact on OVHcloud’s accounts?
According to EuroLand Corporate’s analysis relayed by Boursorama, not significantly before 2028. Development costs, excluding acquisitions, are said to have remained limited.
Sources
- OVHcloud AI Summit 2026: 55 launches to take AI into production — OVHcloud Blog, 2026-11-19
- OVHcloud, 55 nouveautés produits cloud et IA lancées ! — Solutions Magazine, 2026-11-20
- OVHcloud : AI Summit : le cloud autour de l'IA, la vision se précise — Boursorama (EuroLand Corporate analysis), 2026-11-20