OVHcloud launches an agentic AI suite and its own models
OVHcloud unveiled a family of LLMs, an “AI OS,” and several new AI services. The group aims to sell its own models, not just host those of others.
On November 19 in Paris, OVHcloud presented a roadmap that shifts it from a hosting provider to a publisher of AI models and services. The Roubaix-based group is launching a family of LLMs called Olala, a workspace designed for agentic AI, and a series of products across its entire tech stack. The message is clear: infrastructure alone is no longer enough; capturing software value is also essential. Without Mistral AI, with whom discussions did not succeed, and with a committed in-house strategy.
Olala, the bet on a model trained on internal use cases
The centerpiece is called Olala, a new family of models developed after acquiring Dragon LLM. The first version comes in five variants: base, instruction, translation, summary, and mini. Its most compact version features 7 billion MoE parameters, with 1 billion activated at inference, and can run on a simple CPU. OVHcloud initially uses it as a showcase of expertise, but the roadmap goes further with a version 1.1 expected by the end of the year, a context window extended to 128,000 tokens, then a version 2.0 with 100 to 300 billion parameters in early 2027.
Octave Klaba aims even higher. The CEO mentioned a model of 3 trillion parameters eventually, explaining that surpassing the trillion threshold is necessary to go far in cybersecurity and 24/7 operations. To train this generation of models, OVHcloud relies on 27 years of internal DevOps logs and reinforcement learning, meaning training where the model is rewarded when it successfully completes a defined task. Customer data, however, is not used.
An “AI OS” for offices that no longer want to choose between email and agent
The group does not stop at producing models. It also unveils OVHai Workspace, a collaborative suite natively designed for AI, which its founder prefers to call an “AI OS.” The tool combines messaging, calendar, storage, text editing, and videoconferencing in an interface designed in AI-first mode. The idea is to operate the suite in “prompt first” mode, with an experience close to Claude or ChatGPT, but in a Europe-hosted environment.
The alpha version is available for free now, with 5 GB of storage per user. OVHcloud is pushing here a different economic model than classic office suites. Octave Klaba sums up the logic in one sentence: “Today, with Office 365, you pay 10 or 15 euros per month for email, Drive, videoconferencing. And you wonder what you bought: software. I don’t believe that’s the value of tomorrow. Tomorrow, you will pay for an agent, the AI, and other services will come around it.”
Why OVHcloud insists so much on sovereignty
The issue goes beyond the product range. OVHcloud wants to become, within five years, a sovereign reflex for European companies, including AI. This means no longer just hosting others’ models, but offering its own and controlling more technical components. The group already relies on several acquisitions, including Gladia for voice and Seald for encryption, and plans to accelerate further. Octave Klaba speaks of more than six acquisitions per year today and a goal of twelve.
The strategy is also reflected in the organization: three OVHai labs are announced, one for LLMs, another for voice, and a third for document research. The latter feeds the knowledge platform, capable of storing, indexing, and searching text, spreadsheets, PDFs, audio, video, and images. In this context, the word “sovereignty” is not just for sounding serious in meetings. It mainly refers to control over data, models, and critical services.
A stack of AI services designed for businesses, not for demonstration
OVHcloud also detailed a series of complementary products. AI Fabric is intended to enable inference as a service with automatic GPU selection and temporary sandboxes for agents. AI Endpoints will integrate third-party models via pass-through, while AI Routers will automatically choose the model suited to the request. AI Agents targets large-scale deployment of natural language AI agents, with tools like MCP and scheduled tasks. AI Search provides a managed RAG, that is, an engine combining generation and document search to respond with more context.
The rest of the offering follows the same logic of progressive normalization of AI in everyday use. Website Builder is designed to create a site in natural language, including hosting and domain name. Receptionist promises an AI telephone switchboard available day and night. Gladia Endpoints adds real-time transcription, with voice synthesis announced next. Finally, OVH is preparing a redesign of its OVH.com site, while OVHcloud.com will be refocused on developers and the new agentic experience will primarily target businesses.
Key points
- Olala starts with five versions, including mini, instruction, and translation.
- The mini version activates 1 billion parameters out of 7 billion.
- OVHcloud announces a context window of 128,000 tokens.
- The group targets 2 billion euros in revenue.
- Octave Klaba mentions more than six acquisitions per year.
- OVHai Workspace is available in alpha for free now.
By the numbers
- November 19 — presentation date in Paris, according to OVHcloud.
- 27 years — internal DevOps logs used for training.
- 250 MW — electrical power available, according to Octave Klaba.
- 140 MW — power already used by OVHcloud.
- 5 GB — storage offered per user in the OVHai Workspace alpha.