OVHcloud AI Summit 2026: one model shipped, four layers announced
I was at the OVHcloud AI Summit on 19 November 2026. The opening keynote laid out a European AI stack in five layers: GPUs, inference, agents, search and SaaS. One line came back again and again: an open, reversible and predictable European AI is an architecture, not a slogan.
Most of it lands in 2027. One piece shipped on the day: Olala v1.0 mini, the first language model OVHcloud trained from scratch. The weights are public under Apache 2.0, it runs on a CPU, and it scores 72.2 on MATH with only 1 billion active parameters.
1. GPUs: on demand, then serverless
- 110 MW of power across the datacenters, enough for up to 60,000 B300 or 35,000 H200 GPUs, as long as it stays economically viable.
- NV72 clusters: 72 GPUs that act as one system, for bigger models and more agents in parallel.
- Serverless GPU jobs: about ten lines of Python to run a massive batch, like analysing hundreds of thousands of phone calls, without managing a single GPU.
2. Olala, the first in-house model
AI Endpoints, the inference API, goes from about 20 models to over 100, with voice and image, and gains users, quotas and per-model usage. The headline was Olala ("oh là là" on stage), published on Hugging Face.
- Size: 6.8B parameters, about 1B active per token. A hybrid of Mamba-3 and attention layers, with 256 experts per layer (6 active).
- Data: 12.8T tokens in 37 languages. The data recipe is published.
- Training: on three European supercomputers, JUPITER, Jean Zay and MeluXina, for about 40 t CO₂e.
- Speed: about 50 tokens/s on a CPU and 160 tokens/s on an RTX A4500.
- Context: 32K tokens. 128K is announced.
OVHcloud benchmarked the base model against eleven others. A few columns, sorted by average over 13 benchmarks:
| Model (base) | Params | Avg | MATH | MMLU | HumanEval |
|---|---|---|---|---|---|
| Nemotron-3-Nano-30B | 30B, 3B active | 76.8 | 83.0 | 78.5 | 74.4 |
| Qwen3.5-35B-A3B | 35B, 3B active | 74.1 | 60.5 | 82.9 | 64.6 |
| Qwen3.5-9B | 9B dense | 68.9 | 55.6 | 72.4 | 61.6 |
| Qwen3.5-4B | 4B dense | 64.7 | 54.4 | 73.3 | 50.6 |
| Falcon-H1-3B | 3B dense | 59.5 | 49.6 | 68.3 | 57.3 |
| Olala v1.0 mini | 7B, 1B active | 57.1 | 72.2 | 57.5 | 42.1 |
| Luciole-23B | 23B dense | 55.3 | 31.8 | 64.3 | 37.2 |
| LFM2.5-8B-A1B | 8B, 1B active | 54.6 | 46.0 | 63.2 | 55.5 |
| SmolLM3-3B | 3B dense | 54.1 | 42.3 | 61.4 | 35.4 |
| Qwen3.5-2B | 2B dense | 48.8 | 35.8 | 54.5 | 32.9 |
| Qwen3.5-0.8B | 0.8B dense | 38.0 | 24.6 | 48.6 | 20.7 |
| OLMoE-1B-7B | 7B, 1B active | 38.0 | 17.1 | 55.1 | 11.0 |
In its class (about 1B active), Olala comes first. It beats LFM2.5-8B-A1B, trained on three times more tokens, and Luciole-23B, the other French model, with a third of the parameters. It is second on MATH, but weak on knowledge, code and long context. Qwen3.5-4B stays ahead overall. Note that OVHcloud ran the numbers itself, and Mistral, Gemma, Llama and Phi are not in the table.
It already runs in production: Qwant uses it for quick answers, 20% faster and three times cheaper than the model it replaced, per ITPro. Olala 2.0, early 2027, adds DevOps and cybersecurity.
3. AI Agents: sandboxes with access to nothing
AI Agents runs agents serverless, sized for millions. An event (a webhook, a Slack or Telegram message, a GitHub issue, a cron) starts a sandbox in about 2 ms. The agent works, then the sandbox is gone. Nothing runs between two requests.
The security model is the part I liked. By default the sandbox has no Linux machine, no SSH, no network and no disk. It gets three capabilities only: an outbound HTTP client (for AI Endpoints or an MCP server), a SQLite database per agent, and a workspace file system. The database is encrypted and shipped to Object Storage on every write, so a crash loses nothing.
One question I still have: that HTTP client is the way out for an agent fooled by a poisoned message. I did not hear whether it can be limited to a list of hosts.
4. AI Search: three paragraphs out of millions
Due January 2027. Giving an agent all your data burns context and gets it lost, so AI Search fetches, parses and indexes your sources, then serves "the three right paragraphs" over an API or MCP. It mixes semantic and exact-word search, targets tens of millions of documents per collection, and indexed data will not train models. Reranking and per-token multi-vector search come later.
5. SaaS: Website Builder and AI Workspace
- Website Builder (beta, available now): describe your project, get a live site. A side chat edits it. Domain, mail, backups and security updates are included.
- AI Workspace: mail, calendar, drive, documents and meetings, with AI in each. Thread summaries, common slots, folder clean-up, live meeting notes. The assistant runs on AI Agents and indexes everything with AI Search.
For 2027: end-to-end encryption with keys OVHcloud cannot access, and third-party apps (CRM, HR, accounting).
6. Cloud, private or on-prem
The whole stack comes three ways: public cloud (pay-as-you-go), private tenant (fully isolated, for banks, hospitals or local government) and on-prem racks, GPUs and AI Workspace included, air-gapped. Behind it, three labs: LLM (Olala), voice (from the Gladia acquisition) and a knowledge platform. Side note: ovhcloud.com now targets developers and agents, and ovh.com comes back for businesses.
My takeaways
- Usable today: Olala v1.0 mini and Website Builder. The published Olala is a base model with no safety training, so fine-tune it before any user sees it.
- Coming in 2027: AI Search in January, Olala 2.0 early in the year, Workspace encryption and third-party apps later.
- Worth watching: the agent sandbox design, and whether end-to-end encryption ships before people put sensitive mail in Workspace.