OVHcloud Just Shipped 55 AI Products: B300 GPUs, a CPU-Runnable LLM, and Europe's Big Sovereign AI Bet
Nothing here is sponsored — no affiliate links on this page. Every fact below comes from OVHcloud's official press release of November 19, 2026, and Reuters' same-day coverage. Capability and capacity figures are OVHcloud's own claims, not independently verified. We have not tested any of these products hands-on — this is researched analysis, not a review.
Europe's largest cloud provider just made its biggest AI move yet. At its AI Summit in Paris on November 19, OVHcloud unveiled 55 new cloud and AI products in one shot — from NVIDIA's newest GPUs to its own in-house language model family, a Python SDK for agent sandboxes, and a native agentic workplace. Reuters reported the launch the same morning. The headline number is the GPUs. The interesting part is the shape of the whole stack: models, inference, agents and infrastructure, all hosted in Europe, from a single provider.
🔍 How we put this together
We used OVHcloud's official press release of November 19, 2026 — "OVH takes the AI turn at full speed" — plus Reuters' same-day coverage of the AI Summit announcement. Where a figure is OVHcloud's own claim — the GPU capacity, the Olala roadmap, the alpha dates — we say so. Nothing here is independently verified yet, and we have not run any of these products ourselves.The GPUs: Blackwell B300 and Rubin R200 land in Europe
The compute news is straightforward. OVHcloud is adding NVIDIA RTX 6000 Pro, Blackwell B300 and Rubin R200 GPUs to its Bare Metal and Public Cloud offerings. The capacity figure the company quotes is striking: its data centres can supply up to 60,000 B300 GPUs or 35,000 R200 GPUs.
Read that number the way it is written. That is company-stated capacity, not installed inventory and not a count of GPUs already running customer workloads. It is a statement about the scale OVHcloud says it can reach, and it is the number that matters for anyone planning a large training or inference deployment: the ceiling is high, but the press release does not say how much of it is live today. If you are sizing a real workload, ask for the available-now figure before you plan around the headline.
The supporting network work matters as much as the silicon. The VPC stack gains a WAF, VPN, VPC peering, VPC endpoints, an application load balancer, a Global Load Balancer and private DNS, plus BGP Service and BYOIP so customers can announce and keep their own IP ranges. OVHcloud Connect now reaches 40 points of presence worldwide, and object storage gets faster with egress traffic still included. None of that is glamorous. All of it is what decides whether an AI workload in production is actually operable.
Olala: OVHcloud's own LLM family, and the small one is the story
The most developer-interesting announcement may be the least flashy. OVHcloud is launching Olala, an in-house family of large language models built on technology from its Dragon LLM acquisition. Five variants are planned: base, instruction, translation, summarization, and mini.
Olala mini is the one to watch. It is a 7B mixture-of-experts model with only 1B active parameters — and OVHcloud says it runs on a standard CPU. A capable small model that does not need a GPU changes the economics of the most common enterprise AI jobs: summarization, translation, classification, document triage. You can run those on ordinary compute, inside your own perimeter, without renting accelerator time for every request.
The roadmap is aggressive: Olala v1.1, with a 128k context window, is promised by the end of the year, and v2.0 — a 100 to 300 billion parameter family — in early 2027. Treat the dates as a roadmap, not a delivery. The v1.1 and v2.0 models do not exist yet, and the jump from a 7B MoE to a 100–300B family is a large one. The test that matters is not the announcement but whether v1.1 lands in December with real benchmarks and a usable API.
AI Fabric, AI Agents, AI Search: the developer stack — with one honest caveat
This is where the launch becomes a platform story rather than a hardware story:
- AI Fabric — built with Blade.sh, a Python SDK to launch inference-as-a-service and ephemeral agent sandboxes. The pitch is that you can spin up an isolated environment for an agent in Python, run it, and tear it down.
- AI Agents — hosts, scales, secures and monitors agents connected to enterprise tools through nearly 480 connectors and MCP servers. OVHcloud says a WebAssembly sandbox spins up in two milliseconds, and the agent only costs money when it is working.
- AI Search — explore private documents by combining semantic and lexical search across hundreds of millions of documents per collection.
- AI Endpoints — around a hundred new models, including voice and image models, plus governance tools. With Gladia, the offering extends to speech transcription and voice generation, in real time and on recordings, in more than 100 languages. AI Routers round out the inference layer.
Here is the caveat, and it matters: independent coverage notes that AI Agents and AI Fabric are listed as December alphas. Do not plan a production build on them today as if they were generally available. What OVHcloud announced is the shape of the stack and the intent to ship it; the December alpha is when developers will actually get their hands on it. The right move now is to read the docs as they land, test the sandbox model in the alpha, and keep your production agents where they already run.
Three OVHai research labs back the roadmap — dedicated respectively to LLMs, voice, and ontology-enriched document search — which at least signals this is a sustained research investment, not a one-day press release.
The agentic workplace, and the sovereignty pitch
OVHcloud also announced a native agentic workplace — email, documents and videoconferencing with built-in AI assistants — available in alpha free from today with 5GB per user, data hosted in Europe. Partnerships with Shadow, Qwant and EUSP, the European search index, are part of the same push.
This is the part of the launch that is aimed squarely at a European audience, and it is worth understanding why. The sovereignty argument runs through the whole announcement: services are available in Public Cloud, single-tenant Private Cloud and On-Prem, including with SecNumCloud qualification — the French state certification — and OVHcloud's certified cloud already hosts France's e-invoicing platform. The company describes itself as the only European cloud provider able to operate its services in Europe, North America and Asia. Its own framing, from CPTO Yaniv Fdida: most European AI projects never get beyond the prototype stage because moving into production means juggling several products, several providers, pricing grids and data transfers — and a single provider that masters the whole chain offers an integrated, open, reversible and predictable European alternative.
That is a genuine pitch with a genuine audience. European companies, public bodies and regulated industries have real reasons to want models, inference, agents and GPUs under European jurisdiction. The honest question is the same one that applies to every sovereign-cloud claim: how much of the stack is actually usable today, at what price, and with what latency — and how much is roadmap. The GPU and VPC pieces are concrete. The agent pieces are December alphas. The workplace is an alpha with 5GB per user. The sovereign story is real, but the product maturity is uneven across the 55 items.
Who this fits — and who should wait
This fits European teams and regulated organisations that need AI infrastructure under European hosting and want one provider for GPUs, inference and agents. It fits developers who want to try a CPU-runnable small model for document work — Olala mini is the cheapest experiment in the whole announcement. And it fits anyone already on OVHcloud who has been waiting for a managed inference and agent layer instead of stitching together third-party services.
Wait, or pilot carefully, if: you need AI Agents or AI Fabric in production now — they are December alphas, not live products. If you need a large-model API today, the Olala v2.0 family is an early-2027 promise, and the ~100 new AI Endpoints models need a close look at which models, what pricing and what rate limits actually apply. And if your workloads are already running well on your current GPU cloud, the 60,000-B300 capacity figure alone is not a reason to migrate — it is a reason to get a quote.
The honest limits
- Capacity is not inventory. Up to 60,000 B300 GPUs is the capacity OVHcloud says its data centres can supply. It is not a count of GPUs installed and available today.
- Two headline developer products are December alphas. AI Agents and AI Fabric are announced, not generally available. Judge them when the alpha opens.
- The Olala roadmap is unproven. A 7B MoE that runs on a CPU is useful now; v1.1 by year-end and a 100–300B v2.0 in early 2027 are promises. No public benchmarks for Olala were in the launch material.
- 55 products is a lot of surface. A launch this wide always mixes mature products with early ones. The VPC, storage and security items look like shipping products; the agent and workplace items are alphas. Read each one on its own merits.
- Everything here is the vendor's word. Pricing for the new AI services was not in the launch release, and nothing has been independently tested yet.
What to do this week
If you build with AI infrastructure in Europe, the move is a small, targeted pilot — not a migration:
1. Try Olala mini first. A 7B MoE with 1B active parameters that runs on a standard CPU is the lowest-cost experiment in the announcement. Test it on your real summarization or translation workload and compare quality against the model you use now.
2. Get a real GPU quote. If you have a training or inference workload, ask OVHcloud for available-now B300 capacity and pricing — not the 60,000 figure, the number you can actually book.
3. Sign up for the December alphas. AI Fabric and AI Agents are the pieces that decide whether this stack is a platform or a parts list. Get in the alpha queue now so you can test the Python SDK and the sandbox model when it opens.
4. If you are in Europe, try the agentic workplace alpha. Free, 5GB per user, data hosted in Europe — it costs nothing to see whether the email, docs and videoconferencing plus AI assistants fit a small team.
5. Check the SecNumCloud angle only if it applies to you. If you work in French public sector or regulated industry, the certified-cloud hosting of the e-invoicing platform is the signal that matters. For everyone else, it is context, not a reason to switch.
Sources: OVHcloud, "OVH takes the AI turn at full speed," press release, November 19, 2026; Reuters coverage of the OVHcloud AI Summit announcement, November 19, 2026 (08:43 UTC). Capacity, capability and roadmap details are OVHcloud's own claims; they have not been independently verified. Full version with FAQ on this page.