What was launched
OVHai Workspace is built from three named parts. Productivity bundles collaborative tools and AI assistants. Web AI covers creating and managing websites with AI. The Knowledge Platform handles enterprise search. Workspace was first teased in June 2026 and has now moved from preview into an active public alpha, as SMBtech reported. The 55-product wave spreads across bare metal, public cloud, inference, agent orchestration and application security. That is a big release, and most of it is plumbing rather than demos. Plumbing is what decides whether an AI pilot survives contact with a procurement department.Free alpha, 5 GB per user
The Productivity alpha costs nothing and gives each user 5 GB of storage hosted in Europe. Five gigabytes is not a file server. It is enough for documents, notes and a searchable knowledge base for a pilot team, and not enough to migrate a decade of shared drives. The free tier still matters. Most seat-based AI assistants in European offices are charged per user per month, so a 40-person pilot means a real invoice before anyone knows if the tool works. A free alpha pushes that decision back by months. OVHcloud has not published a paid Workspace tier, a seat price or a general availability date in this announcement, so the alpha is not a pricing signal for later.The agent numbers
OVHcloud says its AI Agents connect to roughly 480 enterprise connectors and MCP servers, with support for more than 29,000 tools. These are vendor figures and we have not tested them. Read them as an integration claim: an agent that cannot read your ticket system and your CRM is a chat window with extra steps. Model Context Protocol support is where that battle is being fought right now, and 480 connectors puts OVHcloud in the same conversation as the hyperscalers rather than outside it. Speech transcription comes through Gladia in more than 100 languages, which is a practical detail for anyone running meeting notes across French, German and Czech teams.Capacity in megawatts and GPUs
The capacity figure OVHcloud presented at its Paris summit is about 110 MW, enough for up to 60,000 Nvidia Blackwell B300 or 35,000 Nvidia Rubin R200 GPUs. Dividing that facility power by the accelerator counts gives roughly 1.8 kW per B300 and 3.1 kW per R200. That is our own rough calculation from the facility power number, not a chip's thermal figure, and it says nothing about how much of that capacity is already allocated. Capacity announcements are inventory, not availability.The self-hosting break-even
The alternative to a hosted API is open weights on rented European bare metal. The two are billed on different clocks: an API charges per token, a GPU rental charges per hour whether or not anyone sends a request. That makes the decision a utilisation question rather than a benchmark question. A model that idles most of the day will lose to per-token billing almost every time; a model that runs close to flat out for months can beat it, because the fixed hourly cost is spread over far more work. The break-even point therefore depends on your own request volume, your own sequence lengths and how tightly you can pack users onto one accelerator — numbers only you have. On our own AI Arena rig with an RTX 5060 Ti 16 GB, small models run fine for internal jobs, but anything at production volume needs a rented GPU or an API.The European AI Alliance
OVHcloud, Qwant, Shadow and the European Search Perspective consortium formed an alliance with one stated aim: European data stays hosted, processed and operated under European jurisdiction. Qwant brings a search index, Shadow brings cloud gaming and desktop infrastructure, and EUSP brings the search consortium angle. Search is the part with teeth. A sovereign assistant that quietly forwards every query to a US search API is not sovereign, whatever the marketing says.Compliance and the AI Act
In September 2026, OVHcloud's SNC Cloud Platform obtained SecNumCloud certification from France's ANSSI, and further sovereign expansions are planned for Germany, Italy and Poland. That certification is the kind of paperwork that wins public sector tenders. The regulatory calendar matters too. Article 50 transparency obligations under the EU AI Act became enforceable on 2 August 2026, which means machine-readable marking of AI-generated content and clear notices when users interact with an AI system. The Commission's Digital Omnibus proposal would move the Annex III high-risk deadline to 2 December 2027, while general-purpose AI enforcement and the GPAI Code of Practice stayed active. In practice, a company running a customer-facing assistant now needs a disclosure layer and a documented data flow. Knowing which jurisdiction the inference runs in is the first line of that document, and it is the argument OVHcloud is selling.What a team can test this week
Start with the free Productivity alpha and one narrow use case, such as internal document search on the Knowledge Platform, because 5 GB will not carry more. Keep the AI Endpoints catalogue in view, since the roughly 100 multimodal models added on 19 and 20 November sit behind a standard API and are easy to swap. Do not plan a migration on a 110 MW capacity figure. The alpha requires no payment and the per-user storage sits in OVHcloud data centres. No paid tier, seat price or general availability date was part of the announcement.What is SecNumCloud and why does it matter for AI?
SecNumCloud is a French certification issued by ANSSI for cloud providers that meet specific requirements on data protection and extra-EU legal exposure. For AI workloads it gives public sector buyers and regulated companies a document they can point to when justifying where model inputs and outputs are processed.
Is the OVHai Workspace alpha limited in time?
The announcement does not state an end date for the alpha, nor a migration path to a paid plan. Treat the free Productivity tier as a pilot environment and keep your data exportable.
Do I need OVHcloud infrastructure to use the AI Endpoints catalogue?
No. AI Endpoints is an API product, so you call hosted models without renting GPUs. Self-hosting open weights on OVHcloud bare metal is a separate option and only makes sense at sustained utilisation.