Scaleway adds AMD MI355X and NVIDIA B300 GPU zones in Stockholm and Copenhagen

What Scaleway announced

At its ai-PULSE 2026 event on 13 October, French cloud provider Scaleway unveiled a set of GPU infrastructure expansions, according to a dev.to report on the vendor's announcement. Stockholm will host a new AMD Instinct MI355X cluster, which Scaleway presents as among the first of its kind from a European cloud provider, while Copenhagen gets a cluster of several thousand NVIDIA B300 GPUs. Both join a GPU Zone already operating in Paris.

Scaleway frames a GPU Zone as a section of a cloud region purpose-sized for model training and inference, which distinguishes it from general-purpose compute. Across the Paris, Copenhagen and Stockholm zones, the devto report says the vendor expects eventual capacity in the tens of thousands of GPUs, tens of exaflops of compute and tens of megawatts of power draw. Neither new cluster comes with a published delivery date or price.

Two accelerator families under one roof

Putting AMD and NVIDIA silicon in the same catalogue is the commercially interesting part. Workloads differ, and some inference jobs simply run cheapest on one family for a given model shape; a provider offering both lets buyers price them against each other. Scaleway itself presents the wider architecture mix as a way to keep customer choice open.

The dev.to analysis adds an operational caveat. A second accelerator family imports a second software stack, and the cost of maintaining images, drivers and build pipelines for both falls on the customer. Migration plans written for one family typically need a second pass before they work on the other, and that friction can decide whether teams ever reach the cheaper option.

Milan reaches three zones, more planned for 2027

Away from the GPUs, the announcement adds two availability zones in Milan, bringing the first Italian region to three. That threshold matters for production: an application spread across three zones can absorb the loss of one. A further six availability zones are pencilled in for 2027 across Stockholm and Frankfurt.

The Italian expansion runs through Qarnot Computing, a high-performance computing and waste-heat specialist Scaleway acquired earlier in 2026. In Brescia, Qarnot has teamed with utility A2A to feed heat from its computing infrastructure into the city's district heating network, an approach that only works where a heat customer sits within pipe distance.

A European inference chip, without specifics

Scaleway also named a silicon partnership with VSORA and ZML to build a European inference offering around the Jotunn8 processor. The announcement includes no specifications, timeline or performance figures, so the practical reading is intent rather than product. The European AI Gigafactory project, which Scaleway and iliad Group have joined, sits in the same early-stage category.

A wider model catalogue

On the software side, Scaleway's Generative APIs now list more than 40 models. Three were added with the announcement: LightOn OCR 3 from the French company LightOn for text recognition and extraction from complex documents, DeepSeek V4.1-Flash with native visual understanding, and GLM 5.3 from Z.ai, an open model aimed at software development and multi-step agent tasks. The broader catalogue spans Qwen, Gemma, DeepSeek, GLM, Mistral and Devstral families, with cached-token support and adjustable rate limits on the inference platform, plus a Dataiku partnership for orchestration.

Why it matters

European teams training or serving models have historically had limited alternatives to the large US hyperscalers, and Scaleway's dual-silicon GPU zones widen that field. The company points to selection under the European Commission's Cloud III framework, the Health Data Hub and an announced Airbus agreement as evidence of traction. The gap between announced and deliverable capacity is where buyers should focus: no dates or prices exist yet for the new clusters, and the Jotunn8 effort is far from shipping. The questions that can be answered today are which availability zone a workload lands in, which accelerator it needs and at what hourly rate, and what moving it out again would cost in data format and interface terms.