Scaleway to build AMD MI355X supercomputer in Sweden, Europe's first commercial AMD AI cluster

Scaleway says it will build an AI supercomputer in Sweden using AMD Instinct MI355X chips, and calls it Europe's first commercial AMD AI supercomputer. The French cloud provider, part of the iliad Group, announced the project on October 13, 2026 at its annual ai-PULSE event in Paris. It is the company's first supercomputer outside France. What the announcement leaves out is as telling as what it says: there is no GPU count, no power figure, no price and no date.

What was announced

Scaleway's press release says the system will give startups, AI companies and enterprises across Europe access to AMD Instinct GPU capacity at scale, running on AMD's ROCm open software ecosystem. The hardware partner list is specific. Bull will handle cluster integration and deployment, including racks, cooling integration, data center installation and ongoing hardware maintenance. Giga Computing supplies the servers, HPE Networking the Ethernet networking, VAST Data the data platform and Motul the liquid cooling fluid.

Reuters, which covered the announcement, describes the MI355X as AMD's competitor to Nvidia's Blackwell line. It adds that AMD chips already power LUMI, a research supercomputer in Finland, and that Oracle and Vultr rent AMD chips to European customers.

Why Sweden

The release says Sweden is a strategic market for Scaleway's growth in Northern Europe, and cites its AI and technology ecosystem and largely low-carbon electricity mix, framed as part of a "sustainable scale" approach. The project builds on Scaleway's partnership with Tele2, through which it entered the Swedish market in May, Reuters reports.

Tele2's B2B head Stefan Trampus said the project "sends a clear signal" about Sweden's attractiveness for AI infrastructure. That is a statement about positioning, not about megawatts. The release gives no site, grid connection or power figure, so we cannot say how large the facility is.

The software question is the real story

The most interesting line in the Reuters report is not about hardware. AMD's ROCm is open source and runs popular AI tools that also work on Nvidia chips, while Nvidia's CUDA is proprietary and runs only on Nvidia hardware. For a European buyer, the pitch is a second source of supply and a stack that is not locked to one vendor.

Whether that pitch holds depends on things the announcement does not address: how much of a customer's existing code runs unchanged, how the cluster performs against Nvidia-based capacity on real training and inference jobs, and what Scaleway charges per GPU hour. Without those, "choice" is a promise, not a product. AMD's Stephanie Dismore framed it that way, saying the project is "expanding the choice of AI infrastructure available to customers across Europe".

How to read "first"

"Europe's first commercial AMD AI supercomputer" is Scaleway's own description. Reuters reports it as Scaleway's claim and notes that others already rent AMD chips to European customers. The word "supercomputer" is also doing work. The release gives no performance figure, so there is nothing to compare against LUMI or against Nvidia-based European clusters. We would read "first" as "first branded as a supercomputer by a European cloud provider using this chip", which is narrower than it sounds.

What it means for the compute race

Europe's AI compute is thin compared with US hyperscalers, and sovereignty arguments often turn on who owns the hardware and where it sits. A French company placing AMD capacity in Sweden touches those arguments without settling them. We unpack the distinction between funded and announced compute in our sovereign AI explainer, and the grid side of the same problem in our reporting on grid interconnection. This announcement fits the pattern: a named chip, named partners and a named country, with the capacity numbers that would make it a real addition to Europe's compute still missing.

What a buyer should ask

A startup weighing the cluster has a short list of questions, and none is answered yet. How many MI355X GPUs, in what configuration, and on what interconnect? HPE supplies the Ethernet networking, and real training performance will depend on how the whole system is tuned. What does an hour of capacity cost, and for how long can it be reserved? And when does it open for customers?

Reuters notes that Bull, which works with Scaleway in the AION consortium, called the contract a positive step for its own strategy. That is a supplier's view of a sale, not evidence of demand. The release also describes the partner roles, but not how much capacity, if any, has been pre-sold. Customers should treat the announcement as a signal of intent and ask for dated commitments before planning around it.

The European compute context

Scaleway frames the Swedish site as part of a European build-out, and Tele2's executive said the project sends a clear signal about Sweden's appeal. What distinguishes this announcement is the chip choice, not the geography. A second supplier in the market gives European buyers leverage on price and delivery, which is worth something even before the cluster switches on, but only if the capacity is real and rentable.

Our take

Treat it as an intent to build, not capacity you can book. The case for it is real: a credible non-Nvidia option for European customers, powered by low-carbon electricity. The case against waiting is equally real: no number in the announcement lets a buyer plan anything. We would watch for four disclosures: GPU count, power draw in megawatts, a go-live date, and a published price per GPU hour. Once those exist, the AMD-in-Europe story becomes one of performance and price. Until they do, it is a press release with a good partner list.