Scaleway chooses AMD for a Swedish AI supercomputer outside Nvidia’s orbit

The story

French cloud provider Scaleway will deploy an artificial-intelligence supercomputer in Sweden using AMD Instinct MI355X accelerators, making it the company’s first supercomputer outside France. Announced at Scaleway’s ai-PULSE conference in Paris on October 13, the system is intended to give European startups, AI companies and enterprises access to AMD computing capacity for training and inference. Scaleway describes it as Europe’s first commercial AMD AI supercomputer, a company claim that has not yet been independently demonstrated.

The choice is consequential because Europe’s AI-compute expansion has largely followed a global market led by Nvidia hardware and its proprietary CUDA programming platform. Scaleway is pairing AMD’s MI355X with ROCm, AMD’s open-source software stack. Reuters noted that common AI tools can run through ROCm as well as on Nvidia systems, although moving a production workload between ecosystems is rarely automatic: optimized libraries, kernels, monitoring and staff expertise can still create substantial switching costs.

Scaleway is assembling a multi-vendor system rather than buying a single turnkey machine. Bull will integrate and deploy the cluster, connect racks and cooling, install it in the data center and handle hardware maintenance. Giga Computing will supply servers; HPE Networking will provide high-performance Ethernet; VAST Data will provide the storage and data platform; and Motul will supply liquid-cooling fluid. The architecture shows how an AI supercomputer depends on data movement, networking and heat removal as much as on the accelerator itself.

Giga Computing disclosed that the deployment will use its 4U direct-liquid-cooled servers, each carrying eight MI355X GPUs and two AMD EPYC 9005 processors. Each MI355X has 288 gigabytes of HBM3E high-bandwidth memory. That is useful detail at the node level, but it does not establish the system’s overall size: neither Scaleway nor its partners have said how many servers or GPUs will be installed. Data Center Dynamics also reported that expected compute capacity was not provided.

That omission limits what can responsibly be concluded. Scaleway has not published an opening date, procurement schedule, peak performance, availability target, power envelope or customer price. It has not named launch customers or disclosed reserved capacity. The word supercomputer therefore describes the planned cluster, not a benchmarked service already available to users. The claim that it will be Europe’s first commercial AMD AI supercomputer likewise depends on the company’s definition and on when the system enters service.

AMD accelerators are not new to European high-performance computing. Reuters pointed to Finland’s LUMI research supercomputer, which already uses AMD technology, while US cloud providers Oracle and Vultr rent AMD accelerators to customers in Europe. Scaleway’s distinction is the combination of a European commercial cloud operator, a dedicated Swedish deployment and MI355X capacity marketed broadly to European businesses. It is an expansion of choice, not the arrival of AMD on the continent.

Sweden adds a second strategic layer. Scaleway cited the country’s technology ecosystem and largely low-carbon electricity mix, and said the project builds on a partnership announced with telecom operator Tele2 in May. Locating compute near abundant lower-carbon power can reduce the emissions intensity of energy-hungry AI workloads, but the announcement provides no site-level energy data, power-usage effectiveness, water consumption or hourly carbon accounting. Those measurements will be needed to test the sustainability case.

The system also illustrates the boundaries of the term sovereign AI. Scaleway can offer operation in Europe, EU jurisdiction and a European commercial counterparty, while ROCm provides source-level openness that CUDA does not. Yet the central accelerators and processors are designed by AMD, an American company, and the supply chain spans multiple international vendors. Sovereignty here means more control over hosting, contracts and software choice; it does not mean a fully European hardware stack.

INNOVOX analysis: the strongest case for the Swedish project is competitive pressure. A credible MI355X service could give buyers leverage on pricing, availability and software terms while encouraging developers to avoid hard-coding models to one accelerator ecosystem. But optionality exists only when models train and serve reliably at a competitive total cost. Scaleway will have to prove performance per euro, utilization, support quality and workload portability, not simply install different chips.

What to watch next is the conversion of an announcement into measurable capacity. The decisive disclosures will be the GPU count, commissioning date, access model, independently reproducible benchmarks and customer pricing. Power sourcing, cooling efficiency and water use should follow. The longer-term test is whether European developers actually port demanding workloads to ROCm and keep them there. If they do, the project could diversify Europe’s AI infrastructure; if not, it risks becoming a symbolic alternative in a market still organized around Nvidia’s software gravity.

INNOVOX analysis

The strategic value is not a claim of European technological independence; the accelerators still come from a US chipmaker. It is a more practical form of resilience: another accelerator architecture, another software stack and a European operator under EU jurisdiction. That can reduce single-vendor dependence only if customers can move real workloads without prohibitive software and retraining costs.

What to watch

Watch for the installed GPU count, service date, independently reproducible training and inference benchmarks, hourly pricing, customer commitments, power and water disclosures, and evidence that ROCm workloads can move between this system and other clouds without extensive rewriting.