US Edition
Your source for latest news
TechnologyAI infrastructure

Europe turns to AMD, not Nvidia, for its next AI supercomputer

The European Union's high-performance computing agency signed a €387.8 million contract to build LUMI-AI in Finland, a machine designed to cut the bloc's reliance on Nvidia chips and American cloud providers.

PT
By PressTemps Technology DeskPublished Today, 13:20 ET · 6 min read
Europe turns to AMD, not Nvidia, for its next AI supercomputer
The Renforsin Ranta site in Kajaani, Finland — a converted paper mill that already houses the LUMI supercomputer and will host its successor, LUMI-AI. (Credit: AMD Newsroom)
What to know
The EuroHPC Joint Undertaking signed a €387.8 million contract with French systems maker Bull to build LUMI-AI, a supercomputer powered by AMD chips instead of Nvidia's.
The system, to be installed at a former paper mill site in Kajaani, Finland, is expected to deliver ten times the AI capacity of Europe's current LUMI machine when it comes online in the second half of 2027.
Funding is split evenly between the EU's Digital Europe Programme and a six-nation consortium — Finland, Czechia, Denmark, Estonia, Norway and Poland.
The deal is the sixth in a planned network of 19 EU "AI Factories" meant to give European researchers and startups an alternative to Nvidia-built, US-hosted AI infrastructure.

The European Union's supercomputing authority has signed a €387.8 million contract to build a new AI-focused supercomputer in Finland using chips from AMD rather than market leader Nvidia, a deal that officials and industry analysts described as part of a broader push to reduce Europe's dependence on a single American chipmaker for artificial intelligence infrastructure.

The EuroHPC Joint Undertaking, the EU body responsible for the bloc's shared computing infrastructure, announced on August 31 that it had selected French systems integrator Bull to design, deliver and maintain the machine, known as LUMI-AI. It will be installed at CSC – IT Center for Science's data center in Kajaani, in a converted paper mill complex known as Renforsin Ranta, with deployment scheduled for the second half of 2027.

The numbers behind the deal

LUMI-AI will run on AMD's next-generation Instinct MI430X accelerators paired with sixth-generation AMD EPYC processors carrying 256 cores each, according to an announcement posted to AMD's newsroom. The system is expected to deliver roughly ten times the AI capacity and nearly double the general-purpose high-performance computing power of the existing LUMI supercomputer, which has occupied the same site since 2022. Bull's press materials describe the build as a BullSequana XH3500 system cooled with the company's direct liquid cooling technology, networked over its BXI interconnect and Nokia data-center switching, with storage handled by IBM Storage Scale.

The €387.8 million price tag — roughly $451 million at current exchange rates — covers acquisition, delivery, installation and multi-year maintenance. It is split evenly: half comes from EuroHPC JU through the EU's Digital Europe Programme, and half from the six-country LUMI AI Factory consortium of Finland, Czechia, Denmark, Estonia, Norway and Poland, which will host and operate the system once it is running.

How Europe got here

The original LUMI supercomputer, inaugurated in Kajaani in June 2022, was built around roughly 360,000 AMD EPYC processor cores and more than 10,000 AMD Radeon Instinct MI250X graphics accelerators, delivering about 550 petaflops of peak performance. It briefly ranked as the fastest supercomputer in Europe and third-fastest in the world, and its waste heat still supplies about a fifth of Kajaani's district heating. LUMI-AI is a separate, purpose-built addition rather than an upgrade of that machine, reflecting how quickly EU computing priorities have shifted in four years from general-purpose scientific computing toward AI training and inference capacity specifically.

LUMI-AI is the sixth contract EuroHPC JU has signed under its AI Factories initiative, a program that now aims to establish 19 such facilities across the bloc, each anchored by a EuroHPC supercomputer and intended to give researchers, public bodies and small and mid-sized companies access to large-scale AI computing without relying on commercial hyperscale clouds based outside the EU. That effort sits alongside a broader European push, spanning the EU Chips Act and the Digital Europe Programme, to build domestic capacity in semiconductors and computing infrastructure after years of near-total reliance on American and Asian suppliers.

The choice of AMD over Nvidia, which supplies the large majority of the world's AI training chips, was noted by hardware press as a deliberate diversification rather than a one-off procurement decision. Unlike Nvidia's proprietary NVLink, the BXI interconnect Bull is using in LUMI-AI is not locked to a single accelerator vendor, a design choice German hardware outlet Igor's Lab said gives EuroHPC more flexibility to switch chip suppliers in future rounds of procurement rather than being locked into one vendor's ecosystem for a decade or more.

Who stands to use it

Access to LUMI-AI will run through the LUMI AI Factory, which is intended to serve:

  • Academic and public research institutions across the six consortium countries and the wider EuroHPC membership
  • AI and deep-tech startups that lack the capital to build or rent equivalent computing capacity from US cloud providers
  • Established European industry seeking sovereign, EU-based infrastructure for AI training and inference workloads

CSC, which will host and operate the machine, said the system's capabilities will be paired with LUMI-IQ, a quantum computer also planned for the Kajaani site, so that researchers can combine classical AI computing with early-stage quantum resources on the same campus. That combination is intended to appeal in particular to smaller EU member states and to companies that could not otherwise justify building or leasing dedicated AI data-center capacity of their own, giving them a publicly funded alternative to buying compute time from Amazon, Microsoft or Google.

Anders Jensen, EuroHPC JU's executive director, framed the contract as a milestone in the joint undertaking's build-out, noting that LUMI-AI represents its sixth AI Factory procurement to date. Kimmo Koski, managing director of CSC, said LUMI-AI's capabilities, combined with the planned LUMI-IQ quantum system, would help European researchers "tackle unforeseen challenges." Bull chief executive Emmanuel Le Roux called the deal a continuation of the company's track record delivering AI Factories and exascale systems for European clients.

"Europe's next wave of AI innovation will depend on infrastructure that is open, efficient and trusted," said Thomas Zacharia, AMD's senior vice president of global public sector and strategic partnerships, in the company's announcement of the deal.

Industry coverage of the announcement, including from trade outlet HPCwire, largely framed the win as significant for AMD's push into European public-sector AI infrastructure, an arena Nvidia has dominated globally through its GPU and networking ecosystem. Tech Times, in its coverage of the deal, cast it explicitly as Europe "defying" that dominance by betting a nine-figure sum on a challenger chipmaker for one of the continent's marquee AI projects, rather than following the pattern set by most large commercial AI data centers built in the United States over the past several years.

What happens next

Bull and its partners now move into the engineering and construction phase, with the contract calling for the system to be operational at Kajaani in the second half of 2027, roughly a year and a half from now. The facility is designed to run entirely on renewable electricity, with waste heat piped into Kajaani's district heating network, mirroring the sustainability approach the original LUMI system pioneered. With LUMI-AI counted as the sixth of 19 planned AI Factories, EuroHPC JU still has 13 more such procurements to award, a pipeline of decisions in the coming months and years that will help determine how much of Europe's future AI computing capacity ends up running on chips other than Nvidia's.

More on this story

All Technology