Bull just landed the biggest contract in its short life to build a new EuroHPC supercomputer, and the number that jumped out at me wasn’t the €387.8 million price tag. It was the chip list. EuroHPC picked the French state-owned company, spun out of debt-laden Atos only months ago, for up to €404 million, to build LUMI-AI: the sixth machine under the bloc’s AI Factories program, installed alongside the existing LUMI system in Finland, online sometime in the second half of 2027. Storage comes from IBM, networking from Nokia, and the compute itself runs on AMD, not Nvidia.

That last detail matters more than the euro figure. Nearly every European sovereignty announcement I’ve covered this year runs into the same wall eventually: the funding gets approved, but the actual hardware defaults to Nvidia anyway, because switching away from CUDA is expensive even when the political will to diversify is real. Picking AMD here is one of the only visible instances this year of European public money actually diversifying its compute vendor, rather than just diversifying its press releases about compute vendors.

I still can’t get past the arithmetic, though. €387.8 million split between EuroHPC and a six-country consortium (Finland, Czech Republic, Denmark, Estonia, Norway, Poland) is real money, but it’s a rounding error against the $10.2 billion Alibaba raised in a single Hong Kong afternoon last week, and even that only buys the company about one quarter of its own compute burn. LUMI-AI is the sixth machine under a program that has already built 19 AI factory centers and 12 supercomputers off an €8.2 billion budget spread across seven years. Compare that to what a single hyperscaler spends before its next earnings call, and the gap stops looking like a funding delay and starts looking structural.

None of that makes LUMI-AI pointless. OVHcloud has already trained a foundation model on Jupiter, one of the network’s existing machines, and Domyn, the Italian startup behind the Commission-backed EUROPA consortium, is using EuroHPC capacity to chase a model targeting 400 billion parameters. Neither company has shown a working model yet, which is the part nobody wants to say out loud: Europe has mostly figured out the hardware procurement pipeline, well into the seven-year budget period that funds it, and is still waiting on someone to ship a frontier model on top of it that actually competes with what Mistral is doing, and even Mistral’s own answer to the sovereignty question turned out to be hosting a Chinese lab’s open weights unmodified rather than training something purely its own.

I’m not digging into whether Bull, barely free of Atos’s balance sheet, can actually deliver LUMI-AI on schedule. What I keep coming back to is that this is the first EuroHPC procurement in a while that reads like infrastructure instead of theater: a contract, a vendor, a chip choice that isn’t Nvidia by default, a delivery date. AXA’s Copilot rollout showed what sovereignty theater looks like when a real compliance framework gets quietly abandoned for convenience. LUMI-AI is the opposite failure mode: real money, real hardware, still not enough of either to close a gap measured in tens of billions.

Sources