The joint European undertaking for high-performance computing signed a contract on 31 August worth around €387.8 million to build an AI-focused supercomputer in Finland. The machine will target roughly ten times the AI capacity and twice the scientific computing throughput of the current facility. It will be equipped with AMD accelerators and processors.
We regularly document US and Chinese infrastructure investments. Here is the European commitment, and two details deserve attention.
The first detail: the order of magnitude
€388 million is a considerable sum for a European public research facility. Compared with the private investments we cover, it is something else.
We have reported on a $9.1 billion data centre lease signed by a single lab, and a $3.5 billion partnership for a company defending its position.
The gap is a factor of ten to twenty-five. That does not make the European investment pointless — a public supercomputer serves research, not a commercial service — but it does put the means in perspective. Europe is playing in a different category, and it is better to say that clearly than to celebrate a catch-up that is not happening.
The second detail: the supplier
This is probably the most significant piece of information. The machine will rely on AMD accelerators rather than Nvidia.
All week we have described Nvidia's dominant position and its customers' efforts to break free of it: OpenAI's in-house chip, Google's and Amazon's accelerators, Chinese silicon.
This choice fits the same trend. It shows that a credible alternative exists for this type of workload, which was not a given two years ago, since Nvidia's software ecosystem accounts for most of its advantage.
Let us take stock honestly. Europe has a regulatory framework, critical links in the semiconductor chain — notably the lithography equipment without which no advanced chip is made anywhere — and credible players such as Black Forest Labs in imaging. What it lacks is large-scale production capacity and private capital. A public supercomputer does not fill that second gap; it gives research something to work with. That is useful, and it is not the same thing.
What a public supercomputer is actually for
It is worth being precise about its use, because confusion is common.
This machine will not be used to compete with a commercial service. It will serve public laboratories, scientific projects, and companies that cannot afford to rent large-scale compute.
Concretely: training a specialised model on European medical data, running climate simulations, developing models in underserved European languages — a topic we covered on Wednesday regarding the languages that large models ignore.
This is the kind of use that finds no private funding because it has no immediate profitability, yet it produces lasting capabilities.
What to take away
This investment is coherent and modest. Coherent, because it funds what the private sector will not fund and diversifies suppliers. Modest, compared with the sums committed elsewhere.
The question that remains open is that of the model. Europe is funding research capacity while others are funding production capacity. These two strategies do not produce the same results.
There is, however, a path we see emerging that could suit Europe: the one described on Tuesday, where half of large US companies favour open models hosted in-house. In that world, what matters is no longer having the biggest model, but having the compute and the skills to use the ones that exist. That is precisely what this contract funds.