LUMI-AI and the Cloud and AI Development Act: what European sovereign compute can actually replace

Two European sovereignty stories landed within a week of each other and they point in opposite directions. On the build side, EuroHPC picked Bull to construct LUMI-AI, a EUR 387.8 million system next to the existing LUMI supercomputer in Finland, with AMD CPUs, IBM storage and Nokia networking, targeted to be operational in the second half of 2027. Czechia, Denmark, Estonia, Norway and Poland are among the co-funders, and EuroHPC now counts 19 AI factories. On the policy side, reported on September 5, defence ministries in several EU states are pushing back against parts of the planned Cloud and AI Development Act, the legislation meant to reduce dependence on US hyperscalers, citing operational risk and NATO interoperability.

I run workloads for Czech teams, so the question I care about is narrow: what could a team here move to sovereign compute, and what could it not.

What LUMI-AI is and is not

It is a shared HPC resource. Access will run through allocation calls, the way LUMI's does today: a project proposal, a review, and a time-bounded grant of node hours. It is training and inference capacity for research groups, startups and public bodies that can plan a workload in advance. It is not a cloud with an API key, a credit card and a five-minute signup. If your mental model is "AWS but European", adjust it now.

The date matters too. Second half of 2027 means anything you plan for it is a 2028 workload. Until then the sovereign option for a Czech team is existing AI-factory capacity and national centres, and those are also allocation-based.

Sovereign compute is real for the workload you can schedule and imaginary for the one you cannot.

What moves

• Fine-tuning and evaluation runs. Batch jobs with a defined start and end, a dataset that must stay in the EU, and no latency requirement. A mid-size fine-tune of an open-weight model fits allocation-based HPC perfectly.
• Large offline inference. Embedding a corpus, classifying a backlog, generating synthetic data. Anything you would otherwise run as a nightly batch.
• Research prototypes where the point is to prove something before you buy hardware or sign a cloud contract.

For all of these the serving stack is the one you would run on rented GPUs: a vLLM-style server per node as in H100 serving for a team, or a lighter Ollama team server for evaluation harnesses. The code does not care whose datacentre it is.

What does not move

• Interactive production serving. A chat product or an agent with a human waiting needs elastic capacity and an SLA. Allocation-based HPC provides neither.
• Anything with a hyperscaler dependency you have not untangled. Managed databases, identity, queues. Moving the model and leaving the rest in a US region buys nothing on sovereignty and costs latency.
• Defence-adjacent work, ironically. The ministries objecting to the Act are the ones whose workloads are meant to be most sovereign, and their argument is that the strictest rules would cut them off from allied systems.

The Act, as reported, classifies workloads by sensitivity and applies the strictest rules to a small share of services. That is the sensible design. It also means most commercial workloads will face no mandate at all, so "we have to move" is not a reason a Czech SaaS company can give its board.

How a Czech team should prepare

Write the allocation proposal for the batch workload you already have, because review cycles are long and the capacity is real. Keep production wherever it runs today, but make the model layer portable: same container, same weights, same serving config on the HPC node and on the cloud GPU. Track the Act as a procurement input, not a deadline. And measure what sovereignty actually costs you: the wall-clock delay from "I need GPUs" to "I have them" on each path.

The honest limitation

I have the announcement numbers and the press reporting, not the allocation terms, pricing or eligibility rules for LUMI-AI, and those decide everything for a small team. The defence objection is reported second-hand and the Act is a draft that can change. A 2027 delivery date for a large HPC procurement should also be read with the usual scepticism; slips are the norm in this class of project.