ATAILA Newsroom · Budapest · 2026-08-23

Europe is building AI gigafactories — but confidential workloads need a home today

The Hungarian news site infostart.hu describes exactly the tension we run into in customer conversations every week: Europe is finally putting serious money into its own AI infrastructure, yet in the public debate sovereignty keeps collapsing into hardware procurement. The article takes that conflation apart, and we agree with its diagnosis. We would add one thing — the thing that is our daily work: sovereignty is not decided by whose silicon spins in the accelerator, but by who operates the confidential workload, under which jurisdiction, and starting when.

What infostart.hu writes

In “Európa saját AI-t akar – rengeteg energiával és amerikai gyorsítóval” (Europe wants its own AI — with vast amounts of energy and American accelerators), the article reports that the European Commission opened the call for AI gigafactories on 30 July 2026: up to seven such facilities may be built, with up to €10 billion in EU and member-state support expected to mobilise at least €20 billion in private capital.

The lower tier of the network is meanwhile already being built. Hungary, the article notes, is one of seven member states — alongside Belgium, Cyprus, Ireland, Latvia, Malta and Slovakia — designated to host an AI Factory antenna; according to the announcement the article references, the Hungarian antenna is led by HUN-REN SZTAKI. The antenna logic: domestic researchers and businesses get access to high-performance resources while the supercomputer itself runs elsewhere.

The article registers two reservations. The first is hardware: in parallel with the tender, the Commission signed a letter of intent with three American semiconductor makers — AMD, Nvidia and Qualcomm — on access to the required processors. In the article’s wording, “the most important hardware does not necessarily come from Europe”.

The second is energy. Citing International Energy Agency data, the article puts the increment in Europe’s data-centre consumption at more than 45 TWh — a rise of roughly 70 percent between 2024 and 2030. On when the gigafactories will actually enter service, the article gives no date.

We agree with the diagnosis — and draw a different conclusion

We agree that Europe needs its own compute capacity, and that this is a strategic question. We agree with the energy warning too: the power an AI workload draws is a real physical constraint, not an accounting line. And we agree with the article’s paradox — we just draw a different conclusion from it.

A gigafactory is a supply-side answer. It is built for model training, research and large-scale experimentation, on a long horizon — and for that it is irreplaceable. But the 20-to-200-person law firm, accounting practice, clinic or engineering shop that today does not dare put its contracts, patient records or test data into an American chat interface is not looking for gigawatts. It is looking for a running, operated system that lives in the EU and can carry its own workflow — now, not in the next investment cycle.

The antenna model illustrates the gap well: it grants access to compute. Access, however, is not a platform. Between compute capacity and a confidential workflow sit identity management, networking, storage, logging, permissions, release management, backups — and the few people who pick up the phone at two in the morning. The sovereignty debate routinely skips this layer, yet it is precisely what the customer experiences as the service.

That is why we do not think sovereignty is a 2030 topic. Part of it is. The other part is decided today, with every workload someone places somewhere.

What ATAILA does

1. We treat sovereignty as an operations question, not a procurement line

ATAILA Cloud is an EU-resident cloud we maintain with the same daily responsibility as our own systems. We do not rent out machine time: we operate the running system end to end, and the data, the embeddings and the logs stay inside the platform. That is not the same claim as “stored in a European region” — who may touch the system, and through what process, is decided by the operator, not by the name of the region.

2. A GPU is not a platform

Our private AI platform guide distinguishes five layers: model serving, the knowledge layer (private knowledge base, RAG), applications, release management and operations. Most “sovereign AI” promises stop at the first: they hand over hardware or access and leave the remaining four layers to the customer. We ship all five, in one contract. The value is not in the individual layers but in the fact that they work together — and that someone operates them.

3. The right answer to the energy question is sober sizing

The article’s energy figures describe the hyperscale, training-centric world. The confidential workflows of a 20-to-200-person company — contract reading, document search, meeting-minute summaries, an internal assistant — can be served with open-weight models on a modest GPU footprint. The goal is not for every company to have its own gigafactory; it is for nobody to pay hyperscale prices for a well-scoped task. With us that is a fixed monthly fee, with no token-based surprises.

4. The proof is not a deck — it is a system in production

The platform proves itself on our own products: the open*.hu family — OpenMath and OpenChef today, more coming — was built by ATAILA Coder on private AI and runs in production on the platform; the full picture is on the references page. These are not gigawatt stories. They are exactly the stories a data-sensitive mid-sized company is looking for.

Our limits

We do not claim ATAILA replaces the European gigafactories. We do not train foundation models and do not want to: we operate open-weight models on our own infrastructure. If someone needs large-scale training or tens of thousands of euros of inference a month, we say honestly that they are not our customer — and there the European capacity build-out genuinely is the right answer.

Nor do we claim that European silicon works in our machines. It does not: our servers carry American-designed GPUs too, just like the European capacity the article describes. Europe’s independence in chip supply is a decades-long industrial question, and no platform provider will solve it.

If this sounds familiar

If you have a recurring, expensive workflow you would gladly hand to AI but cannot put into a public cloud because of the data, that is worth a conversation. We do not start with a deck; we start with an honest discussion of whether the task is worth doing with private AI at all — and if it is, what it looks like in production. The conversation is confidential and carries no obligation.

Get started →