ATAILA Newsroom · Budapest · 2026-08-18

96 megawatts of iron is not yet sovereignty. An operated platform is.

On 14 August the Hungarian business news site portfolio.hu published a rare kind of analysis about the AI campus planned next to Paks, the town that hosts Hungary’s nuclear power plant. Rare, because it neither celebrates nor scaremongers — it calculates: water demand, drought exposure, infrastructure financing. We take the same line of thought one layer up. If Hungary finally gets its own sovereign compute capacity, that is good news — but the iron on its own solves nothing. The real question is who will actually operate confidential company workloads on it, end to end.

What the portfolio.hu analysis calculates

The piece is an analysis by András Baráth, CEO of the water-industry company Vízipari Holding. Its starting point is a letter of intent signed in 2025 by Germany’s ParTec and Hungary’s 3D Lézertechnika: a multi-billion-euro AI and energy-infrastructure project next to Paks, centred on an AI data centre with a power draw of up to 96 MW, alongside a 530-hectare agri-photovoltaic park and battery storage. The developers also plan to reuse the waste heat in agriculture. The analysis’s strength is its water arithmetic: by the article’s numbers, a 100 MW data centre’s daily water demand hinges on the cooling technology.

The exposure behind those numbers: according to an OECD report cited by the article, surface waters supply 97 percent of the country’s freshwater, and more than 95 percent of that surface supply arrives from beyond its borders. And at the end of July 2026, with the Danube at a record low, three of the Paks nuclear plant’s four blocks were temporarily shut down.

The author’s conclusion is not that the data centre should not be built. It is that the investment’s value depends on whether a dedicated pipeline gets built for a single user, or a multi-purpose regional water system that also serves agriculture and the surrounding towns — one that remains standing even if the anchor tenant one day moves on. One of the section headings sums it up: whoever asks for water should show their water plan.

We agree with almost all of it — and the same holds for compute

Translated, the analysis’s central idea reads: infrastructure by itself is not a result. What matters is what system gets built around it, whom it serves, and who keeps it alive in the long run. That sentence carries over from water to compute word for word.

We are glad that sovereign AI capacity may get built in Hungary. Every piece of European iron strengthens the direction we work in — we wrote a separate response on why sovereignty is not a single-layer question. But a GPU hall is exactly as much of a solution for confidential company work as a dedicated pipeline is for Hungarian water management: it serves one purpose and leaves all the others open.

Look again at the article’s two water figures. The same 100 megawatts consumes 5,520 cubic metres of water a day in one case and 648 in the other — an eightfold difference at identical load. That is not a hardware question but an operations one. Hardware can be bought; the difference comes from who runs it and how. On our side the same relationship holds: between two equally capable infrastructures, what decides the outcome is whether someone operates the layers above them.

There is a second, less conspicuous parallel in the piece. The water system, the author argues, pays off when it is not built on a single tenant but backed by durable, contracted demand. Sovereign compute is in the same position: it becomes a result for the national economy when real, everyday company workloads run on it. Today those workloads sit overwhelmingly on American hyperscalers — not because the megawatt is cheaper there, but because there is a finished, operated platform on top of it.

Because a 20–200-person organisation working with contracts, patient data or financial data will not rent a rack in an AI campus. It does not need megawatts; it needs a working system: model serving, database, application, access control, logging, backups, updates — in one pair of hands, under one responsibility, with a written commitment. The iron is the bottom layer of that.

What ATAILA does — the layers above the iron

Capacity is not yet capability

ATAILA Cloud is not a plan and not a letter of intent — and it is not a cheap VPS either: a sovereign, production-grade cloud, the same stack ATAILA itself runs in production. Data stays in the EU; there is no American provider in the processing chain. It is not 96 megawatts and is not trying to be; it is built so that a company’s confidential workload runs today, runs tomorrow, and someone is accountable for it. We do not sell GPU rental and we do not hand over projects — we operate a running platform.

The gap between capacity and a working workflow

Between the megawatt and the application people use every day sits the whole delivery problem: environments, releases, permissions, audit, rollback. ATAILA Factory covers exactly that: the application is described once in a manifest, and the engine builds it, deploys it environment by environment and audits it.

The confidential workflow is the actual bottleneck

Your team uses AI anyway; the question is where the data goes while it does. A prompt is data processing too — even when you just ask something in a chat window. That is why every layer we build stands on the same principle: private AI for the work you cannot put into the cloud; responsibility and data stay in the same hands. The EU-hosted private assistant and Studio’s private AI coding stand on that same principle — data does not leave the platform. The fee is a fixed monthly amount, not a function of query volume, and the data and the code remain the customer’s throughout, in exportable form.

One row, not one column

Private AI, hosting and the operations team are in one pair of hands, under one contract, with a written SLA — not a dozen services to audit one by one. The platform comes in two editions: as a fully managed cloud, or as a licensed Enterprise version running in your own server room — the depth of sovereignty is a choice. The value is in the combination, not in the individual parts.

This is not theory: our open*.hu reference product family runs in production on this same platform — its newest member reached production in hours by reusing the finished modules.

Where we draw the line

We do not claim to have answers to the water questions around Paks. Those belong to water engineers, permitting authorities and municipalities, and the article is right to raise them. Nor do we claim to be a competitor to a 96-megawatt campus: we do not train foundation models — we operate open-weight models on our own hardware, as a managed service. If you want to train a foundation model, or you need capacity in the tens of megawatts, we are not the right partner — and we say so up front. And we certainly do not claim the iron is unnecessary. On the contrary: every capacity built here and elsewhere in Europe is good news for us too.

Our claim is narrower and verifiable. Compute capacity by itself is a cost centre; what turns it into a product is the platform running on top of it, operated end to end. And for that, a 20–200-person, data-sensitive Hungarian organisation does not have to wait for a national-scale build to finish: its confidential workflow can run on private AI, on EU infrastructure, today.

Start with one workflow

If your company has a workflow you cannot put into a public cloud — contracts, patient data, financial or technical customer data — we will gladly talk it through in confidence: which parts of it can run on private AI, and which cannot. We will also tell you if we are not the right fit for the task. We start with the project description, not a price list.

Get started →