ATAILA Newsroom · Budapest · 2026-08-02
Four machines can run the big model. The hard part starts afterwards.
The Hungarian news site hvg.hu reports on work by researchers at EPFL in Lausanne: large open-weight language models running on a handful of ordinary machines, inside an organisation’s own local network. For us this is not a theoretical question — the entire ATAILA platform is built on exactly this claim. Confidential workflows do not need a hyperscale data centre; they need a few GPUs, and someone who operates the system end to end.
The article we are responding to
„Adatközpont? Ugyan már! Irodákba költözhetnek a szuperszámítógépek”
hvg.hu · 2026-07-28
What hvg.hu reported
The starting point is a real tension. Large AI data centres draw growing criticism for their energy and water consumption, while the obvious alternative — running the model on a single personal computer — demands serious hardware. According to the article, EPFL’s researchers show a third way: a software layer (Anyway Systems) that joins several ordinary machines into a single local cluster and distributes the model’s execution across them. The article’s numbers:
The researchers say the most important advantage is not speed — they themselves budget for some latency — but privacy and sovereignty. The question is not whether the model fits in the office. It is whether the data has to be handed over at all.
Where we agree, and the one point we sharpen
This is not a rebuttal. The situation the article describes is what we see in our own customer base. On three points we say the very same thing:
- Inference is the real workload. Most everyday AI use is querying, not training — and querying does not need hyperscale capacity.
- Open-weight models have crossed the “good enough” threshold. For the typical tasks of a 20–200-person organisation — reading documents, summarising, making them searchable, coding — the largest closed model is no longer required.
- In many regulated industries, data residency is not a preference but a condition. With legal files, patient data and financial records, convenience does not decide.
On one point, though, we would sharpen the picture — and it is the point that matters.
The half hour is true — of installation. Not of operations. A running model is not yet a platform: it is a running program. The difference shows up on day 40, when the model needs updating, someone sees a document without the rights to it, the disk fills up, or a new feature has to go live without the system stopping. The research did not take that part on — nor is it its job. Inside a company, that is exactly where the cost begins.
What ATAILA does with those four machines
The article's finding
No hyperscale data centre needed — four ordinary machines are enough.
Our answer
We agree, and we add: this is the first of five layers. A real private AI platform consists of five — inference, knowledge (RAG over the company’s documents), applications, delivery and operations — and most offers in this market stop at the first. GPU rental or a model runner, with the rest left to the buyer. ATAILA ships all five, in one contract with one responsible party — we do not sell one layer and hope somebody else solves the other four.
What counts as a private AI platform →The article's finding
The data never leaves the local network.
Our answer
That is our starting point too: data stays where it is created. In practice we ship it in two editions. ATAILA Cloud is multi-tenant and fully managed — it runs on European hardware, with no iron and no operations on your side. ATAILA Enterprise is a single-tenant licence model in your own data centre: the data never leaves your servers, and operations are run either remotely by us or by your own team, with training and support. This is a business decision, not an article of faith — for many 20–200-person organisations the managed European cloud is cheaper and safer than a server rack nobody patches. We do not push either way.
The two editions →The article's finding
It installs in half an hour.
Our answer
Yes — and the real work starts afterwards. For us the delivery layer means a new feature reaches production on a versioned, audited, human-approved path, through SANDBOX, DEV, UAT and PROD environments — not on whichever machine it happens to be running on. The operations layer means monitoring, backups, updates, security and capacity are somebody’s named responsibility. This is the layer most offers forget — and the reason we charge a fixed monthly fee instead of billing per query.
How Factory and the Release Manager work →The article's finding
It has been in testing for months at companies and in public administration.
Our answer
Months of testing is good news — but the most expensive stretch lies between the test and production. Most enterprise AI initiatives die not because the model is bad, but because nobody funds and nobody owns the gap between a working prototype and a running system. For us the production environment is the first step, not the last: the platform’s open*.hu reference products run in production today, and the newest of them went live in hours through reuse.
Live references →What we do not promise
We do not claim four machines are enough for everything. Distributed local inference has a price: some latency, bounded concurrency, and hardware requirements that grow when you work with large contexts, many simultaneous users or heavy retrieval workloads. If your scale genuinely calls for hyperscale capacity, we will say so — and we will not sell you four machines.
Nor are we a model maker. We do not train foundation models: we run and operate open-weight models — our own merit is everything around them.
And we do not claim every workflow is worth automating. If the data is not there, or the process does not pay off, we say so in the first conversation — not in the third month.
What we do commit to: once you have decided that a piece of data must not go into public AI, what you get from us is not an installation guide but a running, operated system.
Where to start
If you have that one confidential workflow you have long meant to automate — the one you would never paste into a public chat window — tell us about it. We will say honestly whether a few GPUs are enough for it or not. And if they are, we will take it all the way there.
Source: Adatközpont? Ugyan már! Irodákba költözhetnek a szuperszámítógépek — hvg.hu, 2026-07-28
← Back to the Newsroom Press inquiries: contact us