ATAILA Newsroom · Budapest · 2026-08-28

Two-thirds are running AI projects. A fifth are building anything underneath them.

A small survey run by SUSE among Hungarian IT and business decision-makers at this year's Bitport CIO conference reports that 34.1% run an AI solution as a pilot and 34.1% already run one in production, while only 20.5% named preparing AI or ML infrastructure as a modernisation plan. We will take the honest position on the numbers first, and then argue that the gap they point at is real anyway.

First, what this survey is and is not

It is a show of hands at one conference, run by a vendor, published alongside that vendor's product. The percentages back-solve to a sample in the low forties. There is no published methodology and no margin of error, and the subgroups are small enough that a single respondent moves them by several points.

So we are not going to quote it as evidence of anything about the Hungarian market, and we would gently suggest nobody else does either. We are treating it as what it honestly is: a prompt for a conversation that is worth having. The reason we think the conversation is worth having is that a much larger dataset says the same thing.

The pattern that does hold up

The same article cites the Stanford HAI 2026 AI Index: 88% of organisations used AI in at least one business area in 2025, and only 3–10% reported full enterprise-scale rollout. That is a serious dataset, and the shape matches — near-universal adoption, single-digit consolidation.

Two things are true at once, and they are not in tension. Starting with AI has never been cheaper. An API key and an afternoon gets a working demonstration. Keeping it has not got cheaper at all — and almost none of the cost of keeping it is the model.

What “infrastructure” actually means here

When only a fifth of a room says they are preparing AI infrastructure, it is worth being concrete about what the other four-fifths will discover they needed. In our experience running this, the list is dull and non-negotiable:

• Somewhere to keep the weights, with enough space for the next model, which will be larger.
• An inference endpoint with an identity in front of it, so “who asked what” is a query and not a guess.
• Logs that outlive the model version, because the question arrives a year later.
• A second serving engine you have actually tested, so one project's roadmap is not your roadmap.
• Someone on call. This is the line item that turns a pilot into a service, and the one that never appears in the pilot's budget.

None of that is exotic. It is simply the difference between a demonstration and a system, and it is the reason the gap between the pilot count and the infrastructure count is not a survey artefact.

Which is why we sell the boring half

ATAILA exists on the far side of that gap. We run open models on our own GPUs in our own datacentre, and what a customer buys is not the model — it is the part that comes after: the storage, the gateway, the identity, the logs, the patching, the on-call rota and the SLA. That is the substrate the fifth of that room is thinking about, and the four-fifths will meet on the day their pilot is asked to become a service.

And it is the reason we say production is where we start rather than where we hope to arrive.

A better question than “are we using AI yet?”

“If the person who built our AI pilot left tomorrow, would it still be running next month — and could anyone else change it?” The answer separates the pilots from the systems far more reliably than any adoption percentage, and you do not need a survey to find it out.