Home/ What it is used for
Who actually
needs this?
AI computing capacity is not bought for its own sake. It is bought by hospitals, weather services, factories, banks and universities that have a specific job to do. Here is what those jobs look like, and why the location of the machine keeps coming up.
01Start with something ordinary
Five ordinary things. Five different reasons the machine has to be here.
Pick one. The three bars show what it asks of the machine. They change from one to the next, and that is the whole argument.
One swipe, one fresh prediction, made before the screen finishes moving.
Tallinn to Helsinki is stated at under 3 milliseconds.[1]
It works today because somebody else built the building. The rules and the money live there too.
It answers from your own contracts and email. Nothing leaves the building.
Public models are ruled out where confidentiality is required.[13]
The choice is two. Send the contracts abroad, or have no assistant.
Text goes in, Estonian comes out, faster than it is spoken.
Estonian language models are built in Tartu and at TalTech.[13]
No foreign company has a reason to do this for a million people.
The model marks what the radiologist should look at first.
An AI rack draws 120 to 130 kW, an ordinary one 8 to 12.[12]
Without a machine under Estonian rules, the hospital simply waits.
Thousands of chips, wired as one machine, running for weeks.
Europe calls the shortage of this capacity a bottleneck.[14]
You use somebody else's forecast, drawn for somebody else's coastline.
The bars compare the five with each other. They are a rough profile, not a measurement. The sourced detail is in the sections below.
02Where it gets used
The European AI factories were set up around specific sectors.
Health
Reading scans, sorting patient records, and supporting diagnosis. Health is named repeatedly in the EU programme, and one selected site is dedicated to it.[7][15] Patient data is exactly the kind that cannot easily leave the country.
Weather and climate
AI driven forecasting is listed as a high impact use case in the EU network.[15] Better short range forecasts matter for farming, shipping, energy trading and storm warnings.
Manufacturing
Spotting defects on a production line, predicting when a machine will fail, and planning maintenance before it stops. Manufacturing and automotive are among the named focus areas.[15]
Finance
Detecting fraud in payment streams and assessing risk. Finance is a named sector in the EU AI factory network, and it is heavily regulated on where data may sit.[7][15]
Public services
Government bodies use AI for translation, document handling and citizen support. The EU network explicitly includes public bodies alongside startups and researchers.[7]
Startups and small firms
Buying a cluster of AI chips outright is beyond most small companies. Shared capacity is how they get access, which is the stated purpose of the EU scheme.[7]
03The Estonian case
A language with a million speakers has to look after itself.
Large AI models are trained mostly on English. They can produce Estonian, but they often do it by translating from English rather than reasoning in the language itself, which loses the way things are actually said.
Estonian research groups have been working on this directly. TartuNLP and TalTechNLP have built Estonian language models, funded by the Ministry of Education and Research under the national Language Technology Programme running from 2018 to 2027.[13]
There is a second reason beyond language quality. Public models run by large foreign companies cannot be used for work that requires confidentiality, such as national defence or healthcare. That rules them out for a large part of what a state actually needs.[13]
Associate Professor Kairit Sirts of the University of Tartu has put the argument plainly: for technology companies, the Estonian language and cultural background is not something that matters, so Estonians have to look after these things themselves.[13]
Training a language model needs somewhere to run it. That is the link between a research group in Tartu and a building full of power and cooling in Hüüru.
04Why location keeps coming up
The rules that apply to your data depend on where the machine is.
If a hospital in Tallinn sends patient records to a computer in another jurisdiction, it takes on a legal problem that has nothing to do with technology. Which courts decide disputes. Which government can compel access. Which regulator audits the arrangement.
Keeping the machine inside the European Union removes most of that question. This is the practical meaning of the phrase sovereign compute, and it is why governments are funding buildings rather than only funding software.
Europe has been open that it is short of this capacity. The European Commission calls the shortage a major bottleneck for its ability to compete and for its strategic autonomy.[14]
Its answer has two parts. The EuroHPC AI Factories network gives shared access to publicly funded supercomputers, with nineteen sites selected by October 2025.[7] The InvestAI initiative aims to mobilise up to €200 billion, including a dedicated fund for much larger AI gigafactories.[14]
Estonia signed the EuroHPC joint procurement agreement alongside seventeen other member states, but has not been selected as an AI Factory host site.[7][8] Neighbouring Lithuania has been, through the LitAI Factory selected in October 2025.[7] Anyone claiming the Hüüru site is the first AI factory in the Baltics should be asked which definition they are using.
05The fair objection
These buildings use a lot of electricity. That objection is legitimate.
The International Energy Agency put data centre electricity use at 415 terawatt hours during 2024, about 1.5% of world consumption. It expects that to more than double to around 945 terawatt hours by 2030, slightly more than Japan uses today, and to reach roughly 1,200 terawatt hours by 2035.[11]
Demand from AI optimised data centres alone is expected to more than quadruple by 2030.[11] In advanced economies, data centres account for more than a fifth of all electricity demand growth to 2030.[11]
What honest answers look like
There are three real levers, and all three are measurable rather than rhetorical. Where the electricity comes from, how efficiently the building uses it, and whether the leftover heat is captured instead of thrown away.
On the first two, the Hüüru operator publishes a 100% renewable supply and a target PUE under 1.2.[3] On the third, we found no public evidence either way, so it belongs on the unknown list rather than in a selling point.
The IEA also notes the other direction of the relationship. AI is being used to forecast renewable output and manage grids, and it estimates that up to 175 gigawatts of transmission capacity could be unlocked with such tools.[11]
Capacity opens
in Q4 2026.
The waitlist is not open yet. When it opens in the fourth quarter of 2026, places will be handled in the order they arrive.