Join the waitlist →

Home/ What is an AI factory

What is an
AI factory?

A building designed around one job: running the computers that train and operate artificial intelligence. The name is borrowed from manufacturing, because it works like a production line that never stops.

01A word with two meanings

Careful: "AI factory" is used in two different ways.

1. The industry term

Popularised by the chip maker NVIDIA. Any computing site built to take in data, train models and answer questions at volume. Output is measured in tokens per second and tokens per unit of energy.[10] Anyone may use the phrase this way.

2. The official EU programme

A EuroHPC AI Factory is a designated national hub built around a publicly funded supercomputer, open to startups, small companies and researchers. Nineteen sites had been selected across Europe by October 2025.[7][9]

Fact check

Estonia signed the EuroHPC joint procurement agreement, but no Estonian site is on the list of selected hosts.[7][8] Hüüru is an AI factory in the first sense: a commercial building being fitted out for AI work.

02The difference

An ordinary data centre stores and serves. An AI factory calculates.

Ordinary data centre

Holds data, serves it, moves it. What comes out is what went in.

AI factory

Takes data in, trains a model on it, improves it, then runs it to produce answers and predictions.[10]

Ordinary cabinet8 to 12 kW
Where plain air stops being enoughabout 20 kW
One current AI cabinet120 to 130 kW

Power drawn by a single cabinet, drawn to scale.[12]

Ordinary data centre AI factory
Runs Websites, email, databases, business software Model training, then answering requests at volume
Cooling Air, up to about 20 kW[12] Liquid at the chip, standard by 120 to 130 kW[12]
Load pattern Rises and falls with the working day Near full load for days or weeks
Main processor General purpose CPUs GPUs working in large groups[10]
Measured by Uptime, storage, bandwidth Output per second, output per unit of energy[10]
Put simply

A normal data centre is a library. An AI factory is a foundry.

03What is inside

Seven parts. Get any one of them wrong and the rest is useless.

01

Power

A grid connection large enough and steady enough. Usually the hardest part to arrange, because capacity is granted by the network operator and can take years. The figure quoted for a site, such as 31.5 MW at Hüüru, is the ceiling for the whole campus.[3]

Rule of thumb
1 MW
is roughly a few hundred homes, running constantly.
02

Cooling

Nearly all the electricity a computer uses turns into heat. Remove it too slowly and the chips throttle themselves. Plain air stops being enough past roughly 20 kilowatts per cabinet, and by the 120 to 130 kilowatts a current AI rack draws, the industry runs coolant straight to the chips.[12] This is the biggest physical change when a data centre is converted for AI.

Efficiency measure
PUE
Building power divided by power reaching the computers. 1.0 is perfect. Hüüru targets under 1.2.[3]
03

The chips

Graphics processors and similar accelerators, good at doing the same sum across enormous amounts of numbers at once. They are usually not owned by the building: the operator provides space, power, cooling and network, and the customer installs the hardware.[3]

Why they cluster
1000s
of chips may work on one model together, so they sit close and wire fast.
04

The network inside

Thousands of chips on one problem must constantly exchange results. If the internal wiring is slow, expensive chips sit idle waiting for each other.[10]

Analogy
Kitchen
A hundred chefs cook faster only if they can pass things quickly. Otherwise they queue.
05

The network outside

Data has to reach the building and answers have to get back, so distance to the major hubs matters. A carrier neutral site lets customers pick their own providers rather than being tied to one.[1][3]

Tallinn to Helsinki
< 3 ms
About a thousand times faster than a blink.[1]
06

Storage and data

Training means reading very large amounts of data over and over, fast enough to keep the chips busy.[10] This is also where the legal question lands, because the physical location of the storage decides whose rules apply.

Why location matters
GDPR
The EU rules follow the data. Storage inside the EU keeps one framework in charge.
07

Security and certification

Anyone who can walk up to a machine can eventually get into it, so sites use fencing, guarding, monitored access and duplicated equipment. Independent certification is how a customer checks those claims, and EN 50600 is the European standard.[3][4]

At Hüüru
EN 50600
Level 3 to 4, reported as the only such certificate in the Baltics and Finland.[3][4]

04Plain words

The words you will hear, without the jargon.

Megawatt (MW)

How much electricity is being used at any moment. The size of the tap, not the size of the bucket.

Terawatt hour (TWh)

How much electricity was used over time. That is the bucket. The world's data centres used 415 TWh during 2024.[11]

PUE

Total building electricity divided by the electricity that reaches the computers. 1.5 means half as much again goes on cooling and losses. Lower is better.

GPU

A chip built to draw video game images, which turned out to suit AI: both jobs repeat one calculation across huge numbers of values.

Token

The unit an AI model reads and writes, usually a word or part of one. NVIDIA measures an AI factory's output in tokens, the way a mill measures tonnes.[10]

Colocation

Renting space, power and cooling for your own computers. The customer owns the machines, the operator owns the building.

Carrier neutral

The building does not force you to buy connectivity from one company. Customers can choose, and can switch.[3]

Sovereign compute

Capacity that sits inside one legal area, so only that area's laws and courts govern the data on it. Legal control, not national pride.