Home/ What is an AI factory
What is an
AI factory?
It is a building designed around one job: running the computers that train and operate artificial intelligence. The name is borrowed from manufacturing because the building works like a production line. Power and data go in at one end, and finished output comes out at the other, without stopping.
01A word with two meanings
Careful: "AI factory" is used in two different ways.
1. The industry term
The chip maker NVIDIA popularised the phrase. In its definition, an AI factory is computing infrastructure built to manage the whole life of an AI system, from taking in data through training to answering large volumes of questions. It says the output is intelligence, measured in tokens, and that the useful yardsticks are tokens per second and tokens per unit of energy.[10]
Anyone can use the phrase this way. It describes a type of building, not a licence or a title.
2. The official EU programme
The European Union runs a scheme with the same name. An EuroHPC AI Factory is a designated national hub built around a publicly funded supercomputer. It gives startups, small companies and researchers access to that machine, plus data, training and technical help.[7][9]
By October 2025, nineteen sites had been selected across Europe, in countries including Finland, Germany, Spain, Italy, Poland and Lithuania.[7][8]
Estonia has signed the EuroHPC joint procurement agreement, but no Estonian site appears on the list of selected AI Factory hosts.[7][8] The Hüüru project is therefore an AI factory in the first sense, a commercial building being fitted out for AI work. It is not an EU-designated AI Factory. Anyone reading about the project should keep the two apart.
02The difference
An ordinary data centre stores and serves. An AI factory calculates.
Both are large buildings full of computers. The difference is how hard those computers work and how much heat they make while doing it.
| Ordinary data centre | AI factory | |
|---|---|---|
| Main job | Hosting websites, email, databases and business software | Training AI models and answering large volumes of AI requests |
| Power per cabinet | About 8 to 12 kilowatts[12] | 120 to 130 kilowatts for a current high-end AI rack[12] |
| Cooling | Air is usually enough below about 20 kilowatts per cabinet[12] | Above about 75 kilowatts, only liquid cooling works[12] |
| Load pattern | Rises and falls with the working day | Close to full load for days or weeks without a break |
| Main processor | General purpose CPUs | GPUs and other accelerators working in large groups[10] |
| Measured by | Uptime, storage, bandwidth | Output per second and output per unit of energy[10] |
A normal data centre is a library. An AI factory is a foundry. Both hold valuable things, but only one of them runs hot enough that the building has to be designed around the heat.
03What is inside
Seven parts. Get any one of them wrong and the rest is useless.
Power
Everything starts with a connection to the electricity grid that is large enough and steady enough. This is usually the hardest part to arrange, because grid capacity has to be granted by the network operator and can take years.
The number quoted for any site, such as 31.5 megawatts at Hüüru, is the ceiling for the whole campus, covering the computers plus the cooling and everything else.[3]
Cooling
Nearly all the electricity a computer uses turns into heat. Remove the heat too slowly and the chips slow themselves down to avoid damage, so the whole investment underperforms.
Air can carry heat away up to a point. Above roughly 75 kilowatts per cabinet, only liquid cooling is able to cope, which means pipes and coolant running to the chips themselves.[12] This is the single biggest physical change when a normal data centre is converted for AI.
The chips
AI work is done by graphics processors, called GPUs, and similar accelerator chips. They are good at doing the same simple sum across enormous amounts of numbers at once, which is exactly what training a model requires.
Importantly, the chips are usually not owned by the building. The site operator provides space, power, cooling and network, and the customer installs their own hardware.[3]
The network inside
When thousands of chips work on the same problem, they must constantly exchange results. If that internal wiring is slow, the expensive chips sit idle waiting for each other. Specialised high speed interconnects exist for exactly this reason.[10]
The network outside
Data has to reach the building and answers have to get back to users. Distance costs time, so the delay to major internet hubs matters. Estonia sits close to the Nordic backbone, with under 3 milliseconds quoted between Tallinn and Helsinki.[1]
A site described as carrier neutral lets customers choose their own network providers rather than being tied to one.[3]
Storage and data
Training a model means reading very large amounts of data over and over. Storage has to feed the chips fast enough to keep them busy, otherwise the same idling problem appears.[10]
This is also where the legal questions land. The physical location of the storage decides which country's rules apply to the data on it.
Security and certification
Physical protection matters as much as digital protection, because anyone who can walk up to a machine can eventually get into it. Sites use fencing, staffed guarding, monitored access and duplicated equipment so a single failure does not stop the service.[3]
Independent certification is how a customer checks these claims. EN 50600 is the European standard for data centres, and its levels describe how much redundancy is built in.[3][4]
04Plain words
The words you will hear, without the jargon.
Megawatt (MW)
A unit of power, meaning how much electricity is being used at any moment. One megawatt is a million watts. It describes the size of the tap, not the size of the bucket.
Terawatt hour (TWh)
A unit of energy, meaning how much electricity was used over time. That is the bucket. All the world's data centres used 415 TWh during 2024.[11]
PUE
Power Usage Effectiveness. Divide the total electricity the building uses by the electricity that actually reaches the computers. A result of 1.5 means half as much again is spent on cooling and losses. A result of 1.2 means only a fifth extra. Lower is better.
GPU
Graphics Processing Unit. Originally built to draw video game images, it turned out to be ideal for AI because both jobs involve doing the same calculation across huge numbers of values at once.
Token
The small unit an AI model reads and writes, usually a word or part of a word. NVIDIA describes an AI factory's production in tokens, the way a mill measures output in tonnes.[10]
Colocation
Renting space, power and cooling in someone else's building for your own computers. The customer owns the machines, the operator owns the building.
Carrier neutral
The building does not force you to buy internet connectivity from one company. Several providers are available, so customers can choose and can switch.[3]
Sovereign compute
Computing capacity that sits inside a particular country or legal area, so that only that area's laws and courts govern the data on it. The point is legal control rather than national pride.