Published: 21 September 2026
The Netherlands now knows where its AI supercomputer will stand. On 21 September 2026 SURF awarded the housing contract to Eurofiber’s existing datacenter campus on the Zernike Campus in Groningen, for a machine of roughly 1,800 GPUs inside a 201 million euro programme. It should be fully operational in early 2028. We work in Groningen, so this one lands close to home, and it is worth separating what it changes from what it does not.
What was decided on 21 September
Three sites were in the running. The former De Halm strawboard factory in Hoogkerk, an undisclosed location elsewhere in the province, and Eurofiber on Zernike. Eurofiber won because the building is already there and already connected.
That existing connection is the whole argument. The datacenter is on the electricity grid with the capacity it needs, so the supercomputer adds no new congestion. Nearly all of its waste heat goes into the WarmteStad network, enough for around 2,000 homes. Cooling uses neither surface water nor drinking water.
“Vandaag markeren we een belangrijke stap richting de realisatie van de AI-fabriek”, said SURF board member Hans Louwhoff, marking the tender award as a step towards realising the AI factory. The machine goes in during 2027. The expertise centre, the human half of the project, has been running in the Niemeyerfabriek since April 2026.
The Dutch AI factory in numbers
These are the figures that decide whether the project is relevant to your company.
| Aspect | Value |
|---|---|
| Total investment | 201 million euro |
| Where it comes from | 71 million from the EU, 70 million from the Dutch state, 60 million from Nij Begun |
| Size of the machine | About 1,800 advanced GPUs |
| Fully operational | Early 2028, with the datacenter ready in 2027 |
| Reserved for Noord-Nederland | 23 percent of the Dutch half of the capacity |
| Homes heated by its waste heat | About 2,000, through WarmteStad |
The funding split is from RTV Noord’s reporting of the announcement; the capacity allocation and the technical figures come from the project partners. Half of all capacity goes to European users outside the Netherlands, which is the condition of the European money.
How a company gets compute, and who it is not for
Access runs through the European EuroHPC framework rather than a Dutch scheme, and for AI SMEs and startups the compute is free. There are three tiers:
- Playground: up to 5,000 GPU-hours, reviewed within days, no consortium needed
- Fast Lane: 5,000 to 50,000 GPU-hours, reviewed every two months, needs a justified plan
- Large Scale: more than 50,000 GPU-hours, long review, usually a consortium
Companies that are not AI SMEs pay per use. SMEs can apply continuously through vouchers, assessed by an independent committee, and a project graduates out of the facility once it becomes commercially viable.
Be honest about the threshold, though. Fred Hassert, chair of Samenwerking Noord, has said the facility will not spend capacity on “een zelfstandige die even een jaarrekening wil maken”, a sole trader doing their annual accounts. This is built for training your own models on your own data, not for running inference you could buy from an API. The example the partners keep returning to is the UMCG’s Lifelines population study: a dataset you cannot legally ship to a US cloud.
What the announcement leaves out
The hardware has not been bought. The tender for the supercomputer itself does not close until January 2027, and in August SURF warned that it will buy less than planned because component prices have risen sharply. Some memory chips cost five times what they did a year ago, and new production capacity is not expected before 2029 or 2030.
The timeline has also moved. In June 2025 the machine was due to be fully operational in early 2027. It is now early 2028, roughly a year later. That is not scandalous for infrastructure of this size, but it does mean any plan that assumed Dutch sovereign compute in 2027 needs redrawing. The expertise centre is the one part that arrived early.
What this means
For a Groningen AI company this is good news, and the reason is not the hardware. It is that 23 percent of the Dutch capacity is earmarked for Noord-Nederland, in a country where this kind of facility normally lands in the Randstad. Combine that with an expertise centre in the city that is already staffed and already guiding projects onto other European machines, and the practical advantage starts in 2026, not 2028. If you are a Dutch company with sensitive data, healthcare records, IP-protected industrial data, anything you cannot put on a US hyperscaler, this is the route that opens.
Do not wait for 2028 to act on it, and do not overestimate what arrives then. Around 1,800 GPUs is respectable for a national facility and it is not a frontier training cluster: it will not train a model that competes with what OpenAI or Google ship, and it was never meant to. Treat it as the place to fine-tune on data you cannot move, and to run the experiments a cloud bill would kill. The step we would take this month is the boring one: apply for Playground hours on an existing EuroHPC machine, because 5,000 GPU-hours reviewed in days is available now, most Dutch companies do not know it exists, and the teams that learn the process now will be the ones ready when the Groningen machine switches on.
For more information, visit the official announcement of the AI factory location on the SURF website.
