If you want to train your own AI model without a US cloud, the first practical question is where the compute comes from. Europe has built an answer, and the Netherlands is getting its own site in Groningen.
The EU AI Factories are AI-capable supercomputers run under the EuroHPC Joint Undertaking, where European companies, researchers and public bodies can apply for compute time to train and develop AI models. There are now 19 across Europe, plus 13 smaller "antennas". For SMEs the compute time is free. For a company with 250 or more staff, it isn't.
That last point rarely makes the coverage, and for many mid-sized companies it decides whether an AI Factory is an option at all.
What is the Dutch AI Factory in Groningen, and when can you use it?
EuroHPC selected the Dutch AI Factory in October 2025. According to the Dutch government's letter to parliament of 19 December 2025 it costs just over €202 million: €71.7 million from EuroHPC, €70.9 million from the Dutch state and €60 million from the Nij Begun fund for Groningen. Over €150 million goes to hardware and operations, €50 million to an expertise centre.
That letter planned the expertise centre to open in April 2026, in the former Niemeyer factory in Groningen. The supercomputer itself goes to the Zernike campus, in a Eurofiber data centre. SURF announced that choice on 21 September 2026: installation in 2027, fully operational in early 2028.
How many GPUs the machine gets isn't settled. The purchase budget is €64 million and GPU prices have risen sharply since the plans were made. The AI Factory itself writes that earlier GPU numbers can no longer be taken as the starting point and that the machine will be smaller than originally planned. The tender is expected to close in January 2027. Any GPU count you see online today is speculation.
Until the machine runs, the expertise centre helps companies apply to other European AI Factories. So for a Dutch company in 2026 and 2027, the route is: apply to EuroHPC, compute in Finland, Spain, Italy, Germany or Bulgaria.
Who gets free access to an EU AI Factory?
Free access to an EU AI Factory is only for SMEs and startups. EuroHPC's access policy says it plainly: industrial innovation access is "open and free-of-charge to AI SMEs (including startups) for innovation purposes. Other industrial applications can benefit from pay-per-use commercial access."
SME here means the EU definition: fewer than 250 staff and turnover up to €50 million or a balance sheet up to €43 million. A 300-person company pays. EuroHPC can reserve up to 20% of capacity for commercial use, at a price based on acquisition and operating costs and "aligned with commercial offerings of the Hosting Entity". There's no public price list. You ask for a quote.
The Dutch AI Factory follows the same line. According to its own FAQ, use is free in principle, particularly for innovative SMEs doing training and R&D, and the compute counts as state aid. Once a model goes into commercial use, it's expected to move to a production environment.
Which access modes exist, and what are the conditions?
EuroHPC offers three access modes for industrial innovation, with different sizes and turnaround times:
| Access mode | Size | Turnaround | Duration |
|---|---|---|---|
| Playground | entry level; per the Dutch AI Factory up to 5,000 GPU hours | access within 2 working days | 1 to 3 months |
| Fast Lane | up to 50,000 GPU hours | approval within 4 working days | up to 3 months |
| Large Scale | more than 50,000 GPU hours | approval within 10 working days of a cut-off, twice a month | 3, 6 or 12 months |
For a business, the conditions matter more than the size. From the Terms of Reference for the access calls:
- Data and models stay yours, and SMEs may use the results commercially.
- You publish the results, except confidential parts, and acknowledge EuroHPC. Within three months of the end you deliver a final report, which EuroHPC may publish after a year.
- Under Large Scale, SMEs don't have to release the trained model itself, only a report on the results.
- Proposals are in English, for civilian use only, and the project must comply with the AI Act.
- There are no extensions or top-ups. Usage is monitored monthly, and repeated under-use cuts your allocation.
Demand is high, too. On 9 October 2026, EuroHPC's pages say MareNostrum 5 and the JUPITER Booster are unavailable for Fast Lane, and MareNostrum and LUMI-G for Large Scale, "due to exceptionally high demand".
What an application looks like in practice is described by the Dutch AI Factory through the first northern Dutch SME to get access. Its first application, asking for the maximum, was firmly rejected. And the application asked for insight into the program code, which the company said isn't something an SME shares lightly. Expect to justify your request and to be open about your approach.
Time saved
Save 16 hours per week on preparing a EuroHPC application and justifying the compute needs of a training project
How much compute does your project actually need?
Most companies that want their own model don't need a supercomputer. What they want is to adapt an existing open-weight model to their own tasks, and that takes hours, not months. Three reference points, priced at OVHcloud's public rate of €2.80 per hour for one H100 GPU (excluding VAT, October 2026):
| Task | Compute | Indicative cost if rented |
|---|---|---|
| Fine-tune an open-weight model with QLoRA | in the QLoRA paper, 24 hours on one GPU for a 65-billion-parameter model | 24 × €2.80 ≈ €67 |
| Maximum Fast Lane allocation | 50,000 GPU hours | 50,000 × €2.80 = €140,000 |
| Pretrain an 8B model from scratch | 1.46 million H100 hours for Llama 3.1 8B | 1.46m × €2.80 ≈ €4.1 million |
The sources for the first and last rows are the QLoRA paper by Dettmers et al. and the Llama 3.1 model card. The costs are indicative: a real project runs several experiments, and GPU prices are rising. Nebius raised its H100 price from $3.85 to $4.50 per hour on 1 October 2026.
The table shows where an AI Factory makes sense. Fine-tuning a model for one task costs about as much as lunch. A company with 50 to 500 staff doesn't pretrain from scratch. In between sits the work AI Factories are built for: continued training on a large domain corpus, such as tens of thousands of contracts or technical files, or distilling a large model into a small one, with many experiments and large volumes of synthetic training data. Those are projects of thousands to tens of thousands of GPU hours, where a free allocation quickly saves tens of thousands of euros.
When is an AI Factory the wrong choice?
An AI Factory is the wrong choice for running a model in production. The factories are meant for development in what the Dutch government calls the pre-competitive phase. Answering staff or customer questions day to day, the inference, belongs on your own hardware or with a European cloud provider. What that takes is covered in AI hardware for business.
You're also better off renting or buying when:
- You're not an SME. You'd pay a price you only learn after applying, through the same application process. A European provider with a public price list is more predictable.
- You can't publish the results. A report on your approach and results becomes public. For a model that's your competitive edge, check carefully what you can mark as confidential.
- Your project fits in a day on one GPU. The application takes longer than the compute.
- You need it within a few weeks. The big machines are full, and the Dutch one won't run until 2028.
For gigafactories, the larger commercial tier Europe has now opened a call for, Dutch companies can't count on the government. The Dutch cabinet wrote on 31 March 2026 that the current budget has no room to buy compute at a gigafactory. According to the AI Factory, the Netherlands also won't bid to host one.
How does this fit a local AI strategy?
An AI Factory serves one phase of a local AI project: the point where you build a model that's better at your task than anything you can download. For most companies that phase only comes after an existing open-weight model with good retrieval over your own documents turns out not to be enough. When training your own AI model pays off and when it doesn't is covered separately, as is the role of Dutch models such as GPT-NL.
The full picture, from model choice to operations, is in our guide to local AI for business. If you want to know whether your project fits a rented GPU or is worth a EuroHPC application, we'll work through the numbers with you as part of local AI.