AI Factories need a compute brief, not a wish list

Europe’s AI Factories are becoming less abstract, but they still do not turn an SME’s loose AI ambition into a usable experiment by themselves.

EuroHPC now describes a network of 19 AI Factories and 13 AI Factory Antennas offering free, customised support to SMEs and startups. That matters because the public promise is no longer only policy language. There is an access route, a support layer and a set of modes that smaller firms can actually read.

Compute is not the first question

The phrase “AI Factory” makes the offer sound like capacity first. Capacity is part of it, but it is the wrong first question for most SMEs. The useful first question is narrower: what workload is ready enough to justify shared European infrastructure?

An SME does not need a full machine-learning department to answer that. It does need one clean page that names the task, the data boundary, the current baseline and the decision the experiment should unlock.

Brief line	What to write
Workload	Fine-tuning, evaluation, synthetic data test, inference benchmark or prototype run
Data	Where it sits, whether personal data is involved, and what can leave the company
Scale	Small playground test, medium experiment or large model/application run
Support needed	Data preparation, model choice, deployment advice, evaluation or compliance route
Success test	The fixed task set that decides whether the experiment worked
Business decision	Buy, build, stop, apply for larger access or move to production partner

Source: EuroHPC AI Factories pages and access-mode descriptions. Last verified 2026-07-01.

That table is deliberately plain. It prevents the application from becoming a wish list for “AI capacity” when the company only has three untested prompts and a folder of documents.

Free access still has a shape

The EuroHPC access page draws an important line. Industrial Innovation access modes are open and free of charge to AI SMEs, including startups, for innovation purposes. Other industrial applications can use pay-per-use commercial access.

That distinction matters for ordinary companies. A small manufacturer that wants to test whether a support assistant can read maintenance manuals is not in the same position as a startup training a model as its product. Both may be interested in European compute, but their access case, evidence and expected support differ.

The smaller end of the route is also useful. Playground access is aimed at SMEs, startups and entry-level users, with rapid access, hosting-factory guidance and allocations for 1, 2 or 3 months. That is closer to how a cautious SME should start: prove one workload before writing a grand adoption plan.

The Apply AI link is practical

The Commission’s Apply AI Strategy frames AI adoption as a sector and SME competitiveness problem. AI Factories are one part of that stack: infrastructure, expertise and support around EuroHPC capacity.

The risk is that SMEs hear “AI adoption” and jump straight to tools. The more useful reading is operational. If the company has a real bottleneck, an AI Factory experiment can test whether the model layer is the bottleneck at all.

Sometimes it will not be. The problem may be data quality, unclear ownership, missing evaluation, weak security, bad procurement or a workflow that should be simplified before any model is trained. Finding that out cheaply is still a good result.

What to prepare before applying

The smallest credible compute brief has five attachments or links:

• a one-paragraph workflow description;
• a sample of the data shape, not necessarily the data itself;
• a baseline method or current tool;
• ten to fifty representative test cases;
• one named owner who can decide what happens after the experiment.

That is enough to make a first conversation concrete. It also protects the SME from treating free access as free strategy. Compute can be subsidised. Internal clarity cannot.

The signal for buyers

AI Factories are not a general-purpose outsourcing lane for every AI problem. They are most useful when the technical question is real enough that compute, model expertise or evaluation support can change the decision.

For smaller firms, the next step is not to chase the biggest allocation. It is to write the smallest workload brief that would survive expert scrutiny. If that page cannot be written, the company probably needs workflow cleanup before it needs an AI Factory.

If it can be written, Europe has made the next question easier: which access mode is small enough to learn from without pretending the experiment is already a production system?