How to Get EU AI Compute for Your Startup Before the Gigafactories Arrive

Knowing how to get EU AI compute just became a real competitive advantage. At the EU’s first AI summit on November 13, 2026, Ursula von der Leyen said seven AI Gigafactories are being planned, backed by around 30 billion euro.

As The Sofia Globe reported, she said these sites “will provide the computing capacities our startups need to build the most advanced AI models,” and estimated that Europe needs at least 500 billion euro for full tech sovereignty. Gigafactories will take years to build. But public AI compute in Europe is not a future promise only. The EuroHPC Joint Undertaking already runs supercomputers and AI Factories that startups, SMEs and researchers can apply to use today. If you prepare now, you are first in line when gigafactory capacity opens. Here is how to do it.

What You Need Before You Start

• An EU or associated-country entity: most public access calls are aimed at organisations established in EU member states or countries associated with the EuroHPC programme. Check eligibility for each call.
• A clear project description: what model you are training or fine-tuning, on what data, and why it matters.
• A compute estimate: in GPU-hours or node-hours. Reviewers want numbers, not “a lot.”
• Containerised code: Docker or Apptainer images that run your training jobs reproducibly.
• Basic AI Act awareness: the EU AI Act obligations for general-purpose AI models started applying in August 2025, so your data and documentation should already be in order.

Step 1: Estimate Your Compute Needs Honestly

Start with a small benchmark. Run your training script on one or two GPUs for a short time, measure throughput (tokens or samples per second) and extrapolate to your full dataset and number of epochs. Add 30 to 50% for failed runs, hyperparameter tuning and evaluation.

Split the result into phases: experimentation, main training, and fine-tuning or inference testing. Public calls often have different tracks for small exploratory projects and large-scale runs, and you will have a better chance asking for the right size in the right track than asking for everything at once.

Step 2: Find the Right Access Call

Go to the EuroHPC Joint Undertaking website and look at the access calls section. EuroHPC runs calls for its supercomputers and dedicated AI Factory access, with modes that range from fast, small allocations for testing to large-scale allocations for serious training runs. The European Commission’s AI Factories page explains how the AI Factories network fits together and where the sites are.

Also check your national AI Factory or supercomputing centre. Several countries host AI Factories that run their own onboarding, support desks and training for startups. They can often help you write a stronger application.

Step 3: Write a Strong Technical Application

Treat the application like a grant proposal. Reviewers look for three things: technical readiness, impact and a realistic plan. Show that your code already runs on GPUs and scales across multiple nodes (a scaling chart from 1 to 8 GPUs helps a lot). Explain the problem your model solves and who benefits. Give a timeline with milestones.

Be specific about software. Name your framework (PyTorch, JAX), your distributed training approach (FSDP, DeepSpeed, Megatron-style tensor parallelism) and your data pipeline. Vague applications get smaller allocations or none at all.

Step 4: Make Your Workloads Portable

Supercomputers are not cloud VMs. Most use Slurm for scheduling, shared parallel file systems and strict security rules, and you usually cannot run Docker directly. Convert your images to Apptainer (formerly Singularity), test them on a Slurm cluster if you have access, and write job scripts that checkpoint regularly so a failed job does not waste days of allocation.

Keep your code hardware-flexible too. Different EuroHPC systems use different accelerators, and gigafactories may use yet another generation. Avoid hard-coding vendor-specific tricks where you can, and test on more than one GPU type if possible.

Step 5: Get Your Data and Compliance House in Order

Public compute comes with expectations. Document where your training data comes from, what licences apply and how you handle personal data under GDPR. If you are building a general-purpose AI model, the AI Act requires technical documentation and a summary of training content. Doing this now saves you a painful scramble later, and it strengthens your application.

Also think about data transfer. Moving terabytes onto a supercomputer takes planning. Check the centre’s data transfer options early and clean your dataset before upload, not after.

Step 6: Plan a Hybrid Setup Until Gigafactories Open

Public allocations are time-limited and competitive, so do not bet the company on one call. Run a hybrid plan: public EU compute for big training runs, European cloud providers for steady inference and development, and hyperscalers where they make sense. Design your stack so models and checkpoints move easily between them. When gigafactory capacity becomes available, you will be able to shift workloads in weeks, not months.

A practical way to do this is to keep three things provider-neutral: your container images, your experiment tracking (MLflow or Weights & Biases) and your checkpoint format (safetensors is a good default). Store datasets and checkpoints in S3-compatible object storage so any site can pull them. Then write a short runbook for moving a job from one environment to another, and test it once a quarter. It sounds like overhead, but it is what lets a small team jump on a new allocation the week it is granted.

Finally, keep an eye on the Commission’s follow-up. Von der Leyen promised “an ambitious proposal” on financing, and the Commission said the first part of its sector-focused AI work would land within weeks. Sign up for EuroHPC and national AI Factory newsletters so you hear about new calls the day they open, not a month later.

Common Mistakes to Avoid

Applying without benchmarks. A request for “100,000 GPU-hours” with no scaling evidence looks like guesswork. Run small tests first and show the numbers.

Assuming supercomputers work like AWS. No root access, batch queues, shared storage. Teams that do not prepare for Slurm and Apptainer lose a big part of their allocation to setup problems.

Leaving compliance until the end. Data provenance and AI Act documentation take time. Start them alongside your technical work, not after training finishes.

Key Takeaways

• Do not wait: EuroHPC supercomputers and AI Factories offer access calls now, well before gigafactories exist.
• Quantify needs: Benchmark and estimate GPU-hours by phase.
• Write like a grant: Show readiness, impact and a realistic timeline.
• Go portable: Apptainer, Slurm job scripts and frequent checkpoints.
• Stay compliant: GDPR and AI Act documentation are part of being ready.

Need Expert Help?

If this feels like a lot to manage alone, TecniForge can handle the heavy lifting. Our team specializes in custom software development and AI integration. Get in touch with our experts.

Also read: EU AI Gigafactories: 7 Big Takeaways From Europe’s First AI Summit, our earlier coverage on why this matters today.

Your move: this week, run a one-hour benchmark of your main training job and write down a GPU-hour estimate. With that single number, you are ready to open the EuroHPC calls page and work out how to get EU AI compute for your next model.