The European Union is trying to turn its high performance computing network into a practical advantage for homegrown generative AI startups. The Commission has adopted an AI innovation package that aims to help startups and scale ups train models, build applications and compete in a market where access to computing power has become a central constraint.
The move reflects a clear gap in the bloc's AI strategy. The EU already has a network of supercomputers, including the newly inaugurated MareNostrum 5, but that infrastructure was historically shaped around scientific users and scientific workloads. Generative AI has changed the compute equation.
Why Supercomputers Are Now an AI Priority
Generative AI models are trained on very large data-sets and can produce outputs such as text, imagery or audio on demand. That makes computing power a core input, not a secondary technical detail. Commission officials acknowledged that lawmakers were caught off guard by the sudden rise of tools such as OpenAI's ChatGPT last year.
The Commission is now trying to adapt existing strategic infrastructure to serve AI companies as well as researchers. The package focuses on high performance computing infrastructure for model training, access to talent and skills, and services that can help startups use the systems effectively.
Margrethe Vestager, the EU's digital chief, framed the issue directly: "You need computing power to develop AI. A lot of it. So we want to give SMEs and start-ups privileged access to the network of European supercomputers. We are committed to innovation of AI and innovation with AI. And we will do our best to build a thriving AI ecosystem in Europe."
What AI Factories Are Meant to Do
A central idea in the package is the creation of "AI Factories" around the bloc's supercomputers. These are intended to combine the main inputs that AI startups need: computing power, data, algorithms, talent and practical support.
Thierry Breton, the internal market commissioner, described the factories as a "one-stop shop" for Europe's AI start-ups. The goal is to make it easier for companies to develop advanced AI models and industrial applications without having to assemble every piece of infrastructure and support alone.
The package includes several connected measures:
- AI-focused hardware upgrades for existing supercomputers.
- Privileged access for AI startups and the broader innovation community.
- Support services, including supercomputer-friendly programming facilities and other AI enabling services.
- Measures to foster AI applications based on General Purpose AI models.
The Commission had already launched the Large AI Grand Challenge in November, offering AI startups financial support and supercomputing access. The new package extends that direction into a broader policy effort.
The Capacity Problem Is Immediate
The strategy depends on upgrading infrastructure that is already under pressure. Commission officials said the EU's supercomputers are overbooked by a factor of at least two. They also said the systems were not designed with this type of AI model training in mind.
To move faster, the Commission has proposed amending the Regulation establishing the European High Performance Computing Joint Undertaking. The change is meant to make it easier to upgrade existing supercomputers instead of waiting for a new AI-dedicated system to be procured.
One Commission official said it would be more cost efficient to double the capacity of existing supercomputers by adding AI capabilities now rather than wait another two or three years. Another official suggested that, with more flexibility, the EU could almost double capacity within a year.
In practice, the upgrades would mean adding many racks packed with accelerators, including GPUs, and rewiring components so data can move efficiently. That technical work matters because model training depends not only on raw compute, but also on getting data to the right place at the right speed.
Free Access Could Help, But Demand May Surge
Research uses of the EU's supercomputing infrastructure are free, and AI startups doing model pre-training or training will also be able to benefit from free access. That could help reduce the gap between startups that have close cloud computing partnerships and those that do not.
But free access also increases demand. If many AI startups try to use the same infrastructure, the bottleneck could simply move from cost to availability. That is why the success of the plan rests heavily on whether the EU can deliver upgrades quickly.
Chip supply is another risk. Commission officials admitted that delays are possible because of high global demand for GPUs. They said the bloc has been seeking commitments from Nvidia in relation to supercomputer procurement, while also looking to longer-term measures such as the Chips Act to support European access to accelerators.
One official summed up the uncertainty plainly: "There's no guarantee we will not suffer similar shortage or delay as other clients for Nvidia. But we're working on a number of mitigation measures, short term or long term,"
Beyond Compute, the Package Targets Data and Applications
The AI innovation package is not only about machines. The Commission also plans extra financial support for AI startups through the existing Horizon Europe and Digital Europe programs, with a specific focus on generative AI. It expects this to generate additional overall public and private investment of around €4BN until 2027.
The EU also says it will seek more public and private investment in AI start-ups and scale-ups, including through venture capital or equity support. The article identifies the EIC accelerator Programme and InvestEU as examples of channels for this support.
On data, the package aims to accelerate the development and deployment of Common European Data Spaces for the AI community. The Commission is also flagging a GenAI4EU initiative to support novel use cases and emerging AI applications across Europe's 14 industrial ecosystems and the public sector. Application areas named by the Commission include robotics, health, biotech, manufacturing, mobility, climate and virtual worlds.
Two European Digital Infrastructure Consortiums are also being established. The Alliance for Language Technologies will focus on common European infrastructure in language technologies and on addressing a shortage of European languages data for AI training. CitiVerse will support AI tools for city planning and management, including ways to model and optimise processes such as traffic management and waste management.
How This Fits With EU AI Regulation
The support package is designed to sit alongside the AI Act, the EU's emerging regulatory framework for artificial intelligence. The Commission presents the AI Act as a way to support adoption by building user trust, while the new package focuses more directly on AI builders.
The AI Act is still in draft form, even though a political deal was reached last month. The Commission is publicly confident it will be adopted, while officials privately described themselves as "hopeful".
At the same time, the Commission is setting up the AI Office, a new public body within the Commission. It will enforce the AI Act for general purpose AIs and act as a central coordination body for AI policy at EU level. The decision to establish the AI Office enters into force today, with operations expected to begin in the following months.
The supercomputing rule changes still need backing from the European Parliament and Council. That means the EU's plan has two tracks: move ahead with coordination and support now, while seeking the legal flexibility needed to upgrade the compute backbone that generative AI startups need.