A European Union initiative aimed at bolstering local AI startups by granting them access to the bloc’s supercomputers for model training has gained momentum since its announcement last September and official launch last month. France's Mistral AI has already engaged in an initial pilot phase, as revealed by a recent EU update. A key takeaway from this early stage is the necessity for added support tailored specifically for AI startups to maximize their usage of the EU’s high-performance computing resources.
“One important insight we’ve identified is that we need to offer not only access but also the necessary facilities—especially skills, knowledge, and experience from hosting centers—to help users leverage our computing power effectively,” noted an EU official during a press conference today.
The EU plans to establish "centers of excellence" dedicated to developing specialized AI algorithms optimized for operation on its supercomputers. Many AI startups have become accustomed to utilizing dedicated compute hardware from U.S. hyperscalers, rather than exploiting the capabilities of supercomputers for their model training. To address this gap, EU officials are enhancing the high-performance computing access program with additional support measures, coinciding with the unveiling of MareNostrum 5, a pre-exascale supercomputer, at the Barcelona Supercomputing Center in Spain.
“We're setting up facilities to help small and medium enterprises (SMEs) understand the best ways to utilize supercomputers and how to optimally parallelize their algorithms for AI model development,” stated a Commission official. "By 2024, we anticipate a significant increase in these outreach efforts.”
“AI is now a strategic priority for the Union,” they continued. “Alongside the AI Act, we're aiming to provide substantial innovation capacity to our SMEs and startups, enabling them to fully leverage our advanced machines and public infrastructure to compete globally in creating safe, trustworthy, and ethical AI algorithms.”
Another EU official confirmed the creation of an “AI support center” specifically designed to assist SMEs and startups in maximizing their use of the EU’s supercomputing resources. “We must acknowledge that the AI community hasn't extensively employed supercomputers over the past decade,” they remarked. “While they are experienced with GPUs, they are new to working with supercomputers, so guidance is essential.”
“While the AI community has considerable knowledge about optimizing GPU use, the supercomputers consist of multiple GPU containers, requiring additional skills and resources for effective scaling,” the official added.
The EU has significantly increased its supercomputing investments over the last five years, expanding to a cluster of eight interconnected machines planned to create a federated cloud-based supercomputing resource accessible throughout Europe. The first EU exascale supercomputers are set to launch soon, one in Germany (expected next year) and another in France (projected for 2025). The Commission also intends to invest in quantum computing, with plans for hybrid systems integrating quantum simulators alongside supercomputers to enhance classical computing capabilities.
Among the exciting applications being developed on the EU's high-performance computing infrastructure are initiatives like Destination Earth, aimed at simulating Earth’s ecosystems to improve climate change and weather modeling, and a digital twin project for the human body to advance medical science, including drug discovery and personalized medicine. The push to utilize supercomputing resources specifically for AI startups underscores a new strategic priority following the EU president's announcement of the AI model training program last fall.
Last month, the EU unveiled the “Large AI Grand Challenge,” a competition focused on AI startups experienced in large-scale models, which will select up to four promising European startups to receive a total of 4 million hours of supercomputing access for foundational model development. A prize pool of €1 million is also allocated for the winners, who are expected to release their models under an open-source license for noncommercial use or publish their findings, as outlined by the Commission.
Previously, the EU had a program allowing industry users core access to supercomputing resources through project calls. Now, the focus is intensifying on commercial AI with dedicated programs and resources—to strategically position its expanding supercomputing network as a power source for developing "Made in Europe" general-purpose AI. Notably, Mistral, an AI startup looking to compete with U.S. giants like OpenAI and claiming to offer “open assets,” is an early recipient of this supercomputer access program. Some may find it noteworthy that a tech firm recently raising €385 million in Series A funding from U.S. investors is prioritized for this EU initiative. Nonetheless, this reflects the high-level strategic investments in “large AI” ecosystems.
As the EU’s “supercomputer for AI” initiative is still in its infancy, the extent of model training benefits from dedicated access remains to be seen. Mistral co-founder and CEO Arthur Mensch mentioned, “We utilized Leonardo, one of the EU’s current supercomputers located in Bologna, for a few small experiments over the summer while the cluster was ramping up. It was a productive collaboration, yielding valuable feedback and results. However, all our models were trained on our own cluster.”
The Commission hopes that channeling support toward AI startups to utilize its high-performance computing investment, coupled with supercomputer hardware designed for AI model training, will give a competitive edge to a local AI ecosystem that currently trails behind U.S. hyperscalers. "As we lack the large hyperscalers that the U.S. has for foundational model training, we are committed to utilizing and evolving our supercomputers to be AI-compliant," an official explained. “Starting in 2024, we will focus on more SMEs harnessing supercomputers for foundational model development.”
Future plans include acquiring “more dedicated AI supercomputing machines that lean heavily on accelerators rather than standard CPUs.” The newest machine, the Barcelona-based MareNostrum 5, has been specifically designed for AI developers, per the Commission.
How the EU's AI support strategies align with national ambitions to cultivate local AI champions is yet to be determined. This was a focal point during recent discussions on the EU's AI regulation framework, where France advocated for special provisions for foundational models—a stance met with criticism from SMEs. However, Mistral's involvement in the EU’s supercomputing access program might indicate a shared vision.
This article was updated to include insights from Mistral.
Ranking
