The European Union has announced the winners of the “Large AI Grand Challenge” initiated earlier this year to boost innovation by large-scale AI model makers in Europe.
Four startups have been selected to receive €1 million in prize money and 8 million GPU hours to train their models on leading high-performance computing (HPC) supercomputers within the EU over the next 12 months. This initiative aims to significantly reduce model training times, as highlighted by the Commission.
The winning startups, in alphabetical order, are: Lingua Custodia from France, Textgain from Belgium, Tilde from Latvia, and Unbabel from Portugal. These startups specialize in various aspects of natural language processing (NLP) applications, including financial document processing, text analysis, language translation, and AI-driven chatbots.
A total of 94 proposals were submitted for the AI Challenge, with Unbabel being one of the most well-known winners. This translation startup, backed by Y Combinator, has raised close to $100 million in funding over the years.
After the training period, the EU expects all winners to release their developed models under open-source licenses or publish their research findings for noncommercial use.
EU supercomputers to support AI startups
Through the European High Performance Computing Joint Undertaking (EuroHPC JU), the EU is providing access to supercomputing resources for AI startups. The initiative aims to support “ethical and responsible AI startups” and prioritize their access to computational resources.
The EuroHPC JU currently operates eight operational (nine procured) supercomputers, with two of them, Lumi in Finland and Leonardo in Italy, offering 8 million GPU hours to the winners. Another startup, Multiverse Computing from Spain, will receive 800,000 computational hours on MareNostrum 5 supercomputer.
These startups are part of a broader effort to reconfigure HPC infrastructure for generative AI capabilities, aligning with the EU’s January announcement of AI innovation measures to boost accessibility and upgrades to supercomputers.
