The European Union has revealed the winners of its “Large AI Grand Challenge,” launched earlier this year to boost local innovation among large-scale AI model creators. Four startups will share €1 million in prize money and, crucially, gain access to 8 million GPU hours for training their models on select high-performance computing (HPC) supercomputers over the next 12 months. The Commission anticipates that this support will drastically reduce model training times “from years to weeks.”
The four winning startups, listed in alphabetical order, are:
- Lingua Custodia (France): A fintech company that specializes in processing financial documents using natural language processing (NLP).
- Textgain (Belgium): This startup also employs NLP for text processing with a focus on analyzing unstructured data, including monitoring social media for hate speech.
- Tilde (Latvia): A language expert emphasizing Balto-Slavic languages, offering machine translation and AI-driven chatbots in these languages.
- Unbabel (Portugal): Known for merging machine translation with native speaker expertise, Unbabel employs AI to enhance customer service and productivity for businesses.
The Commission noted that the AI Challenge received 94 proposals in total. Among the winners, Unbabel stands out, having been in operation for nearly a decade and raising nearly $100 million, according to Crunchbase. While the necessity of additional funding or GPU training hours might be debated, even seasoned AI startups stand to benefit from such resources, especially considering the rapid advancements in generative AI over the last 18 months.
At the conclusion of the training phase, the EU expects all winners to either release their developed models under an open source license for noncommercial use or publish their research findings.
To facilitate AI startups’ access to supercomputing resources, the EU announced plans last fall in President Ursula von der Leyen's state of the union address, emphasizing support for “ethical and responsible AI startups.” The European High Performance Computing Joint Undertaking (EuroHPC JU) currently operates eight supercomputers (with nine more procured), with two—Finland's Lumi and Italy's Leonardo—providing the 8 million GPU hours to the winners.
Additionally, Spain-based Multiverse Computing, which focuses on enhancing energy efficiency and speed of large language models through quantum-inspired tensor networks, narrowly missed out on prize money but will receive 800,000 computational hours on Spain’s MareNostrum 5 supercomputer.
These emerging European AI startups are not the first to experience the capabilities of HPC hardware. For instance, French AI model creator Mistral participated in a pilot phase last summer, utilizing Leonardo for small experiments, although it hadn’t advanced to model training at that time. EuroHPC JU has provided some capacity to commercial players in the past, but demand typically exceeds availability, allowing AI startups to take precedence.
Recognizing the need to adapt the HPC infrastructure for the generative AI era, the Commission announced a set of “AI innovation” measures in January, which include plans to upgrade supercomputing resources and enhance accessibility for AI startups looking to leverage this vital infrastructure for model training.
Ranking
