The European Union has announced the winners of its Large-Scale AI Grand Challenge, which was launched earlier this year with the aim of accelerating the pace of homegrown innovation from large-scale AI model makers.
The four startups will share a prize of 1 million euros, and, perhaps more importantly, they will also receive 8 million GPU hours to train their models on several of the European Union’s high-performance computing (HPC) supercomputers over the next 12 months. As the European Commission states in its PR, this is expected to reduce model training times “from years to weeks.”
The four winning startups, in alphabetical order, are Lingua Custodia, a French fintech company that uses natural language processing (NLP) to process financial documents; Textgain, a Belgian startup that also uses NLP for text processing but focuses on analyzing unstructured data, such as monitoring social media chats to detect hate speech; Tilde, another language specialist focused on Baltic-Slavic languages, a Latvian startup that offers machine translation and AI-powered chatbots in its target languages; and Unbabel, a Portuguese company that has historically blended machine translation with native speaker expertise to apply AI to customer service and productivity use cases for enterprise customers.
The committee said the AI Challenge received a total of 94 proposals.
Unbabel is likely the most well-known of the four winners: The Y Combinator-backed translation business has been around for nearly a decade and has raised nearly $100 million to date, according to Crunchbase.
Whether Unbabel needs another 250,000 euros or 2 million free hours of GPU training is up for debate, but given how generative AI has developed so rapidly over the past year and a half or so, even veteran AI startups might feel like a little help would be appreciated.
At the end of the training period, the EU expects all awardees to release the models they develop under an open source license for non-commercial use or to publicly publish their research findings.
EU supercomputer to support AI startups
The EU announced plans to give startups greater access to the EU’s supercomputer hardware during President von der Leyen’s State of the Union address last fall, saying the EU wanted to ensure that “ethical and responsible AI startups” were first to receive computing help.
The European High Performance Computing Joint Undertaking (aka EuroHPC JU), the official name of the European Union’s supercomputer initiative, currently has eight operational supercomputers (nine more procured), two of which will allocate eight million GPU hours to four awardees: Finland’s Lumi and Italy’s Leonardo (both of which are pre-exascale HPC supercomputers).
The fifth startup, Spain-based Multiverse Computing, which focuses on using “quantum-inspired tensor networks” to improve the energy efficiency and speed of large-scale language models, narrowly missed out on the prize but has some consolation: the company will be allocated 800,000 computing hours on another Spanish supercomputer (pre-exascale), MareNostrum 5.
The handful of European startups building large-scale AI models are no strangers to the power of HPC hardware: French general AI model maker Mistral took part in an early pilot phase of supercomputing offerings last summer, using Leonardo to “run some small experiments,” co-founder and CEO Arthur Mensch told TechCrunch in December, though he said it wasn’t being used to train models at the time.
EuroHPC JU has historically offered some capacity to commercial players as well, but demand for supercomputers has typically far outstripped supply, effectively pushing AI startups to the front of the queue.
EU policymakers also recognize the need to reconfigure and retool HPC infrastructure for the age of generative AI. That’s why in January the European Commission unveiled an “AI Innovation” package of measures that includes proposals for upgrading supercomputers and building a support layer to make infrastructure more accessible to AI startups.
