View all AI news articles

Unbabel Scores Millions of GPU Hours on EU Supercomputers

June 26, 2024
But does Unbabel, a veteran AI startup, really need the extra help?
  • EU's Large AI Grand Challenge awards €1 million and 8 million GPU hours.
  • Winners: Lingua Custodia, Textgain, Tilde, and Unbabel.
  • Prize includes training on EU supercomputers Lumi and Leonardo.
  • Aim: Reduce model training from years to weeks.
  • Winners to release models or findings under open source licenses.
  • Fifth startup, Multiverse Computing, gets 800,000 GPU hours.

The European Union recently revealed the winners of its “Large AI Grand Challenge” initiated earlier this year. This competition is part of the EU's effort to boost innovation in AI model development by providing substantial computational resources. The challenge saw four startups win €1 million in prize money and an invaluable 8 million GPU hours to train their AI models on Europe's top-tier supercomputers over the next year. These resources promise to cut down model training times drastically.

The winners include French fintech Lingua Custodia, Belgian startup Textgain, Latvian language specialist Tilde, and Portugal’s Unbabel. Each of these companies brings unique AI applications to the table, ranging from financial document processing and social media analysis to machine translation and AI-powered chatbots.

Despite Unbabel's established presence in the AI industry, having raised nearly $100 million to date, the additional resources from the EU come at a crucial time. With rapid advancements in generative AI, even seasoned startups like Unbabel can benefit from every bit of support.

The EU’s initiative, announced during President Ursula von der Leyen’s state of the union address, is aimed at giving ethical and responsible AI startups priority access to computational support. The European High Performance Computing Joint Undertaking (EuroHPC JU) is spearheading this initiative, providing access to its high-performance supercomputers, Lumi and Leonardo, located in Finland and Italy respectively.

A fifth startup, Spain-based Multiverse Computing, focused on enhancing the energy efficiency and speed of large language models through quantum-inspired tensor networks, narrowly missed out on the prize money. However, it was awarded 800,000 computational hours on Spain’s MareNostrum 5 supercomputer.

This initiative isn’t the first time European startups have had access to such resources. Last summer, French AI model maker Mistral participated in an early pilot phase, running experiments on Leonardo. While demand for supercomputing resources often exceeds supply, this initiative ensures AI startups are prioritized.

The European Commission is also planning to upgrade its supercomputing infrastructure to better support the AI industry's evolving needs. This includes proposals for enhancing accessibility, ensuring AI startups can more easily leverage these powerful computational tools.

Does Unbabel, with its significant funding and established market presence, really need the additional support? Perhaps not. But in the rapidly evolving world of AI, where the pace of development can be relentless, even the most successful companies can benefit from every edge they can get. This move by the EU could serve as a model for other regions looking to bolster their own AI ecosystems.

Recent articles

View all articles