The EuroMesh project analyzes whether Europe can train a frontier AI model by federating its public AI compute resources. Using low-communication distributed training (DiLoCo-style), a frontier model could be achieved around 2028, faster than building new gigawatt datacenters (around 2033).
The EuroMesh project published a report analyzing whether Europe can train a frontier AI model by federating its existing public AI compute resources (EuroHPC supercomputers and national AI Factories). Using low-communication distributed training (DiLoCo-style), a frontier model could be achieved around 2028, faster than building new gigawatt datacenters (average 7.6 years grid connection, around 2033).
Europe plans large datacenters to boost AI competitiveness, but grid connection takes years. Meanwhile, Europe already operates tens of exaflops of public AI compute, which can be leveraged via distributed training.
This study offers a practical alternative for Europe to quickly build AI capability without massive infrastructure investment. Optimizing existing resources for frontier model development has significant implications for AI sovereignty and competitiveness.
HN users are predominantly skeptical about the EU's ability to develop AI models. Key issues include lack of cooperation among EU member states, with frequent references to the failed joint fighter jet development by Germany and France, and criticism that political resources are wasted on identity issues like immigration debates. Some compare the EU to the US without a strong federal government, pointing out that competition and regulatory differences among countries hinder AI progress.