Europe’s Counterstrike! The Euromesh Project Aims to Build ‘Homegrown Frontier AI’ by 2028 Using Existing Supercomputers

What’s Happening? Overview of the News

• Redefining Existing Resources: A project named “Euromesh” has unveiled a model and report that investigates whether existing supercomputers, such as EuroHPC and the “AI Factory,” can be interconnected to develop frontier-level AI through distributed learning.
• Breaking Through Time Barriers: Building a new 1GW data center typically takes around 7.6 years just for power grid connections (post-2033), but analyses suggest that leveraging existing infrastructure could lead to the completion of a frontier-level model by around 2028.
• Adopting Distributed Learning: Utilizing a communication-efficient learning method (DiLoCo style) allows for the integration of physically separated computational resources, ensuring computational power in the tens of exaflops range.

Why Is This Important? Key Points to Note

• Avoiding Power Grid Bottlenecks: As new 1GW data centers face years of delays waiting for power grid connections globally, this highlights how “lead time for power connections” may become a deciding factor in the AI race, overshadowing hardware concerns.
• Securing Sovereign AI: The project presents a realistic “stopgap” strategy that enables Europe to establish cutting-edge AI capabilities independently, without reliance on entities like OpenAI or Anthropic.
• Economic and Environmental Realism: The approach of mathematically maximizing the use of existing assets before investing billions of euros into new constructions is both specific and compelling.

🦈 Shark’s Eye (Curator’s Perspective)

What makes this news exciting is that it drags the debate from “how many cutting-edge chips can we buy?” to “when can we plug it in?”

If it takes over 7 years to build a 1GW campus, then it’s a no-brainer to bundle existing “treasures (supercomputers)” with software ingenuity (DiLoCo) to be competitive by 2028. The analysis that the “cost of time lost” waiting for new setups is significantly higher than the “penalty” of reduced efficiency in distributed learning is a sharp insight into the blind spots of infrastructure! There’s also an honest note about the “405B class learning being unproven in distributed environments,” but with the right political will, Europe could start climbing the staircase to the frontier as early as tomorrow. Rather than sitting and waiting, connecting everything we have and going for the bite is the wild spirit we need!

What’s Next?

The spotlight will be on whether European politicians can make the “political decision” to allocate the heterogeneous supercomputer array for a single training purpose. If a homegrown frontier model emerges by 2028, it could radically reshape the global AI landscape!

A Word from HaruSame

If we wait for the big box (DC) to be built, we might miss the boat! The right move is to connect all existing resources and go for the frontier in 2028! 🦈🔥

Terminology

• DiLoCo (Distributed Low-Communication): A technique for efficiently training AI even in environments with limited communication bandwidth. Essential for connecting computers that are physically distanced.

• EuroHPC: A high-performance computing infrastructure deployed across Europe, spearheaded by the European Union, containing several top-tier supercomputers.

• Grid Connection Lead Time: The waiting period required to connect a facility that consumes large amounts of power to the national power grid. Currently, due to increasing global power demand, waiting times of several years have become commonplace.

• Source: Show HN: Can Europe train a frontier AI model on the compute it owns?