Lays out which physical constraint binds first on training scale, helping engineers reason about how fast frontier capability and cost can keep moving.
Key takeaways · AI-distilled
Epoch puts frontier training compute growth at about 4x per year, faster than the peak growth of mobile phone adoption (2x a year, 1980-1987), solar capacity installation (1.5x, 2001-2010) and human genome sequencing (3.3x, 2008-2015).
Epoch says a 2e29 FLOP run would exceed GPT-4 in scale by about as much as GPT-4 exceeded GPT-2, and that if labs pursue it, we might see advances by 2030 as large as the step from GPT-2 in 2019 to GPT-4 in 2023.
The projection folds in planned foundry and chip packaging expansions, utilities' capacity growth forecasts, and spreading data centers across multiple power networks, rather than assuming today's supply stays fixed.
Feasibility is not a forecast: Epoch says whether labs actually scale this far depends on their willingness to invest hundreds of billions of dollars over the coming years, which the report treats only briefly.
