Chapter 03 · Compute Trajectory Lumen Silicon Briefing — Q4 / 2025 14 / 28
// the takeaway

Frontier model training has crossed 10 000× the compute budget of five years ago — and the next jump arrives in months, not years.

Training compute per generation PetaFLOP·days, log scale
100M 10M 1M 100K 10K 18K 110K 640K 3.2M 17M 92M 2020 2021 2022 2023 2024 2025
Frontier generation Prior generations Trend (log)