Epoch estimate: OpenAI spends roughly 1.2–2.4% of compute on safety monitoring

sjgadler · x · 2026-08-20

Josh You at Epoch did rough back-of-the-envelope math on OpenAI's recent safety-monitoring announcement, estimating it consumes roughly 1.2–2.4% of OpenAI's total compute budget.

His reasoning: about half of compute goes to R&D/internal use vs. external inference; a quarter to half of R&D compute is RL for frontier models (the rest is pre-training, sub-frontier, and early-stage work); and half of RL compute is inference. That yields about 6–12% of all compute covered, multiplied by a 20% monitoring overhead to arrive at the final range.

Original post →

More from Infra

Infra channel →