U.S. compute cap
Hard cap on training compute. No training run above a fixed ceiling set at about today's largest runs, and the ceiling never rises.
What it does
No training run may use more computing power than today's largest runs, and the limit never rises.
Applies to: All training runs above the cap (say 10^27 FLOP).
- Prohibit training runs above a fixed compute threshold.
- The threshold does not rise with time.
Where things stand
No jurisdiction has a hard cap. Thresholds in SB 53 and the EU AI Act trigger obligations, not prohibitions.
Why it costs ~13 months
A cap freezes the size of training runs while better algorithms keep coming. A year in, the U.S. frontier is about four months behind where it would have been.
Biggest unknown: How much frontier progress comes from compute scaling versus algorithms and post-training.
Why it lowers p(doom) by ~0.4%
Stops anyone in the U.S. training models much more powerful than today's.
Capping scale directly limits how capable U.S. models get. But it doesn't bind China and pushes progress into better algorithms, so the effect is uncertain.
The strongest case that it costs more
Compute is the input the U.S. leads on. Epoch and RAND put the U.S. hardware lead at about four years and its compute capacity at roughly ten times China's. A cap surrenders the one advantage that export controls protect. The lead cost could be the entire lead within two years.
The debate
For
- ControlAI, A Narrow Path proposes compute limits and a ban on superintelligence for 20 years, 2024.
- MIRI Technical Governance Team drafted an international agreement banning training runs above 10^24 FLOP, 2025.
- Eliezer Yudkowsky and Nate Soares called for treaty monitoring of any cluster larger than eight top chips, 2025.
- Statement on Superintelligence Hinton, Bengio, Wozniak, Bannon, Susan Rice and others called for a ban on superintelligence, 2025.
Against
- Daniel Castro, ITIF called banning superintelligence a historic mistake, 2025.
- Neil Chilson, Abundance Institute called the treaty and GPU proposals astoundingly casual authoritarianism, 2026.
Sources
- Epoch AI, U.S. export controls and China's AI progress: hardware lead of about four years
- Epoch AI, Algorithmic progress in language models: 60–95% of gains from compute and data
- Epoch AI, Training compute of frontier AI models grows by 4-5x per year
- RAND (Heim), China's AI models are closing the gap: roughly ten times the compute capacity
Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.