Global compute cap

What it does

Countries would agree on a maximum amount of computing power any single AI training run may use, and both the U.S. and China would stay under it. Large data centers would be licensed and monitored so no lab could quietly go over the limit.

Applies to: Training runs above the agreed compute limit in any signatory country.

• A treaty capping how much computing power any training run may use, in the U.S., China and everywhere else.

Where things stand

Compute thresholds today only trigger reporting or extra obligations (for example under the EU AI Act), not bans. U.S. export controls limit China's chip access, which is a one-sided limit, not a mutual cap.

Why it costs ~8 months

A cap binds America hardest, because America has far more compute. China, with less to give up, loses less. Net: about two and a half months a year in.

Biggest unknown: Whether algorithmic efficiency gains would let actors reach dangerous capability while staying under any fixed cap.

Why it lowers p(doom) by ~0.75%

Keeps every country's AI models below a size where they could become uncontrollable.

Keeping everyone below a size limit removes much of the risk from the largest systems. Better algorithms keep making each unit of compute more capable, so the cap slowly weakens.

The strongest case that it costs more

A compute cap freezes America's biggest advantage. Algorithmic progress would keep going, so the cap would need constant tightening.

The debate

For

• ControlAI (Andrea Miotti, Tolga Bilge et al.) 'A Narrow Path' (Oct 2024): license any training run above 1e25 FLOP and any data center above 1e17 FLOP/s, then harmonize limits internationally by treaty.
• MIRI Technical Governance Team Draft treaty (Nov 2025) bans training above 1e24 FLOP and monitors clusters above 16 H100-equivalents.
• Al Ramiah, Meindertsma et al. 'Toward a Global Regime for Compute Governance: Building the Pause Button' (June 2025): restrict compute access via a governance, enforcement and verification framework.
• Eliezer Yudkowsky 2023: 'Put a ceiling on how much computing power anyone is allowed to use in training an AI system.'

Against

• Sara Hooker (then Cohere For AI) 'On the Limitations of Compute Thresholds as a Governance Strategy' (July 2024): thresholds are shortsighted because compute is a poor proxy for risk.
• Dario Amodei (Anthropic) Sept 2026: limits on inputs like training compute may be 'more gameable' than limits on external behavior.

Sources

• A Narrow Path, Phase 0: Specific FLOP and FLOP/s licensing thresholds.
• MIRI treaty summary (May 2026): Hard cap numbers and verification design.
• Epoch AI: How many AI models will exceed compute thresholds?: Compute growth about 4.6x per year; frontier runs 1e26 to 1e27 FLOP in 2026.
• Al Jazeera: China vs US, who is winning the AI race, in four charts (Sept 24, 2026): U.S. about 75% of AI compute vs China about 14% (Epoch); China about 8 months behind.