Dave White: a US-China AI training slowdown deal is largely enforceable via zk proofs

_Dave__White_ · x · 2026-10-29

Dave White offers a controversial take: negotiating an AI training slowdown with China would be largely enforceable, with four supporting arguments.
Key points:
We can track where the majority of current compute lives;
Much of the compute needed to train next year's models isn't built yet, creating a window;
Zero-knowledge proofs could demonstrably show most compute is used for inference rather than training (yes, really);
This lets both sides trust the other isn't racing ahead on next-gen pretraining, buying preparation time even as efficiency gains let "dark compute" train frontier models.
He invites pushback on any specific point.

More from Infra