Inference‑only chips enforce AI pause

Updated: 2026.11.19 3H ago 1 sources
The article describes a proposal to halt frontier model training by mandating production and use of chips that can perform inference (running existing models) but cannot efficiently perform training. The idea leverages the concentrated chip supply chain and data‑center economics to make a software policy durable by changing hardware defaults. — If feasible, this reframes AI governance: from slow, contested lawmaking to technical design and industrial policy that can be enforced via production lines, export controls, or procurement rules with broad economic and security effects.

Sources

There is Still Time to Stall the AI Apocalypse
Kristen French 2026.11.19 100% relevant
A working paper by researchers at UC Berkeley, Stanford, Oxford and others proposing a 10‑year pause enforced by replacing 'training‑capable' chips (e.g., Nvidia GPUs used to train GPT‑6 Astra at the Stargate facility) with inference‑only chips.
← Back to all ideas