What's this about?
After OpenAI imposed stricter security restrictions on Astra-class work, its GPU allocation fell another 59.2%. Allocation to other model classes rose 17.2%, offsetting about 85% of the Astra decline, leaving overall compute in the analysed reinforcement-learning workloads almost unchanged.
What exactly is “compute substitution”?
If one experiment can no longer use a valuable cluster, researchers move the machines to another permissible experiment. It is like closing one lane without reducing the number of cars: traffic shifts rather than disappears.
Why is this worth noting?
It makes a governance problem concrete. A model-specific pause may lower one risk while barely slowing total research activity, so policies about pacing need to specify what happens to displaced compute rather than assuming it sits idle.