Regulating by how much computing power was used to train a model, on the grounds that compute is measurable, concentrated in a few supply chains, and correlates roughly with capability. It is the most tractable lever available and a contested proxy for what anyone actually cares about.
Technical view
Thresholds expressed in floating point operations appear in several regulatory instruments as triggers for additional obligations. Advantages are measurability and the physical concentration of supply, enabling monitoring at chip manufacture and data centre level. Objections are that the relationship between compute and capability shifts with algorithmic and data efficiency, that inference-time compute is not captured, and that fixed thresholds age quickly.
Related
Part of AI Terminology Index, section: Governance, Law and Regulation.