Glossary term
Glossary term
Infrastructure and Serving
The computational resources used to train a model. Training compute can be a proxy for potential capability and may influence regulatory classification, internal escalation, and safety review depth. Where training compute is material, it should be visible to risk, security, finance, and leadership rather than treated only as an engineering resource.
EU AI Act Article 51 uses a 10^25 FLOP training compute threshold to designate GPAI models with systemic risk.
US Executive Order 14110 set a 10^26 FLOP threshold for dual-use foundation model reporting requirements.
Epoch AI maintains a public dataset of training compute estimates for leading models from 1950 through the present.