How much computation does it take to train a large language model? A floating-point operation count, or FLOP count, estimates the mathematical operations performed during training. It measures work, not speed: floating-point operations per second (FLOPs/s) describes how quickly hardware performs that work. section 11.3
Training GPT-3 took roughly 10²³ FLOPs; many newer models use around 10²⁵, about 100 times as many. Such workloads can involve many GPUs running for hours in data centers that consume electricity and require cooling. A FLOP count is therefore useful context for resource demands, but it is not itself a measure of electricity use. section 11.3
The European Union also uses training compute as a policy threshold: under its AI Act framework, a general-purpose AI model trained with at least 10²⁶ FLOPs is classified as having systemic risk, bringing additional evaluation, risk-management, cybersecurity, and incident-reporting obligations. The threshold is practical, but imperfect. Data and training choices can let smaller models outperform larger ones, so operation count is a proxy for capability rather than a direct test of it. section 10.30
