The total amount of calculation used to train or run a model, and the industry’s main unit of currency. Training runs are routinely described by how much compute they consumed, because it correlates closely with both cost and capability.
Technical view
Usually measured in floating point operations (‘FLOPs’), with training runs of frontier models now in the range of ten to the power of twenty-five and above. Compute thresholds have started appearing in regulation as a rough proxy for model capability, which is contested, since the relationship between compute and capability depends heavily on data quality and architecture.
Related
Part of AI Terminology Index, section: Hardware, Compute and Training Mechanics.