A count of the raw arithmetic a model consumes.
A FLOP is one floating-point operation. FLOPs count the arithmetic a training run or a request requires; FLOPS (per second) measures how fast a chip can do it.
Training runs are described in the 10^25–10^26 FLOP range, which is also the threshold regulators have started to use when defining models that require reporting.
Published peak FLOPS are a ceiling, not a promise — real workloads land well below it because data movement, not arithmetic, is the bottleneck.