Compute & Cost
FLOP
Floating Point Operations: the count of individual calculation steps. FLOP measures how much computing went into a training run; regulators use thresholds such as 10^25 FLOP to define “large” models. The number says nothing about intelligence but a lot about effort. It is the kilogram weight of the AI industry.
An example
FLOP counts computing operations, the unit in which training effort is stated. The orders of magnitude are beyond intuition: a current frontier model sits at around 10²⁶ operations. For comparison, a single good computer would have to work for longer than there are humans to do it. The number becomes practically relevant because it appears in regulation: the EU AI Act ties additional obligations to a threshold of 10²⁵ FLOP in training.
A closer look
FLOP counts floating-point operations. For models, the total number used in training is usually given, today in the range of 10²⁵ to 10²⁶. The EU AI Act ties additional obligations to such a threshold.
Why it matters
For the vast majority of companies, FLOP is a vendor metric, not their own: you don't train models of this size. Still, it is worth knowing why it shows up in contracts and statutes: it is currently the only halfway objective approximation for “particularly powerful model” and therefore serves as a trigger for review and reporting obligations. If you deploy a model above the threshold, the vendor has additional obligations, and you have one more question for the supplier conversation.
Commonly misunderstood
The number is no good as a measure of quality. Two models with the same computing effort can differ considerably in data quality and method.