GPU counts describe how many chips there are. GPU-hours combine chip count and time: eight GPUs used for ten hours equal 80 GPU-hours. FLOPS means floating-point operations per second, a computing rate; FLOP counts describe a quantity of work. These measures answer different questions about hardware capacity, time used, and computation. None can be substituted for another without additional assumptions.
How to read this finding
An hour on one GPU generation is not necessarily equivalent to an hour on another. FLOPS comparisons must specify numerical precision and whether the number is a theoretical peak, a sparse-workload figure, or achieved performance. Memory limits, communication, software, and workload affect what a system actually delivers. The Compute Index’s cluster charts count GPUs, rather than measuring GPU-hours consumed or achieved FLOPS. Do not infer a performance ranking or training-cost estimate from chip count alone.
Citation estimates use inputs through September 1, 2026; fleet counts have an October 1, 2026 cutoff. The research-topic sample covers January 1 to June 1, 2025. See each answer for its period and limitations.
Benaich, Nathan. “How do GPU counts, GPU-hours, and FLOPS differ?” State of AI Report Compute Index. Web page updated 2026-10-11; data periods as specified above.
