GPU-hours and compute capacity
GPU-hours measure accelerator time used or sold; compute capacity is the total accelerator power a provider can deliver in a period.
One GPU running for one hour is one GPU-hour. Cloud providers and AI developers use the unit to price and plan training and inference. Compute capacity is the amount available to sell, roughly the number of accelerators multiplied by the hours they can run and the share of that time they are used.
Traders care because capacity limits can cap revenue growth for AI cloud providers, and utilization drives the return on AI capex. Companies also disclose capacity in megawatts or GPU counts, and those figures are not directly comparable across chip generations. Check how a company defines its unit before comparing.
Worked example (illustrative numbers, not real data): a provider has 10,000 GPUs running at 80 percent utilization over a 720-hour month. That is 5,760,000 GPU-hours. At an illustrative price of 2 dollars per GPU-hour, monthly revenue is 11.52 million dollars.