Epoch AI data suggests chip stock could support over a billion AI agents

The estimates rest on one deceptively simple figure. Epoch AI, a nonprofit that tracks AI progress, reported that approximately 27.6 million H100-equivalent chips had been sold as of June 2026. How you get from chips to workers H100-equivalent, or H100e, is a standardized unit for counting AI computing power. Once you know how much total compute exists, the next question is how much compute one AI agent needs to do the work of one person. That requirement is measured in FLOP/s, or floating-point operations per second. A September 2026 assessment published on LessWrong took Epoch’s chip figures and ran the math across different efficiency assumptions. Depending on the FLOP/s needed for human-equivalent work, the current and expected hardware stock could support anywhere from 24 million to 24 billion AI workers. The same analysis indicated that the projected cost of running these agents could come in significantly lower than human labor. The growth curve The analysis forecasts the stock of H100-equivalents to reach about 46 million by the end of 2026. At one level of efficiency, that hardware could potentially support close to 410 million AI workers. By the close of 2028, the install...