Epoch AI The chip sales data tracked points to a possible scale of digital labor: within a few years, the number of AI agents may grow from hundreds of millions to more than 1 billion.
This estimate is built on a seemingly simple number. Epoch AI, a nonprofit organization that tracks AI progress, reports that as of June 2026, approximately 27.6 million H100 equivalent (H100-equivalent) chips have been sold worldwide.
The logic of extrapolating from chips to "workers"
H100 equivalent (H100e) is a standardized unit used to measure AI compute.
After grasping the total scale of compute, the next question is: how much compute does an AI agent need to complete work equivalent to one person. This demand is measured in FLOP/s (floating-point operations per second).
An assessment published on LessWrong in September 2026 cited Epoch's chip data and carried out calculations under different energy-efficiency assumptions. Depending on the FLOP/s required to complete human-equivalent work, the current and expected hardware stock could support between 24 million and 24 billion AI workers.
The same analysis also pointed out that the projected cost of running these agents may be significantly lower than the cost of human labor.
The analysis predicts that by the end of 2026, the stock of H100-equivalent chips will reach about 46 million.
At a certain level of energy efficiency, this hardware is expected to support close to 410 million AI workers.
By the end of 2028, the installed stock is expected to swell to about 140 million H100e. Under reasonable energy-efficiency scenarios, this could correspond to about 1.2 billion AI agents.
Infrastructure and the distribution of compute
Epoch AI's GATE model points to rapidly growing infrastructure capable of running AI inference. As of September 2026, Epoch's data center tracking had covered 86 AI facilities. These sites house about 13.9 million H100e GPUs, with a combined power capacity of 13.3 GW.
Epoch estimates that this is roughly equivalent to 44% to 46% of global AI compute.
Background: various forecasts in the agent era
Research firm IDC expects that by 2029 there will be 1 billion active AI agents.
Bill Gates predicts that, beginning around 2026, there will be a major shift toward capable autonomous agents.
Another variable to watch is energy efficiency. The huge range between 24 million and 24 billion potential workers depends almost entirely on how much compute each agent needs.
The figure of 13.3 GW for Epoch's tracked facilities reminds people that compute is ultimately an energy issue. To scale to about 140 million H100e, the required power capacity would be far higher than that of the currently tracked sites.
The next update of Epoch's chip and data center trackers will help narrow this picture. What is worth watching: whether the installed stock can maintain its pace and reach the forecast level of about 46 million H100e by the end of 2026; and whether the coverage of global compute can break through the current 44% to 46%.
