AI agents are moving from demos into always-on workloads, but how many could run at the same time? In a report published on October 2, 2026, titled "How many AI agents could run on the AI chips shipped through 2027?", Epoch AI researcher Jason Li gives a supply-side answer: based on high-bandwidth memory (HBM) shipped from 2025 through 2027, the world's hardware could support roughly 30 million to 170 million concurrent frontier-model agents. Counting only shipments through 2026, the range is 16 million to 56 million.
How the estimate works: HBM is the constraint
The method is deliberately narrow. Epoch uses HBM shipments as the binding constraint and assumes the hardware is fully deployed and devoted entirely to this kind of workload. In other words, it answers "how many could the hardware run at most," not "how many the market will actually need." Running costs also vary widely by workload: in Epoch's analysis of agent trajectories, Codex workloads averaged about $16–18 per hour, while Claude Code workloads ran about $24–50 per hour, so the same hardware can produce very different bills and results depending on what it runs.
Converted into labor: striking hours, with a caveat
An agent can work 168 hours a week, 4.2 times the 40 hours of a full-time employee. Converted into weekly hours, the concurrency above — including 2027 shipments — equals roughly 140 million to 720 million full-time employees. For reference, the U.S. population is 342 million, with about 100 million knowledge workers. Switch to a more efficient model and the numbers climb further: using DeepSeek V4 Pro serving benchmarks, the same hardware could run about 1.9 billion concurrent agents, whose weekly hours would equal those of 8 billion people. Epoch is careful to note, however, that output quality varies and the comparison counts hours only — runtime is not the same as work of equal quality.
Can the money keep up? Supply may outrun demand
Behind the expansion of compute is ever-larger infrastructure spending, and the financing behind it keeps growing in scale, as explored in Broadcom's $42 billion loan to Anthropic: from backer to biggest customer, how deep does the compute business run?. Epoch then does the math on the demand side: using just 20% of its central capacity estimate, priced at API rates, would already imply $2.6 trillion to $5.3 trillion in annual spending — while model developers' revenue, even growing fivefold every year, would reach only about $1 trillion annualized by the end of 2027. Demand may not keep up with supply, leaving a risk of overcapacity. Deployment also does not have to mean large data centers: agents can also run on local hardware, with desktop-class machines becoming another path, as in DGX Spark 64GB edition goes on sale October 23: $4,999 to bring large models back to the desktop. The figures are therefore best read as a ceiling: they mark the boundary of what hardware supply could allow, not a prophecy of future demand.
