Hardware that ships by the end of 2027 could host between roughly 30 million and 170 million frontier-model agents at the same moment, according to a modelled estimate in a report Epoch AI published on 2 October. Nobody has counted agents that exist. Jason Li’s analysis is a capacity calculation, and its most useful finding concerns demand, not hardware.
“Concurrent agents” means AI workers active at the same instant, each one a session like those run by Codex or Claude Code, where a model reasons, calls tools, and keeps going. Software does not sleep, so an agent can put in every one of the 168 hours in a week, 4.2 times a full-time schedule. Epoch converts that into weekly working hours equal to about 140 million to 720 million full-time employees. That is a very wide spread, and Epoch does not hide it. For scale, Epoch pegs US knowledge workers at about 100 million.
The estimate starts from memory, not processors. Epoch counts high-bandwidth memory, the scarce component made by Micron, Samsung, and SK hynix, because memory decides how many sessions fit on a chip at once. It converts shipments into units matching Nvidia’s GB300 and then layers on assumptions. Epoch uses a $30 price tag for each hour an agent is active, a $5 hourly rental price per GB300, and API prices set at five to ten times serving cost. It also assumes newer HBM4 chips hold twice as many agents per gigabyte as older ones. Vary the HBM4 gain from none to fourfold and the range stretches from about 33 million to 171 million.
Several conditions must hold for the top of that range to be real. All the hardware has to be installed and powered, and all of it has to go to this one workload, with models staying as demanding as today’s. Deployment is not guaranteed. Epoch notes that Satya Nadella said in late 2025 that Microsoft had chips in inventory it could not plug in for lack of powered data center space. Cheaper models would change the picture entirely: using DeepSeek V4 Pro serving benchmarks, Epoch calculates roughly 1.9 billion agents on the same hardware, though a weaker model does different work.
Now the demand side. “API-equivalent spending” is what the same work would cost if bought at standard per-token prices from the AI labs, whether or not anyone actually pays it. Epoch takes a deliberately modest case: 40 percent of the hardware earns revenue from inference, and that slice is busy half the time, so only 20 percent of capacity gets used. Even then, the implied bill is $2.6 trillion to $5.3 trillion a year after installation.
Set that against today’s books. Epoch counts over $100 billion in annualized revenue across the leading model developers. If that kept growing fivefold every year, the total would approach $1 trillion around the close of 2027, which Epoch labels an illustrative scenario, not a forecast. Even that generous path sits well below the spending implied by the 20 percent case: about $1 trillion in revenue against $2.6 trillion to $5.3 trillion in implied annual spending. Epoch itself calls the gap significant.
Falling prices complicate the math in both directions. Epoch estimates that the price of reaching a given benchmark score has dropped 47 percent per quarter since 2023. That lowers what the buildout must earn to pay for itself, but it also means each task needs fewer chips. Demand would have to grow fast enough to absorb those efficiency gains, and Epoch points to a Jevons effect, where cheaper work prompts people to delegate far more of it, as the hopeful case.
One more number frames the stakes. In Epoch’s agent traces, Codex workloads averaged about $16 to $18 per hour of continuous activity, and Claude Code workloads ran $24 to $50. Those are hourly rates in the neighborhood of human wages. For the capacity to be used, agents must do work that employers value at that price, across many industries, not only software.
Epoch states the risk plainly: spending on compute may outpace the demand for running it. The figures are not break-even requirements, and the authors do not claim the idle-chip outcome will happen. Still, the arithmetic leaves buyers in a stronger spot than sellers. Anyone signing a multi-year compute commitment or pricing an agent product should treat today’s rates as a ceiling that surplus capacity would push down.
Epoch AI, “How many AI agents could run on the AI chips shipped through 2027?” by Jason Li, published October 2, 2026.
