The AI memory chips shipped between 2025 and 2027 could keep roughly 30 to 170 million frontier-model agents running at the same time — or about 1.9 billion agents running an efficient open model. That is the headline estimate from a new report by Epoch AI, the research group that tracks AI computing capacity, published on October 2.
The report, by Epoch AI researcher Jason Li, treats the AI hardware buildout as a supply question: how much usable memory is shipping, and how many agent sessions can each unit of memory serve. Epoch converts shipments of high-bandwidth memory into equivalents of an NVIDIA GB300 accelerator, arriving at about 60 million GB300-equivalents through 2027 under its central assumption.
Two very different agents
The wide gap between the two figures comes down to which model each agent is running. For closed frontier models — the kind behind tools like Claude Code or Codex — Epoch cannot see inside providers' servers, so it works backward from cost. Measuring real coding-agent traces, it estimates spending of about $30 per continuous agent-hour and assumes providers charge five to ten times their serving cost, against a GPU rental price of roughly $5 an hour. That yields about one to two agent sessions per GB300, or 30 to 170 million simultaneous agents overall (50 to 101 million under the central hardware assumption).
For open models, Epoch uses direct measurements from SemiAnalysis's AgentX benchmark, which replays real coding-agent sessions. DeepSeek V4 Pro managed about 31 sessions per GB300 at 50 tokens per second per user — translating to roughly 1.9 billion simultaneous sessions, falling to about 870 million at 100 tokens per second.
The caveat: demand, not chips
Epoch stresses this is a capacity ceiling, not a forecast. It assumes every shipped chip is installed, powered and dedicated to agent workloads around the clock, while data-centre construction can lag chip deliveries by a long way. The 2026 and 2027 memory shipments are projections, and the closed-model numbers rest on assumed pricing markups.
"The key uncertainty is whether sustained, rapid growth in demand for AI services will justify the investment," Epoch writes. Even at only a fifth of effective use, the capacity would imply $2.6 trillion to $5.3 trillion a year in API-equivalent spending, against roughly $1 trillion in annualised developer revenue by the end of 2027. Code and data for the analysis have been released in a public reproduction repository.
Sources: Ground Truth, Epoch AI
The report, by Epoch AI researcher Jason Li, treats the AI hardware buildout as a supply question: how much usable memory is shipping, and how many agent sessions can each unit of memory serve. Epoch converts shipments of high-bandwidth memory into equivalents of an NVIDIA GB300 accelerator, arriving at about 60 million GB300-equivalents through 2027 under its central assumption.
Two very different agents
The wide gap between the two figures comes down to which model each agent is running. For closed frontier models — the kind behind tools like Claude Code or Codex — Epoch cannot see inside providers' servers, so it works backward from cost. Measuring real coding-agent traces, it estimates spending of about $30 per continuous agent-hour and assumes providers charge five to ten times their serving cost, against a GPU rental price of roughly $5 an hour. That yields about one to two agent sessions per GB300, or 30 to 170 million simultaneous agents overall (50 to 101 million under the central hardware assumption).
For open models, Epoch uses direct measurements from SemiAnalysis's AgentX benchmark, which replays real coding-agent sessions. DeepSeek V4 Pro managed about 31 sessions per GB300 at 50 tokens per second per user — translating to roughly 1.9 billion simultaneous sessions, falling to about 870 million at 100 tokens per second.
The caveat: demand, not chips
Epoch stresses this is a capacity ceiling, not a forecast. It assumes every shipped chip is installed, powered and dedicated to agent workloads around the clock, while data-centre construction can lag chip deliveries by a long way. The 2026 and 2027 memory shipments are projections, and the closed-model numbers rest on assumed pricing markups.
"The key uncertainty is whether sustained, rapid growth in demand for AI services will justify the investment," Epoch writes. Even at only a fifth of effective use, the capacity would imply $2.6 trillion to $5.3 trillion a year in API-equivalent spending, against roughly $1 trillion in annualised developer revenue by the end of 2027. Code and data for the analysis have been released in a public reproduction repository.
Sources: Ground Truth, Epoch AI
