Epoch AI’s estimate makes the potential supply legible while foregrounding utilization as the unresolved constraint.
Observed facts: Epoch AI estimates that HBM shipped in 2025–2027 could support about 30–170 million concurrent frontier-model agents, assuming full deployment and allocation to those workloads. It says a more efficient model benchmark could support approximately 1.9 billion concurrent agents. Mechanism: The report derives concurrency from projected HBM supply and estimated serving capacity, with memory capacity and bandwidth shaping how many sessions can run and at what speed. It then converts continuous agent hours to human workweek equivalents and estimates spending at API-equivalent prices. These are conditional calculations, not counts of agents already deployed.
Why now: The report links a near-term hardware buildout to the scale of agent use that it could technically serve, sharpening a live infrastructure-versus-demand question. Implications: The implied capacity could be substantial, but its economic significance depends on actual workloads, affordability, and utilization; capacity alone does not establish labor substitution or revenue. Epoch AI estimates that using 20% of its central capacity would correspond to $2.6–5.3 trillion in annual API-equivalent spending after deployment, against roughly $1 trillion in annualized developer revenue by end-2027 under its stated fivefold-growth comparison.
Evidence boundary: These are model-based estimates conditional on hardware shipment, deployment, allocation, serving, and pricing assumptions; they are not observed adoption, realized spending, or a labor-market forecast. Watch condition: Compare subsequent HBM deployments and measured agent utilization with the report’s assumptions, especially whether demand can absorb capacity at the implied API-equivalent prices. This begins a distinct analytical record because no prior New Runtime record is supplied; related coverage is not evidence of an existing record to extend.