How many AI agents could the world's chips actually run?

Epoch AI set out to answer a concrete question: given the AI chips shipping between 2025 and 2027, how many agents could actually run at the same time? The estimate is wide but striking. The high-bandwidth memory shipped over that period could support 30 to 170 million frontier-model agents running concurrently by 2027, up from 16 to 56 million in 2026. Swap in more efficient open models and the same hardware could run billions. Measured as labor, that pool could do roughly as much work in a year as 140 to 720 million full-time employees.

The interesting part is the gap between what the hardware can do and what anyone will pay for. Running these agents at just 20% utilization would mean $2.6 to $5.3 trillion in annual API spending by 2027. Developer revenue, by contrast, is on track to reach about $1 trillion by the end of 2027 even at fivefold yearly growth. So the chips can host far more agent-hours than the market looks ready to buy, which is where the oversupply risk comes from.

The figures rest on explicit methods, which Epoch lays out in the full report. They count hardware supply in gigabytes of HBM standardized to 288 GB units, infer closed-model serving capacity from TraceLab agent-session data and API prices, and take open-model capacity from SemiAnalysis benchmarks. They put the cost of an agent-hour at $16 to $50, depending on the model and the task.

Why it matters

If you are sizing infrastructure bets or planning large agent deployments, the binding constraint is demand, not silicon. The hardware ceiling already sits well above what current spending can fill, so the real question is whether paying customers show up, not whether the chips exist.