Epoch AI estimates hardware shipped through 2027 could run 33–171 million concurrent frontier-model agents. Efficient open models could support billions of agents on the same hardware. Using 20% of this capacity would imply $2.6–5.3 trillion in annual API-equivalent spending. Agent-hour costs vary from $16–18 for Codex workloads to $24–50 for Claude Code. The analysis uses HBM supply, serving benchmarks, and agent-hour costs. It assumes full deployment and allocation to agent workloads. The study compares agent capacity to human working hours. It highlights potential overabundance of AI capacity relative to demand. The figures come from an Epoch AI report.

• 33–171 million concurrent frontier-model agents
• 1.9 billion concurrent agents with efficient open models
• $2.6–5.3 trillion annual API-equivalent spending at 20% capacity
• $16–18 per hour for Codex workloads