AI agents use 5x more tokens than humans, headed for 10x

OpenRouter and a16z data show AI agents consumed 7.3 trillion tokens in August versus 1.4 trillion for humans, having first overtaken human usage in February. More than 85% of agent tokens come from cached prompts, driving demand for HBM and memory.

• Agents used 7.3 trillion tokens vs 1.4 trillion for humans in August
• Since February agent usage grew 14x, human usage 2.8x
• Over 85% of agent tokens are cached prompt re-reads
• 40% of large organizations scale AI agents, up from 27% a year earlier