AI agents are consuming five times more tokens than humans

AI agents now consume roughly 5x more tokens than human users, while agent usage has grown 14x since February, according to OpenRouter data cited by a16z.
More than 85% of agentic token consumption reportedly comes from cached prompts as autonomous systems repeatedly iterate toward goals instead of making one-off requests.
That changes the infrastructure economics of AI.
Agents running continuously can generate far more compute, memory and inference demand than traditional chatbot usage.
The trend helps explain why memory demand is accelerating and why AI infrastructure providers are gaining greater pricing power.