In August, agents on OpenRouter accounted for 7.3 trillion tokens. Human use on the same platform was 1.4 trillion. Daniel Newman, chief executive of Futurum Group, called that about five times, and said the multiple would move to ten and then higher.
Tom's Hardware's reading of the chart traces the numbers to an Andreessen Horowitz graphic of OpenRouter data. Agents first passed humans six months before that August snapshot. Peter Walker, OpenRouter's head of insights, published a seven-day average split by type. From the February crossover, agent tokens rose 14 times and human tokens rose 2.8 times. OpenRouter sorts each API key as agentic, mixed, or human, using a score that includes tool-call rate, how many turns a session runs, and the gaps between them.
Most of the agent tokens are a reread
Andreessen Horowitz, an OpenRouter investor, wrote that more than 85% of agent tokens come from cached prompts. The system is sending back text it has already processed. Cached tokens cost less than a fresh prompt and still have to sit in memory. The same note says the KV cache, the stored context, is outgrowing the high-bandwidth memory on the GPUs. A call-center firm that tested DeepSeek on rented Nvidia H200 chips saw the same shape in its own September logs for Claude Code: 96% of the input was rereading an old conversation.
The series is not a straight line. Tom's Hardware notes dips in April and July. The counts are tokens, not dollars, and they are one gateway. McKinsey's 2026 State of AI survey is a second window: 40% of respondents at large organizations said they were scaling agents, up from 27% a year earlier.
The memory has to hold the reread
Micron's last reported quarter is what that memory business looks like when the cache grows. Newman wrote that AI is used by AI five times more than by people, and that the lead accelerates to ten times and then keeps climbing. He also wrote 20 times and 30 times. Those later multiples are not on the August chart. The chart's job is the 7.3 and the 1.4, plus the cache share above 85%. OpenRouter is one gateway. A key labeled mixed, part agent and part person, is a third bucket the piece mentions and does not fully split. Token volume is not a bill. Fairphone has already put memory at 60% of its materials bill. A cached token is the cheap operation that still needs that memory. From the February crossover to the August snapshot is six months, which is the window in which agents went from newly ahead to 7.3 trillion against 1.4 trillion.
Cached tokens are cheaper than a prompt the model has never seen, and they still occupy memory. Andreessen Horowitz ties that storage to high-bandwidth memory and says the KV cache is growing past what GPU memory holds. The call-center log is a second measurement with a different percentage: 96% of Claude Code input in September, on a consultancy's own agents, was old conversation read again. They had been testing DeepSeek on rented H200 chips. OpenRouter's 85% and that 96% are two caches, two companies, the same shape. McKinsey's survey is about organizations, not tokens: 40% of large-company respondents scaling agents, from 27% the year before. Newman's 10x, 20x, and 30x are what he expects after the August table. The table itself stops at 7.3 trillion and 1.4 trillion, with dips in April and July that keep the line from being a smooth climb. Human use on that same chart rose 2.8 times after February, while agent use rose 14 times.
The mixed bucket is the one left unsplit
OpenRouter also keeps a mixed label for keys that look partly like an agent and partly like a person. Tom's Hardware does not publish how that bucket divides, so the 7.3 trillion and the 1.4 trillion are the two piles with a clean name.