New data from OpenRouter, shared by Andreessen Horowitz and cited by Futurum Group CEO Daniel Newman, shows AI agents now consume about five times more tokens than humans on the platform. As of August, agents accounted for 7.3 trillion tokens versus 1.4 trillion for humans, six months after agent usage first surpassed human usage. Newman predicts the ratio will accelerate to 10x and beyond, arguing that ROI discussions focused only on human adoption miss the scale of machine-to-machine usage.
The catch is that most of those agent tokens are not fresh computation. A16z notes that over 85% of agent tokens come from cached prompts, and a call center consultancy testing DeepSeek on rented Nvidia H200s saw 96% of its Claude Code input re-reading old conversation. Cached tokens cost less to process than new prompts, but they still must be held in memory, and that demand is straining high-bandwidth memory (HBM) capacity as models' KV caches outgrow GPU memory.
The data comes from a single platform and measures token volume, not spending, with dips in April and July. Still, the broader trend appears consistent: McKinsey's 2026 State of AI survey found 40% of large-organization respondents scaling AI agents, up from 27% a year earlier. If agent token use keeps climbing, memory makers like Micron expect shortages to worsen in 2027 and 2028, leaving PC buyers competing with AI agents for scarce memory supply.