According to data from the AI model gateway OpenRouter, AI agents currently consume 14 times more tokens than human users. This trend follows a crossover point observed in February. The platform classifies API keys into agentic, mixed, and human categories based on a weighted composite score that evaluates factors like tool call rates and turn counts.
The McKinsey 2026 State of AI survey indicates that 40% of large organizations are scaling AI agents, compared to 27% in the previous year. While token usage is rising, a significant portion of this volume is attributed to cached tokens, which involve models repeatedly reading existing conversation history. For instance, testing on Claude Code showed that 96% of input consisted of re-reading old data.
The rapid growth in cached tokens is driving demand for high-bandwidth memory (HBM). Reports suggest that the KV cache is currently outpacing available GPU HBM capacity. This trend has implications for the broader hardware market, as memory manufacturers prioritize HBM for AI data centers over other segments.
Micron anticipates that shortages in RAM and storage will intensify throughout 2027 and 2028. As AI agents continue to increase their token consumption, experts predict that PC buyers may face heightened competition for memory resources, potentially leading to increased costs for consumers.
