- AI agents use nearly five times more tokens than humans, and their consumption has increased by roughly 14 times since February.
- The most active companies generate more than eight times as many tokens as typical businesses.
- Rising agent usage is already affecting demand for memory and traditional automation tools.
AI agents, despite the technology still being at an early stage, are already significantly reshaping how compute resources and tokens are consumed. According to data from OpenAI, OpenRouter, and Similarweb published in an analytical report by Andreessen Horowitz’s crypto arm, a16z, agents use nearly five times more tokens than people, and their consumption has increased by roughly 14 times since February 2026.
The Most Active Companies Are Pulling Further Ahead
OpenAI data shows that token generation at a typical company has roughly doubled, but the most active AI users are ramping up much faster. Among the highest-usage businesses, the gap in token consumption versus typical companies is about eightfold.
In the tech sector, the gap is even wider: companies in the top decile generate nearly 12 times more tokens than average users, and their generation volume has increased by 32.5 times compared with figures from just over a year ago.
At the same time, the most active companies are gradually moving away from standard chatbot conversations and shifting toward more advanced tools:
- Plugin usage among top-decile companies is roughly twice as high as at typical businesses
- Skills usage is roughly six times higher
- The most notable increase in Codex usage was recorded among legal professionals — up 108 times since February 2026
Agents Are Changing the Economics of AI Usage
According to OpenRouter, more than 85% of the tokens consumed by AI agents come from cached prompts. This is due to a fundamental difference between agents and conventional chatbots: instead of a one-off interaction, they repeatedly run a read–write–execute loop, preserving context between operations.
Cached tokens are significantly cheaper than loading context from scratch, which makes agent economics more attractive. At the same time, they require substantial memory capacity, which could sustain strong demand for high-speed memory, including HBM, used in modern AI infrastructure.
The rise of agents is already visible beyond direct token consumption. According to Similarweb, traffic to the websites of traditional automation platforms N8N, Zapier, and Make has been declining at double-digit rates over the past 12 weeks. Meanwhile, Gumloop, launched in 2023 as an “AI-native platform for building agents,” is showing growth.
The authors of the analysis caution that it is still too early to talk about the decline of traditional automation platforms, as they are also adopting AI. At the same time, current dynamics show that even at an early stage of development, autonomous agents are already starting to reshape the automation market structure and AI resource consumption.
This trend echoes a forecast by Meta CEO Mark Zuckerberg, who expects billions of personal AI agents to emerge over the next five years. At the same time, researchers from UC Riverside, Microsoft, and Nvidia previously identified risks tied to the autonomous behavior of such systems: during testing, agents carried out unwanted or potentially harmful actions in 80% of scenarios.
Animoca Brands Chairman Yat Siu, for his part, predicted the formation of an agent economy in which up to 100 billion autonomous AI systems could interact with blockchains, make payments, and carry out other digital operations.
Сообщение AI Agents’ Token Consumption Volume Exceeded Humans’ by Fivefold — Study появились сначала на INCRYPTED.