a16z says agents are already using nearly 5x the tokens of human users
Andreessen Horowitz (a16z) said in its latest Charts of the Week that AI agents are not only growing fast, but are also consuming nearly five times as many tokens as typical human users. Since February, agent usage has risen by about 14x.
The firm said the shift could also create a hardware winner: high-bandwidth memory, or HBM. That is because agents rely heavily on cached tokens, which is very different from the usual one-prompt, one-response chat pattern.
Enterprise AI usage is splitting apart
Using OpenAI data, a16z said enterprise AI usage is rising, but not evenly. Since April 2025, output tokens for the average company have roughly doubled, while the top 10% of companies have grown by more than 17x.
That leaves a roughly 8x gap between the top 10% and the typical company. In information technology, the gap is even wider, at nearly 12x. In that sector, the top 10% are now generating 32.5x as many tokens as they were about a year ago.
a16z said the pattern suggests enterprise AI may be moving from a question of whether companies use AI at all to a question of how deeply AI has been built into workflows.
The firm said top-tier users are no longer treating AI like a chat bot. In its data, plugin adoption among the top 10% is about 2x that of typical companies, while skills adoption is about 6x.
One of the fastest-moving sectors is not tech, a16z said. Based on Codex adoption as a measure of advanced AI-tool usage, the legal sector has seen adoption jump 108x since February 2026. a16z also said it is still hard to tell how much of that came from Codex’s wider rollout.
Why cached tokens point to more memory demand
a16z cited data from OpenRouter and Peter Walker showing that only a small share of AI users have fully deployed agents, but those agents are already consuming huge amounts of tokens. Agent token usage is now close to 5x that of human users, and usage has increased about 14x since February.
That makes sense, the firm said, because agents work differently from normal chat tools such as ChatGPT and Claude. A user sends one prompt, but an agent may keep reading information, calling tools, writing results, and then looping through another round until the task is done.
OpenRouter data showed that more than 85% of agent token consumption comes from cached prompts, and most of the recent increase in agent token usage has also been driven by cached tokens. At launch, agents need to load a large amount of background material, including system rules, coding standards, company policies, and task context. After the initial prefill, they can reuse cached context instead of reloading everything on every inference step.
Cached tokens usually cost less than running a fresh prefill, which helps lower the inference cost of long-running agents. The tradeoff is higher memory demand. That is why a16z tied the rise of agents to strong HBM demand: as more agents run at the same time and keep reusing context, AI inference infrastructure needs not only GPU compute, but also large amounts of high-bandwidth memory.