AI is already using AI more than its human creators.
New data from OpenRouter, published by venture firm a16z, shows that AI agents are now burning through 7.3 trillion tokens on a seven-day average basis, roughly five times what human users consume through the same period. The line for agentic usage first crossed above human usage on February 6, 2026, and it has not looked back since, growing roughly 14x in about six months.
The chart itself tells a story that would have sounded far-fetched even a year ago. Back in September 2025, agentic token consumption barely registered, a thin flat line hugging zero while human and mixed usage moved along in their own steady bands. Agents caught up to humans by early February, overtook mixed usage soon after, and then broke away entirely through the spring and summer, climbing past 4.5 trillion tokens by June and touching nearly 7 trillion by mid-July before a brief dip and a fresh push to 7.3 trillion in the first week of August. Human token usage, by comparison, has grown at a far more modest pace, edging up from around 0.4 trillion tokens to about 1.5 trillion over the same window.
Why agents burn through so many more tokens
The gap comes down to how agents are built to work. A human typing into a chatbot tends to follow a prompt-and-response rhythm: ask a question, read an answer, ask another question. An agent set loose on a task does not stop after one exchange. It reads, writes, checks its own output, and loops back again, iterating toward a goal across dozens or hundreds of steps before it produces a final result.
That iteration is expensive in token terms, but a lot of the expense is now going toward a specific place. OpenRouter’s data indicates that more than 85 percent of agentic token consumption comes from the cached prompt, and cached tokens account for nearly all of the category’s relative growth. The initial prompt an agent receives, loaded with the context it needs to do its job, whether that is a codebase’s style guide, a company’s internal policies, or a broader task brief, is expensive to generate the first time. After that, the agent keeps reading and writing incrementally against that same context, adding to the cache with every cycle instead of rebuilding it from scratch. Cached tokens cost a fraction of what fresh tokens cost, which is part of why running agents in production has become economically workable at all. But cheap does not mean free of demands elsewhere. Every cached token an agent holds onto is memory that has to sit somewhere, and that is a big part of why high-bandwidth memory has become such a coveted resource in the current AI buildout. Agents that do not want to start over with each cycle need somewhere to keep what they have already figured out.
AI is quietly outpacing humans on several fronts at once
The token chart is a striking number, but it is not an isolated one. Pull back and a pattern shows up across a few different corners of the internet, all pointing the same direction.
Text is one of them. Data shows that AI-generated articles now make up roughly half of everything published on the web, having crossed that line sometime in late 2024. As recently as 2022, AI wrote less than a tenth of new web content. Within a couple of years it had gone from a rounding error to parity with human writers, and separate figures cited around the same period showed AI’s annual written output overtaking humanity’s for the first time in 2025.
Traffic followed a similar arc not long after. Cloudflare reported that bots now account for more than 57 percent of all HTTP requests hitting web content, pushing human traffic below the halfway mark for what the company says is the first time in the internet’s history, a shift that arrived faster than Cloudflare’s own leadership had predicted. Put together with the writing numbers, the picture that emerges is one where a growing share of what gets published online is machine-written, and a growing share of what reads it is a machine too.
None of this happened because anyone flipped a switch. It happened because the underlying capability crossed a threshold where letting a model handle the work, whether that work is drafting an article or running an agent loop, became cheaper and faster than routing it through a person. Anthropic’s Dario Amodei has argued that if AI keeps improving at its current pace for even another few years, systems will be able to work independently on complex tasks and surpass top human performance across most fields. This is already showing up in OpenRouter’s infrastructure billing.
What the trend actually means going forward
There is a practical reading of all this that matters more than the headline multiple. Token consumption is a reasonable proxy for how much actual work is being delegated to autonomous systems, and right now that work is scaling far faster than the number of humans typing prompts into a chat window. Companies building agentic products are not just serving more users, they are serving systems that never stop working once they start, which changes what infrastructure they need to plan around. Memory, in particular, stops being a nice-to-have and becomes the thing standing between an agent that can hold a long-running task in its head and one that has to keep restarting from scratch.
It also raises the obvious question of where the ceiling is. Agentic usage has already grown 14x since February, and the chart shows no sign of flattening. If a similar multiple plays out again over the next six months, the gap between what agents consume and what humans consume will not just widen, it will make human usage look like a rounding error next to it. Whether the rest of the AI stack, from chip supply to memory manufacturing to data center capacity, can keep pace with that curve is turning into one of the more consequential questions in the industry right now.