June 26, 2026
The Token Economy Is Eating the Internet
This post is part of my Medium blog.
Something is happening to information that we don't have a word for yet. Every metric we use to measure the internet — page views, unique visitors, click-through rates, engagement time — was designed for a world where humans typed things and other humans read things. That world is ending, and the signal is in a metric most people have never looked at: token volume.
A token is roughly three-quarters of a word. It's the unit of currency for AI models — the thing you pay for when you call an API, the thing that gets counted when a model processes a prompt and generates a response. Token volume is the fossil record of AI activity. And right now, that fossil record looks like the Precambrian explosion — a sudden, violent proliferation of new forms of life, except the life forms are conversations between machines.
China's daily AI token consumption reached 180 trillion tokens per day in February 2026. Two years earlier, it was 100 billion. That's a 1,800x increase in 24 months. ByteDance's Doubao model alone processes over 50 trillion tokens daily — one app, one company, one country. Google processes 1.3 quadrillion tokens per month across its surfaces — Search, AI Overviews, Workspace, Gemini. That's 43 trillion tokens per day, and it doubled twice in the five months before they stopped reporting the number. OpenAI's API serves roughly 8.6 trillion tokens per day, and that was as of October 2025 — seven months ago. Fireworks AI, an inference platform most people haven't heard of, processes 15 trillion tokens daily.
I'm not listing these numbers to alarm you. I'm listing them because the volume is the story. The way information is generated, consumed, and exchanged is changing in ways that will become apparent only in retrospect. Right now, if you look at token volume, you see an explosion in a kind of virtual conversation — machines talking to machines, agents calling APIs, models generating text that other models will ingest and process. It's happening below the surface of the human-readable web, and it's growing at a rate that makes the early internet look slow.
The shape of the curve
OpenRouter, a routing layer that connects applications to dozens of LLM providers, is the closest thing we have to a public window into token volume. In mid-2025, they processed about 100 trillion tokens in a year. By May 2026, their weekly volume had grown from 5 trillion to 25 trillion tokens — a 5x increase in six months. They're on pace to process well over a quadrillion tokens in 2026. One platform. A routing layer. Not even a model provider.
The Alibaba CEO said it plainly in September 2025: "Token consumption speed doubles every two or three months." That's not a prediction. That's an observation of what's already happening. If token volume doubles every three months, it grows 16x in a year. In two years, 256x. China went from 100 billion daily tokens to 180 trillion in two years — a 1,800x increase. The math checks out.
Microsoft reported over 100 trillion tokens processed in a single quarter, up 5x year-over-year, with 50 trillion in one month alone. That was April 2025 — fourteen months ago. OpenAI named 30 customers who each processed more than a trillion tokens through its models. Not 30 customers total — 30 customers who each individually crossed a trillion tokens. Shopify, Salesforce, Duolingo, Canva — companies that used to measure their API usage in millions of tokens are now in the trillions.
And video is coming. Seedance 2.0, a video generation model, consumes roughly 350,000 tokens to produce a single 10-second 1080p video. A typical animated project runs into the hundreds of millions of tokens. Video AI hasn't fully kicked in yet. When it does, the volume graph will bend again.
What tokens become
Tokens don't disappear. They become things. They become bug reports filed against open source projects. They become AI-generated pull requests. They become crawler requests hitting websites. They become summaries, articles, comments, reviews, and issues. Every token generated by an AI model is an output that lands somewhere — in a GitHub queue, in a web server log, in a maintainer's inbox.
The slop flood I wrote about in the first post in this series — the 241% increase in pull requests at curl, the 95% invalid rate on bug bounty submissions — that's token volume made visible. Each AI-generated bug report is a few thousand tokens. Each AI-generated pull request is a few thousand more. Multiply that across millions of users running AI coding tools, and you get the numbers I pulled from GitHub. The token economy doesn't just produce text. It produces labor for the humans who have to process that text.
The web traffic shift I wrote about in the second post — AI bots exceeding 50% of internet traffic, training crawlers consuming 80% of AI bot activity — that's also token volume. Every page an AI crawler fetches is training data. Every training run is trillions of tokens. The crawl-to-click gap, where Anthropic crawls 38,000 pages for every one visitor it sends back, exists because the tokens flow in one direction. The AI companies take content, convert it to tokens, train models, and generate more tokens. The humans who created the original content get nothing.
The feedback loop
Here's the part that should make you uncomfortable. The token economy is becoming a closed loop, and humans are not in it.
AI models generate tokens. Those tokens become bug reports, pull requests, articles, comments, and web pages. Other AI models crawl those web pages, ingest those tokens, and train on them. The next generation of models generates more tokens, which become more content, which gets crawled again. Each cycle produces more volume. Each cycle reduces the proportion of human-generated content in the training data.
OpenRouter's data already shows the shift. Chinese models surpassed US models in weekly token consumption for the first time in February 2026 — 4.12 trillion vs 2.94 trillion. Four of the top five models by usage were Chinese. These aren't consumers chatting with ChatGPT. These are developers and agents calling APIs, building applications, running workflows. The token economy is global, and it's growing fastest in the places where models are cheapest.
Nobody is scaling the human side. There are not 1,800x more maintainers reviewing pull requests. There are not 1,800x more readers clicking through to websites. There are not 1,800x more humans creating original content. The human capacity to process, review, and verify AI output is fixed. The token volume is not.
The cost nobody counts
When AI companies report their metrics, they talk about tokens served, parameters trained, benchmarks achieved. They don't report the externalized costs: the maintainer hours spent dismissing AI-generated bug reports, the bandwidth costs borne by websites serving AI crawlers, the decline in referral traffic to content creators, the pollution of issue trackers and pull request queues.
These costs are real, and they're growing at the same rate as token volume. Every trillion tokens generated creates a downstream burden that someone has to carry. Open source maintainers carry it in review time. Website operators carry it in bandwidth. Content creators carry it in lost traffic. The AI companies that generate the tokens carry none of it.
This is the same pattern in every domain AI touches. The benefits — speed, scale, automation — accrue to the company running the model. The costs — review, verification, cleanup, maintenance — are externalized to whoever is downstream. In open source, that's maintainers. On the web, that's publishers. In the token economy, that's everyone who has to interact with AI-generated output.
The token numbers are not abstract. 180 trillion tokens per day in China. 25 trillion per week on OpenRouter. 15 trillion per day on Fireworks AI. Those tokens become the bug reports that killed curl's bug bounty. They become the crawler traffic that exceeds human traffic on the web. They become the pull requests that GitHub had to build throttles for. The tokens are the input. The slop is the output. The humans are the cleanup crew.
The question isn't whether token volume will keep growing. It will — the curve says it doubles every three months. The question is whether the humans who carry the cost of that growth have any say in how fast it happens. Right now, they don't. And the Precambrian explosion doesn't care.
In Essential, the third book in The Condition Set trilogy, the law that requires a human to remain in the decision loop is ninety days from expiring. Rob Coleman runs the agency that oversees every AI system in Canada. The question isn't whether the machines work. The question is whether anyone can tell when they stop working in the public's interest.