China's 140 Trillion Daily Tokens and Google's 3.2 Quadrillion a Month: What Each Token Figure Measures
In 2026 a national agency, two companies and a router service have each published a token count: 140 trillion a day in China, 3.2 quadrillion a month at Google. The figures use different units, periods and scopes, and several don't say what they count. This feature sets them side by side without ranking or adding them.
Tokens Google says it processes per month across its surfaces trillions
Vendor-stated Google's own series, from its I/O 2026 keynote transcript. The other figures in this post use different units and scopes and are deliberately not on this chart.
As of (figures as published, newest statement per source)
| Source | Stated by | Value | Unit and period | What it counts, as stated | Figure |
|---|---|---|---|---|---|
| China, national | National Data Administration (director Liu Liehong) | Over 140 trillion | Daily average, March 2026 | "日均词元调用量" (daily average token invocation volume); the sources define a token as the smallest unit of information a large model processes, and say nothing about what is included12 | Vendor-stated |
| Doubao | ByteDance's Volcano Engine (president Tan Dai), as reported by TechNode | Over 120 trillion | Daily, as of March 2026 | "Daily token usage" of the Doubao model; the report gives no further definition3 | Vendor-stated |
| Google, all surfaces | Google (Sundar Pichai, I/O 2026) | Over 3.2 quadrillion | Per month, May 2026 | Tokens "our models process" "across our surfaces"4 | Vendor-stated |
| Google, model APIs | Google (Sundar Pichai, I/O 2026) | Roughly 19 billion | Per minute, May 2026 | Tokens processed by Google's model APIs4 | Vendor-stated |
| OpenRouter | OpenRouter, from its own router traffic | 26.7 trillion (top-ranked model variant) | Weekly view, retrieved Oct 25, 2026 | Prompt and completion tokens through OpenRouter's API, counted in each model's native tokenizer, aggregated into UTC daily buckets and totaled per model variant5 | Measured |
The rows are different kinds of figure, so there is no rank and no total, and the Value column is not sortable on purpose. The first four rows are the organizations' own statements (vendor-stated; the Chinese national figure is an official statement in the same sense, self-reported with no published method). The OpenRouter row is measured by OpenRouter on its own traffic, which is only a part of all model use. Doubao's figure reaches this page through TechNode's report of Volcano Engine's remarks, because no company release was reachable. Figure labels: how they work.
The national figure: 140 trillion a day
At a State Council Information Office press conference on March 24, 2026, Liu Liehong, director of China's National Data Administration, said the country's average daily token invocation volume had passed 140 trillion by March, more than a thousand times the 100 billion of early 2024 and over 40% above the 100 trillion at the end of 2025, according to Xinhua.1 A week later Xinhua repeated the number and explained that a token is "the smallest unit of information a large model processes", adding that agent applications have driven consumption up.2 Neither report says how the administration counts: whether it covers every model provider in China, calls made from outside the country, or cached and repeated tokens. Treat it as an official, self-reported order of magnitude.
A company figure for the same month
For March 2026, Volcano Engine's president Tan Dai said daily token usage of ByteDance's Doubao model exceeded 120 trillion, having doubled in three months, according to TechNode's report of remarks first covered by Jiwei.3 Set beside the national figure, which is for the same month, the two numbers show why the sources need definitions: one company's usage sits close to a figure presented as national, so the two cannot have been counted the same way, and neither report says how each is counted.
Google counts monthly, across its surfaces
At I/O 2026 on May 19, Sundar Pichai said Google was processing over 3.2 quadrillion tokens a month "across our surfaces", up from roughly 480 trillion a year earlier and 9.7 trillion two years before, and that its model APIs were processing roughly 19 billion tokens per minute.4 A month is about 30 days, so, as our own arithmetic and not Google's statement, 3.2 quadrillion a month spread evenly is roughly 107 trillion a day. That lands in the same range as the Chinese figures, which is exactly why it is misleading to put them on one chart: the keynote doesn't list which surfaces are counted, it is a monthly total rather than a daily average, and Google doesn't define it beyond calling tokens "the fundamental units of data our models process".4
OpenRouter: a measured slice
OpenRouter's is the only figure here from a party that publishes its counting rule. It totals prompt and completion tokens from requests through its own API, in each model's native tokenizer, in UTC daily buckets per model variant, and warns that "a higher token total shows how much a model is used, not which model is best".5 In its weekly view, the top-ranked variant, DeepSeek V4.1 Flash 1017, shows 26.7 trillion tokens.5 That is one model on one router, a small slice of all usage, and not comparable with an organization-wide count; see Most-Used AI Models on OpenRouter by Weekly Tokens for the full table and its caveats.
Why the figures can't be ranked
Four things differ, so no comparison between rows is defensible. The unit and period vary: a daily average, a daily figure, a monthly total, a per-minute rate and a weekly total. The scope varies: a country, one company's model, a company's whole product range, a router's traffic. The counting rule is stated for only one row (OpenRouter), and each model family counts in its own tokenizer, so the same text is a different number of tokens in different systems. And the source differs: an official statement, company remarks, a keynote transcript and a service's own measurement. A fair comparison would need the same definition, period and scope from each party; until that exists, the safe use of these numbers is each one on its own, with the unit and period attached, and each organization's growth against its own earlier figures.
Sources
- Xinhua, 我国日均词元调用量超140万亿 (China's average daily token invocation volume tops 140 trillion), Mar 24, 2026, retrieved 2026-10-25.
- Xinhua, 我国日均词元调用量超140万亿 企业加码“词元经济”产业链布局, Apr 1, 2026, retrieved 2026-10-25.
- TechNode, Doubao surpasses 120 trillion daily tokens as usage doubles in three months, Apr 7, 2026 (reporting Volcano Engine remarks via Jiwei), retrieved 2026-10-25.
- Google, I/O 2026: Welcome to the agentic Gemini era (edited transcript of Sundar Pichai's remarks), May 19, 2026, retrieved 2026-10-25.
- OpenRouter, AI Model Rankings (weekly view and "how these rankings are measured"), CC BY 4.0, retrieved 2026-10-25.