China Runs 140 Trillion AI Tokens a Day. Your Brand Visibility Score Samples About 2% of Them.

On March 23, 2026, at the China Development Forum, the head of China's National Data Administration disclosed a number that got reported everywhere and analyzed almost nowhere: China's daily token throughput had passed 140 trillion tokens per day, up more than 1,000x from roughly 100 billion in early 2024. In February and March, China's weekly token volume overtook the United States for three straight weeks.

In the same quarter, QuestMobile counted 440 million monthly active users across China's AI-native consumer apps — Doubao at 345M, the Qwen app at 166M, DeepSeek at 127M.

Those two numbers cannot both be describing the same activity. And the gap between them is where almost every brand visibility program in China is currently pointed in the wrong direction.

The arithmetic nobody runs

Let's estimate how much of that 140 trillion comes from a human typing a question into a chat app.

Take the generous end. Assume all 440M monthly users are active daily (they are not), that each runs 20 conversational turns a day (heavy usage), and that each turn costs 2,000 tokens of prompt plus response. That gives roughly 17.6 trillion tokens/day — about 12.6% of the national total.

Now take the realistic end. Assume a 30% DAU/MAU ratio (~130M daily users), 10 turns a day, 1,500 tokens per turn. That gives roughly 2 trillion tokens/day — about 1.4%.

So consumer chat is somewhere between about 1.5% and 13% of China's daily AI inference, and the realistic figure sits at the low end of that band. Call it under 10%, plausibly nearer 2%.

Every commercial brand visibility tool operating in China — including the panel behind our own /brands scores — measures that slice. We prompt the consumer-facing model, with its default tool configuration, and record whether your brand appears. That is a legitimate and useful measurement. It is also a measurement of the minority surface.

What the other 90% actually is

The remainder is API and agent traffic: coding assistants, customer service bots, in-app copilots, enterprise RAG systems, batch content generation, in-car assistants, phone-OS agents, and multi-step workflows where a single user request fans out into dozens of model calls.

This is why enterprise AI spending in China keeps rising even as unit token prices collapse. As Huxiu documented in its analysis of what it calls the "token price tearing" phase, companies have moved from simple Q&A to agent collaboration and code generation — task types that consume ten to a hundred times the tokens of a single chat turn. Cheaper tokens, far more of them.

For a brand, the relevant question is not how big that layer is. It is: when a model in that layer is asked about your category, what does it use to answer?

Cheap tokens change how the model answers about you

Global LLM inference pricing has fallen roughly 88% since 2023, to around $1.16 per million tokens on average. In China the floor is lower still. DeepSeek's most recent price moves cut cache-hit input from ¥1 to ¥0.1 per million tokens and cache-miss input from ¥12 to ¥3.

At ¥0.1 per million tokens, a deployer running millions of calls a day optimizes hard for three things: aggressive prompt caching, the smallest model that clears the quality bar, and the fewest tool calls per request.

A web search tool call is expensive on all three counts. It adds seconds of latency, adds retrieval cost per request, and injects fresh content into the prompt — which breaks the cache that makes the economics work in the first place.

The practical consequence: on the highest-volume surfaces in China, the model answers questions about your brand from its weights, not from the web. Not from the Baike entry you corrected last month. Not from the SMZDM review you seeded in July. From whatever was in the pretraining corpus, and how prominently.

That splits brand visibility into two layers that behave completely differently:

The parametric deficit test — run it before Q4 budget

You can measure your own exposure to this in an afternoon, on one model, without buying anything.

1. Pick 10 category queries your buyers actually ask. Not branded queries — category queries ("recommend a five-star hotel near Guangzhou Baiyun Airport", "which sunscreen suits sensitive skin").
2. Run each query twice against the same model. Once through the consumer app or API with web search enabled. Once through the raw API with no tools, temperature fixed, no system prompt.
3. Record two things per run: whether your brand is mentioned at all, and its position in the list if it is.
4. Report the delta. Mention rate with retrieval, minus mention rate without.

That delta is your parametric deficit. Four readings:

• High in both — durable visibility. Your brand is in the weights and in the fresh evidence. Rare.
• High with retrieval, low without — the common result for brands with an active GEO program. Your visibility is rented. It works in the chat app and disappears the moment the model is deployed without search, which is most of the time.
• Low with retrieval, high without — legacy brand equity carrying you while your content operation underperforms. The fastest fixable case.
• Low in both — start with the retrieval track. It is the only lever that moves inside a quarter.

Two honest caveats. First, "no tools" API output is a proxy for parametric memory, not a pure measurement of it — post-training and safety tuning also shape the answer. Second, this test is per-model. Because Chinese models ship new versions on a fast cycle, a parametric result has a shelf life; we have written before about how model version churn quietly breaks score comparability across months.

Why the parametric layer is slower — and stranger — to buy

The two layers reward different things, and in at least one case they reward opposite things.

Retrieval rewards recency, citability, and source authority. Pretraining corpora reward the opposite profile: text that is old, widely duplicated across domains, and structurally clean. Encyclopedia entries. Forum threads that get quoted and re-quoted. Product documentation. Wire copy syndicated across hundreds of sites.

Note what that implies. We have previously shown that press release syndication is close to worthless at citation time — a 0.04% citation rate in our sampling. But mass duplication across domains is precisely the property that gives a piece of text weight in a pretraining corpus. The same tactic can be near-useless in one layer and meaningful in the other. We flag this as a working hypothesis rather than a measured finding; it is testable with the deficit test above, on a brand that has run heavy syndication.

The second parametric variable is one nobody budgets for: name consistency. In the retrieval layer, a model resolves 宝格丽 / BVLGARI / 寶格麗 to one entity because the retrieved page makes the link explicit. In the parametric layer, three surface forms with inconsistent co-occurrence can behave like three weaker entities. A brand that has used different Chinese renderings across its Tmall store, its official site, its OTA listings, and its press coverage has been splitting its own weight for years.

Takeaway for Brand Marketers

1. Stop describing your GEO score as coverage of "China AI." It is coverage of the consumer chat surface, which is a single-digit percentage of national inference. Say so in the deck. It makes the number more credible, not less.
2. Run the parametric deficit test before Q4 planning locks. It costs a day of an analyst's time and it will tell you whether the last year of GEO spend bought durable presence or rented presence.
3. Split the budget into two tracks with different cadences. A retrieval track that runs monthly — fresh content, source seeding, Baike accuracy, review-platform presence. An entity track that runs quarterly and is judged on a much longer horizon — canonical naming, structured factual consistency, presence in high-duplication reference sources.
4. Audit your name variants this quarter. Pick one canonical Chinese name, one canonical English name, and one explicit stated pairing between them, and enforce it across every property you control. This is the cheapest parametric intervention available and almost no one does it.
5. When an agency pitches you, ask which layer the deliverable moves. If they cannot answer, the deliverable moves the retrieval layer, and you should price it accordingly.

The 140 trillion figure is not a curiosity. It is the size of the surface your measurement does not reach — and in 2027, when more of China's AI-mediated brand exposure happens inside embedded agents than inside chat apps, it will be the only surface that matters.

Related: see current brand scores across China's six major AI models on hubGEO /brands.

Sources: National Data Administration remarks at the China Development Forum, March 23, 2026 (Xinhua); QuestMobile Q1 2026 AI application report; DeepSeek API pricing announcements; Huxiu analysis of China's token price divergence, 2026.