Executive Summary

Start by pinning down the announcement itself. On March 24, 2026, at the China Development Forum, Liu Liehong, head of the National Data Administration, said China's daily token consumption had reached 140 trillion. That is up 40% from 100 trillion at the end of 2025, and more than 1,000 times the 100 billion of early 2024. In the same remarks, he formalized the Chinese term for token, 词元 (cíyuán), and named it "the settlement unit connecting technological supply and commercial demand."

How steeply the number climbed becomes clear only when you lay it out over time. The table below shows the daily average consumption at the four points the National Data Administration cited.

The multiples in the table don't quite capture the character of this growth. Put the same four points on a logarithmic axis and the picture shifts. On a log scale where each vertical step is a tenfold jump, the line still climbs almost straight — which means this isn't a one- or two-year flash spike but something close to an exponential curve that clears an order of magnitude on a regular cadence.

1.1How "1,000×" Became a Slogan

A small gap stands out here. Divide 140 trillion by 100 billion and you get exactly 1,400×, yet the announcement and the coverage uniformly reached for the phrase "more than 1,000×." You could call it conservative rounding, but the gap between the precise figure and the slogan is itself the first clue to how a metric gets communicated. People don't remember 1,400×. What they remember is the round symbol, "1,000×."

The policy backdrop came with it. The National Data Administration said that as of late 2025 it had built more than 100,000 high-quality datasets and 890 petabytes of data resources, declared 2026 "the year of releasing the value of data as a factor of production," and rolled out six policy initiatives. Within that larger frame, the 140 trillion tokens were positioned as the headline metric that sums up the growth of the intelligence economy in a single number.

When a state builds a new metric, it is usually because whatever the metric measures has become politically important. Just as electricity consumption once stood in for industrialization, China wants tokens to stand in for the intelligence economy. Tokens are a convenient candidate: they are the smallest unit you can measure, price, and trade. When the National Data Administration called tokens a "settlement unit," it was less a metaphor than a declaration of policy design.

Look at where this settlement unit is actually consumed and the picture tilts more sharply than you'd expect. ByteDance's Doubao (豆包) disclosed that as of April 2026 it processes 120 trillion tokens a day — roughly 86% of the national total of 140 trillion, concentrated in a single company. The diagram below shows that skew.

What deserves closer attention is what drives this concentration. The engine behind Doubao's token consumption is not text chat but AI video generation. The industry estimates that ByteDance's video model, Seedance, burns about 40,000 tokens to produce one second of 1080p video — so a few minutes of video overwhelm vast amounts of text conversation. Doubao's consumption later jumped again to 180 trillion at the June 2026 Volcano Engine conference, and Volcano Engine said it holds 49.5% of China's public-cloud MaaS market by tokens. In other words, one of every two tokens a Chinese enterprise uses passes through this single platform.

Once tokens became a settlement unit, capital markets responded. Right after the 140 trillion announcement, the combined market capitalization of Chinese AI-chip firms such as Cambricon and Hygon surpassed 2.5 trillion yuan (about $345 billion). Eight ministries issued guidance to deploy 1,000 industrial AI agents across manufacturing, logistics, and public services, and the 15th Five-Year Plan projected the AI-related industry would exceed 10 trillion yuan (about $1.4 trillion). The token, as a unit, became the axis threading policy, investment, and industrial strategy onto a single line.

The first line of skepticism toward the 140 trillion targets the character of the consumption. In April 2026, Reuters Breakingviews ran a column headlined, in effect, "China's token obsession may be misguided." The original sits behind a paywall, so direct quotation is difficult, but the argument that several secondary outlets cross-cited is clear: if a large share of the total is bots churning out posts for content farms or automated systems trading messages with one another, the number swells while the corresponding economic benefit may be hollow.

This skepticism has data pointing in its direction. As we saw in the previous section, the main engine of the concentration was AI video generation, not conversation with people. On top of that, the share of LLM token use going to coding and programming jumped from 11% in early 2024 to more than 50% in 2026, and by some tallies agent-driven workflows account for over half of all output tokens. The industry estimates that agent workloads use 10 to 100 times more tokens than conversational chatbots. That multiple itself can't be stated with confidence, since the original research isn't fully traceable, but the direction is consistent: the more automation grows, the more tokens get consumed with no bearing on human benefit.

Tellingly, the government's own moves lend weight to this skepticism. That the National Data Administration went out of its way to declare 2026 "the year of releasing the value of data" and put forward six initiatives reads as evidence that policymakers, too, recognize that the phase of growing quantity differs from the phase of extracting value. The signal is that the total is already big enough, and the task now is to turn that total into value.

3.1Hearing the Other Side

That said, skepticism isn't the whole story. For balance, the other side deserves a hearing. A column from the China Europe International Business School (CEIBS) offers a different angle: the very fact that token consumption draws this much attention is itself evidence that tokens have emerged as a core metric for an organization's AI transformation and productivity. AI deployed inside a company doesn't show up on public leaderboards, yet its strategic value can be greater — the same context in which Jensen Huang said at GTC 2026 that "the annual token budget per engineer reaches half of their salary." From this vantage, the token total isn't a phantom figure but a proxy signal of real economic activity.

The two views look contradictory, but they are really two faces of the same fact. Some of token consumption creates genuine value, some of it is waste, and today's aggregate metric can't tell the two apart. The skeptic says don't trust an undifferentiated total; the optimist says there is real value inside that total. Both are right. Which is why the real question isn't "is 140 trillion big or small?" but "can we tell, within it, what is effective and what is waste?"

Here is the center of gravity of this report. If you want to use tokens as an economic metric, what should you measure, and how? Start with three fault lines: the uncertainty of definition, the order-of-magnitude mismatch in the totals, and the divergence between unit price and total volume.

4.1No One Defined What the 140 Trillion Counts

The most fundamental fault line is definition. No primary source from the National Data Administration states what the 140 trillion includes and excludes. Is it pre-training tokens consumed while training a model, inference tokens spent responding to user requests, or communication tokens that agents pass among themselves? Depending on how you add and subtract the three, the same activity can double or halve. No definition, no verification; no verification, no comparison. The first condition of a metric is not a big number but agreement on what is being counted.

4.2Count the Same Thing Differently and the Digits Diverge

Why definition matters becomes plain when you set two actual numbers side by side. The National Data Administration's all-of-China tally is 140 trillion a day. Meanwhile, consumption by Chinese models routed through the third-party platform OpenRouter runs about 13 trillion a week — roughly 1.85 trillion a day. Both are "China's tokens," yet they diverge by more than an order of magnitude. This is not an error. The national tally tries to capture the whole, while the platform statistic captures only one specific channel. Few examples show more sharply that what you count is what makes the number.

4.3Quantity Rose as Value Fell

The third fault line is the most counterintuitive. Over the very two years that token consumption grew more than 1,000-fold, the unit price of a token fell by as much as 1,000-fold. The GPT-3 era's roughly $60 per million tokens came down to less than a thousandth of that by 2026, and the gap is wide even among contemporaneous models: DeepSeek V3.2 runs $0.42 per million output tokens against Claude Opus at $75 — about a 170× difference. If volume rose 1,000× while unit price dropped 1,000×, real spending may not have risen nearly as much as it seems. The diagram below shows this symmetry.

4.4So What Should We Measure?

The three fault lines point to a single prescription: divide the total by value before measuring it. Token consumption is an unclassified aggregate that mixes effective tokens, which contribute to a task, with waste tokens from retry loops, bot-to-bot chatter, and duplicate generation. Fail to separate the effective from the wasteful and you fall into the same illusion — whether in national statistics or a corporate dashboard. The table below contrasts the two ways of measuring.

This problem wears an old face. The critique that GDP measures output but not welfare or distribution, and Campbell's Law — that a social indicator corrupts once it is used for decision-making — both point at the same structure. An aggregate proxy is convenient but misses "activity toward what end." A token metric can become the intelligence economy's GDP, and if so, it inherits the half-century of criticism GDP has drawn. The answer isn't to throw the metric out but to stand a value indicator beside the total.

The national-scale story is really every organization's story. When a company that has adopted AI internally reports results, the easiest number to reach for is token consumption. A dashboard showing twice as many tokens as last month looks, at a glance, like AI adoption has doubled. But this very equation is a miniature of the trap China's 140 trillion faces. Using more is not the same as using well.

Goodhart's Law works even faster inside an organization. Once "token consumption = adoption" becomes a team's KPI, the team makes using more tokens the goal. It pads prompts unnecessarily, lets retry loops run, and leaves agents to converse with one another at length. Costs rise and outcomes stall, yet the dashboard number keeps growing healthily. This is not a hypothetical risk; it is a FinOps reality already observed as agent-based workflows spread.

The prescription is the same one we offered the state: stand a value indicator beside the total. Concretely, measure three things together. First, task completion rate — regardless of how many tokens were spent, did the user's request actually get resolved? Second, outcome per token — if the same result came from fewer tokens, that is improvement. Third, the ratio of waste tokens to effective tokens — separate out and track the share that retries, duplicates, and bot-to-bot chatter take of the total. With these three in place, token consumption itself finds its proper place as a supporting metric for cost management.

This is where data quality intervenes decisively. Separating effective tokens from waste tokens is, in the end, the work of judging with data what contributed to a task. Only with a well-curated evaluation set, logs tied to task success, and data-lineage tracking can you finally divide the total by value and measure it. Refining data to raise verification efficiency, rather than scaling compute without limit, is a practical path open to every actor with finite resources — state or enterprise alike.

Primary announcements & reporting

• 1.National Data Administration (中國國家數據局) briefing. China Development Forum (中國發展高層論壇), 2026-03-24. Director Liu Liehong (刘烈宏) — 140 trillion daily tokens; naming 词元 as a settlement unit. (Primary announcement)
• 2.Yicai Global (2026). China's Daily Token Usage Jumps 40% in Three Months. yicaiglobal.com.
• 3.CEIBS — Yang Wei column (2026). On the token economy and organizational AI-transformation KPIs. ceibs.edu. (Skepticism-vs-rebuttal framing)
• 4.hellochinatech (2026). China token economy 140 trillion. hellochinatech.com. (Goodhart framing; OpenRouter data)
• 5.Reuters Breakingviews (2026-04-15). China's AI token obsession may be misguided. Big View section. (⚠️ Original paywalled — argument based on cross-verified secondary citations, no direct quotation)
• 6.Fortune (2026-04-12). China token economy AI boom. fortune.com. (Alibaba / Z.ai going-private material)

Concepts & theory (classics, established views)

• 7.Goodhart, C. (1975); formalized by Strathern, M. (1997). "When a measure becomes a target, it ceases to be a good measure."
• 8.Campbell, D. (1979). Assessing the impact of planned social change — Campbell's Law (a social indicator corrupts once used for decision-making).
• 9.Kuznets, S. (original caution); Stiglitz, J., Sen, A., & Fitoussi, JP. (2009). Report by the Commission on the Measurement of Economic Performance and Social Progress — critique of GDP as a welfare measure.

Pebblous-adjacent

• 10.Pebblous Blog. Data Curation Bottlenecks and Foundation Models — 38B Tokens Beat 350B Tokens.
• 11.Pebblous Blog. AI Agent Token Costs and Retry Loops — A FinOps View.

Note: The national 140T (Mar 2026) and ByteDance's Doubao 120T (Apr 2026) were reported a month apart, so "~86%" is an approximation. "More than 1,000×" is the reported phrasing; a precise calculation gives about 1,400×. The agent-workload token multiple (10–100×), the Seedance video-token conversion, and JPMorgan's 370× projection are industry estimates whose original research/sources are not fully traceable, and are marked as "estimates" in the text. The OpenRouter figure captures only part of one third-party routing channel.