The invoice arrives on the first Tuesday of the month, and for two years the largest line on it was storage. This month the finance manager at a freight-forwarding company in Shenzhen runs her finger past the object-storage charge and stops at a row she has never had to explain: model calls.
Her team connected a document assistant to customs paperwork in the spring. Nobody built a business case around the meter, because it had only ever gone down. Now the number is large enough that her director will ask on Friday.
The explanation is the Chinese token market in miniature. The price war that made those calls cheap is over — not because demand faded, but because demand became enormous and supply became expensive. Both halves matter, and most commentary keeps only the first.
Key takeaways
- China's average token consumption reached 140 trillion a day by March 2026, against roughly 100 billion in early 2024 — a rise of more than a thousandfold (National Data Administration).
- Doubao alone went from about 100 billion tokens a day in May 2024 to more than 120 trillion by March 2026 (industry data, cited by CGTN).
- In March 2026 Tencent Cloud raised some model input prices by 463%, Alibaba Cloud AI compute and storage by up to 34%, Baidu Cloud by 5–30% (China Development Network).
- Agentic coding consumes 10–50 times the tokens of ordinary chat, which is why enterprise bills rose while unit prices fell (CGTN).
- Alibaba formed a Token Hub business group and Tencent renamed its model platform TokenHub, both selling metered agent capacity rather than seats.
The demand curve bent upward
Average daily token consumption reached 140 trillion by March 2026, up from about 100 billion in early 2024 — a thousandfold increase in roughly two years (National Data Administration). It was also up more than 40% from the 100 trillion recorded at the end of 2025.
ByteDance's Doubao shows the same shape at company level: about 100 billion tokens a day in May 2024, more than 120 trillion by March 2026 (industry data, cited by CGTN).
What changed is not that people type more. It is that software started typing. An agent planning a task, calling tools, reading results and retrying consumes many times the tokens of a single question and answer. The published range for agentic coding is 10–50 times plain chat (CGTN, 18 June 2026), and that multiplier is the demand story.
Then supply got expensive
In March 2026 the direction of prices reversed. Tencent Cloud raised prices on some model inputs by 463%; Alibaba Cloud raised AI compute and storage by up to 34%; Baidu Cloud by 5–30% (China Development Network; China Economic Net). The moves were disclosed and landed within weeks of each other.
The plausible reading is that serving capacity became the constraint. Memory is the binding input: CXMT's 2026 output of about 2.2 million HBM stacks supports only 250,000–400,000 Ascend-class packages, while SMIC can make die for more than a million (SemiAnalysis estimate). Huawei's own pricing tells the same story — the Ascend 950DT reached about RMB 250,000 (≈US$35,200 / HK$275,000) per card in September, up 60% in three months (company pricing).
Electricity is the second constraint. China's data-centre demand for compute (算力) is forecast at about 300 TWh in 2026, against roughly 260 TWh in the United States (Beijing Daily, January 2026). Cheap tokens were always a claim on physical infrastructure, and that infrastructure is now scarce.
The market splits in two
Once supply is short, an undifferentiated price war becomes irrational. What replaces it is a two-tier market.
The frontier tier is sold on capability. Qwen3.7-Max is listed at US$2.50 per million input tokens and US$7.50 per million output tokens (CGTN, 18 June 2026). Buyers at that price are paying for reasoning, long context and reliability on tasks where a wrong answer costs more than the call.
The commodity tier is sold on cents. DeepSeek V3.1 lists at RMB 2 and RMB 8 (≈US$0.28 and US$1.10; HK$2.2 and HK$8.8) per million input and output tokens; Doubao Lite at RMB 0.30 and RMB 0.60 (≈US$0.04 and US$0.08; HK$0.33 and HK$0.66). At that level the model is a utility; margin sits in volume and routing.
The middle is the dangerous place: a model that is neither the best reasoner nor the cheapest router has no story for a procurement team.
The organisational news points the same way. Alibaba created an Alibaba Token Hub business group in March 2026; Tencent Cloud renamed its model platform TokenHub and published an agent governance suite.
Who gets squeezed
Three exposures are worth naming.
- Thin-margin resellers. A product priced as a monthly subscription but billed per token carries an open-ended cost. The 463% increase is a stress test many small teams fail.
- Buyers who budgeted on deflation. Teams that modelled falling unit prices and rising usage now face both curves moving the wrong way.
- Agent builders. The 10–50× multiplier is a feature when you are demonstrating capability and a liability when you are paying for it.
The response is to know which tokens are worth buying: route routine work to the cheap tier, reserve the expensive one for costly errors.
Honest limitations
The 140 trillion figure is a national average published without a breakdown by customer or by paid versus free usage, so revenue per token cannot be derived. The March increases were disclosed by the cloud providers, but the link to memory supply is an inference from timing and Huawei's list prices, not a stated explanation. Agentic token multipliers come from vendor and industry reporting and vary widely by task. Company reorganisations signal intent; they do not disclose margins.
What readers can do now
- Operators: instrument cost per completed task, not cost per token. An agent that solves a ticket in three calls can beat a chat model that needs ten.
- Investors: ask what share of revenue is metered inference and how much of that is contracted at fixed prices. The companies raising list prices are also the ones absorbing memory costs.
- Researchers: report the token budget of an evaluation alongside its score. A result produced with a 50× agentic budget is not comparable to a single-pass one.