Loading today’s stories
Product1 distinct publisher3 min readPublished
The largest token allowances in China go to developers, not shoppers. The real pressure on Western AI pricing sits in API list prices that run 60% to 90% below OpenAI and Anthropic.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
"You're full after eating the dumplings. Go collect some tokens at the cashier so your [AI] agent can be fed with some computing power too." [16] Those cards sit on the tables at Jingu Yuan, a Beijing dumpling restaurant [15]. The card assumes the diner has an agent to feed, and it never says what the credit will get it. That is the flaw in the unit, and it survives every repackaging.
The rates expose odd packaging. China Telecom's entry plan works out at 0.99 yuan per million tokens [18]. Buy through your phone bill from China Mobile in Shanghai, at 400,000 tokens for 1 yuan [12], and the rate is 2.5 yuan per million [19]. China Unicom's two business tiers, 15 yuan for 6 million and 45 yuan for 18 million [13], both land on that same 2.5 [20]: triple the commitment and the unit price does not move, which is the thing a data plan normally rewards. Telecom is roughly 2.5 times cheaper per token than either [21], and none of the three tells a buyer how many questions their allowance answers.
The generous end went to developers. China Merchants Bank's June card offered new holders up to 1.8 billion tokens on MiniMax models, and Shanghai Pudong Development Bank up to 3 billion on Alibaba's Qwen [8][9], which is 180 and 300 times a month of Telecom's consumer allowance [22]. Beijing's AGI Bar serves unlimited DeepSeek V4 Flash from a local Nvidia DGX Spark to anyone who buys a drink and joins the Wi-Fi, and it aims at programmers and startup founders [14]. The people holding the biggest balances already know what a million tokens buys. The genuinely consumer-facing end is a 24-hour coffee shop opening in Changsha that plans to hand out tokens with the coffee [17].
Poe Zhao, who writes the Hello China Tech newsletter, told Rest of World that some of these consumer packages are a "supply-led experiment" [5], and that most ordinary users still meet AI inside an app or a feature "without seeing a token balance" [6]. The volume behind that reading is real: daily token consumption in China grew by a factor of 5,000 between early 2024 and mid-2026 [24]. Capacity priced that low has to go somewhere, and a cafe promotion is a cheap place to put it.
The list price is what should worry a Western pricing owner. Rest of World reports comparable Chinese open-source and open-weight models running 60% to 90% below OpenAI and Anthropic [4], which puts a job costing $10 on the US models at $1 to $4 on the Chinese ones [23]. Nothing in the reporting says how many vouchers or allowances are ever redeemed [26], so the bundles are distribution for spare inference, not proof that households will buy compute by the million.
If you are weighing an allowance for your own pricing page, two axes do the work: whether the buyer can state what the allowance buys before spending it, and who absorbs the overage when it runs out. Seat and message plans sit in the intelligible corner with the vendor carrying variance. The telco token plans sit in the opposite one, an opaque unit with the user carrying variance. The Kimi card, which pays token, agent and coding quotas on card spending [7], dodges the question by being money the holder never budgeted. That is why the telco version travels badly. It asks the buyer to underwrite a unit only the seller can read.
Ranked by verification strength, evidence, and original report placement.
Banks, cafes, telcos and other businesses in China are offering free AI tokens as incentives to attract customers, with some selling tokens much like mobile data plans, offering them as credit card rewards, or giving them out like free Wi-Fi.
An AI token is a unit of computing that an AI model uses to process text; words in a prompt are broken into tokens that the model processes to generate a response.
Daily token consumption in China surged to 500 trillion in mid-2026 from 100 billion in early 2024.
Chinese open-source and open-weight models, despite lagging behind U.S. rivals on some measures of frontier performance, can cost 60% to 90% less than comparable models from OpenAI and Anthropic.
Poe Zhao, a Beijing-based analyst and founder of the newsletter Hello China Tech, told Rest of World that some of the token packages Chinese companies pushed into the consumer market are a "supply-led experiment."
Zhao said: "Most ordinary users still encounter AI through an app or feature, without seeing a token balance."
Distinct publishers with included, body-backed reporting in this cluster.
Follow any of these and your For You feed starts watching them — no settings page required.
product
A Beijing dumpling shop gives away more AI credit with a meal than China Telecom sells for a month1 distinct publisher
invest
Google Ships Flash Instead of Pro While OpenAI Loses Its Two Best Operators1 distinct publisher
product
Swapping a frontier API for a self-hosted open-weight model relocates the audit question1 distinct publisher
product
Baidu's AI line grew 25 percent and still lost the arithmetic1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One outlet, named specifics, unchecked totals
Rest of World names the banks, the carriers, the tariffs and the two Beijing venues, which makes almost every product claim checkable in principle. Nobody has checked them yet, and the two figures doing the most work in the story — 500 trillion tokens a day and the 60% to 90% price gap — appear without a study, filing or company statement behind them.
Every measured figure here is supply-side
Three carriers have live tariffs, three banks have live cards, TokenHub aggregates 142 models, and one dumpling chain is moving over 100 vouchers a day across two stores. All of that measures what is being offered. The single demand number, national consumption, covers the whole Chinese market and cannot be attributed to these bundles.
Consumer framing runs past the usage proof
The story presents tokens as entering everyday Chinese life, while its own detail shows businesses pushing a unit at people who have no way to price it. The largest allowances go to developers, the bar with unlimited model access is filtering for programmers, and the analyst quoted calls the consumer packages a supply-led experiment. The overstatement is modest because the correction is inside the piece.
Every price is a marketing offer
The quantities here originate on the sell side. Banks are buying card activations, China Telecom is openly trying to bill for something other than connectivity, and Moonshot and MiniMax get distribution funded by someone else's promotion budget. Rest of World has no stake in whether the offers work, but it is reporting numbers that exist to attract customers.
Confidence rests on product facts, not behaviour
Who launched what, at what advertised price, is solid enough to act on. The claims about consumer AI habits rest on a single outlet and on allowances quoted without a model, an expiry or a rate limit, so the trend reading stays provisional.