Tokens With Your Dumplings: How China Turned AI Compute Into a Consumer Perk

An AI token — the unit of computing a model consumes to process a request — was, until recently, a term that mattered to developers and nobody else. In China, it is now something you can earn with a credit card, buy in a monthly plan from your telecom operator, receive as a voucher with a plate of dumplings, or even borrow against.

That is the picture Rest of World painted on September 4 in a survey of how “computing power” is being packaged and marketed to ordinary Chinese consumers. The details are sometimes comic — a Beijing dumpling restaurant handing out 10-yuan compute vouchers after meals, a bar running a local DeepSeek model for anyone who buys a drink — but the underlying trend is serious: token prices have become a new front in the global AI race, and China’s aggressively cheap open-weight models have made it possible to treat tokens as a mass-market commodity.

The numbers behind the trend

Daily token consumption in China surged to 500 trillion in mid-2026, up from 100 billion in early 2024 — a 5,000-fold increase in roughly two and a half years. Chinese open-source and open-weight models, which trail U.S. rivals on some frontier-capability measures, are priced 60 to 90 percent below comparable models from OpenAI and Anthropic. That cost gap is the enabler: when a million tokens cost pennies, a restaurant can give away a few thousand with lunch and a bank can hand out billions as a sign-up bonus without blinking.

Poe Zhao, a Beijing-based analyst and founder of the newsletter Hello China Tech, described the phenomenon to Rest of World as a “supply-led experiment.” His point is that ordinary users still encounter AI through an app or a feature, without ever seeing a token balance — the packaging, in other words, is running ahead of the demand. Whether consumers actually want to think about compute the way they think about mobile data remains an open question. What is not in question is that Chinese companies, with cheap tokens in abundance, are running the experiment at scale.

Five channels where tokens meet consumers

Credit cards that reward spending with compute

In July, Moonshot AI — the developer of the Kimi model family — launched a credit card named after Kimi, in partnership with the Agricultural Bank of China and American Express. Cardholders earn Kimi AI tokens, along with agent and coding quotas, on their spending, exactly the way other cards reward users with airline miles or cashback.

Moonshot was not first. In June, China Merchants Bank launched a credit card targeting AI developers that offered new cardholders up to 1.8 billion tokens for use on MiniMax’s models. Shanghai Pudong Development Bank followed with its own AI developer card, offering subsidies worth up to 3 billion tokens for Alibaba’s Qwen models. A billion tokens as a sign-up bonus sounds absurd until you price it: at Chinese open-weight rates, it is a rounding error for the bank and a genuine hook for a developer deciding which model ecosystem to build on.

Telecom plans that sell tokens like data

China’s three major carriers have folded AI usage into their billing in a way that looks deliberately like mobile data. China Telecom sells consumer plans starting at 9.9 yuan (about $1.40) a month for 10 million tokens, with larger allowances for heavier users and businesses. It also operates TokenHub, a subscription platform spanning 142 large models, as the state-owned carrier repositions itself from selling connectivity to selling computing and AI services.

The competitors match it. China Mobile offered 400,000 tokens for 1 yuan (14 cents) in Shanghai, payable through the phone bill. China Unicom introduced monthly plans ranging from 15 yuan ($2.10) for 6 million tokens to 45 yuan ($6.30) for 18 million for business users. The pricing structure is familiar to anyone who has ever bought a data bundle — which is precisely the point.

Free tokens with your meal

Instead of free Wi-Fi, a growing number of Chinese restaurants, cafes, and bars lure customers with free AI computing. Beijing’s AGI Bar — aimed at programmers and AI startup founders — gives customers who buy a drink and connect to its Wi-Fi unlimited access to DeepSeek V4 Flash, running locally on an Nvidia DGX Spark workstation. The Jingu Yuan dumpling restaurant hands out 10 yuan ($1.40) of computing credits after each meal; cards on its tables read: “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.” Its two stores hand out more than 100 vouchers a day. A 24-hour coffee shop under construction in Changsha is planning the same play.

Tokens resold on secondhand marketplaces

On Xianyu, Alibaba’s secondhand marketplace best known for used phones and furniture, sellers now advertise cheap access to AI models in token-denominated packages — millions of tokens for a few dollars in day passes or monthly subscriptions, plus shared access to AI accounts. Some sellers break up model allowances and resell them to students, independent developers, and small businesses; others arbitrage the gap between official token prices and discounted or free access they obtained elsewhere. A gray market in compute, in other words, has emerged organically.

Bank lending scored by token throughput

The most consequential experiment may be in lending. China’s banks struggle to evaluate young AI startups using traditional metrics like physical assets. In August, the Haizhu district of Guangzhou launched a financial product called the “token loan,” which allows banks — including Bank of China, China CITIC Bank, and Bank of Guangzhou — to measure a company’s token production and consumption to decide how much it can borrow. The metric is a proxy for how much inference a startup is actually serving, and therefore for real usage and revenue, in an industry where traditional balance sheets say almost nothing.

Why this is happening in China and not elsewhere

Three conditions converged. First, price: Chinese open-weight models are 60-90 percent cheaper than U.S. frontier APIs, so giving tokens away costs little. Second, distribution: banks, telcos, and payment networks already reach hundreds of millions of consumers and are searching for differentiated perks in a commoditized card market. Third, policy: the state has pushed computing power — suanli — as a strategic resource for years, and carriers pivoting to “AI computing and services” fit that agenda neatly.

The contrast with the West is stark. U.S. and European AI consumption is mediated almost entirely through subscriptions (ChatGPT Plus, Claude Pro, Gemini Advanced) and enterprise APIs. Nobody earns GPT tokens with their credit card. The subscription model bundles capability; the token model unbundles it into a metered commodity. Which approach wins with consumers is an open question — but only one of them is being tested at consumer scale right now.

The skepticism, and the stakes

Zhao’s “supply-led” framing is the right lens for the odder corners of the trend. A dumpling voucher is a marketing gimmick; a token-denominated credit card is a customer-acquisition play; neither proves consumers think in tokens. The daily consumption number — 500 trillion tokens — is real regardless, and the token-loan experiment suggests the financial system is starting to treat compute throughput as an economic signal worth underwriting.

The deeper shift is that China is commercializing AI along the rails of its existing consumer economy — cards, data plans, vouchers, marketplace listings — rather than inventing new ones. If cheap tokens remain abundant, the experiment continues. If token prices rise (as DeepSeek’s own API price hikes earlier this year hinted they might), the freebies vanish first, and what survives will be the parts with real utility: the telco plans, the developer cards, and the loans. Either way, the rest of the global AI industry is watching a live test of what happens when compute becomes a consumer good — and so far, only China is running it.