Token Deals Hit Singles' Day: How to Actually Budget AI Compute
This year on Singles' Day shelves, a new product appeared: Token.
Tmall data shows a 0.01 yuan AI membership monthly card sold 50,000 units in 10 minutes, nearly 100,000 in half an hour, corresponding to about 2 trillion Token credits. On JD.com, AI PC sales bundled with Token packages rose 100% month-over-month. Invisible and intangible compute power, for the first time, sits in the same shopping cart as tissue and laundry detergent.
Token, also called word unit, is the smallest unit of information a large model processes. When you ask AI to write an article or do an analysis, it consumes word units one by one. It is the billing unit of the AI world, analogous to kilowatt-hours of electricity.
The difference: electricity bills are usually settled at the end of the month, while AI's "electricity" is billed per use from the first conversation.
Why is everyone suddenly stockpiling Tokens? Put three signals together and it becomes clear:
- Consumer side: Token officially becomes a mass consumer product; live-stream rooms and e-commerce platforms directly sell token packages.
- Industry side: Data from the China Electronic Information Industry Development Research Institute released at the end of October shows China's AI inference data volume exceeded training data volume for the first time, with daily token calls surpassing 175 trillion, up more than 1,000 times from early 2024. AI's main work has shifted from "learning and training" to "working and reasoning."
- Chip side: At the CIIE expo, Intel and Qualcomm both put AI PC and on-device solutions at center stage. Intel China chairman Wang Zhicong judged: in the next five years, a larger proportion of compute power will be deployed on the inference side, extending toward devices closer to users.
These three things point to one conclusion: inference is becoming the "daily electricity" of the AI era.
But there is another side to the promotion wave worth watching out for.
Besides official stores, a batch of low-priced "AI relay stations" have appeared on e-commerce platforms. Merchants claim "30% of official pricing" and "one yuan for ten million Tokens," but operators hide their identities and have no public filing information.
The few yuan you save come at the cost of every question passing through a stranger's server first. Bare data, privacy leaks, malicious injection, remote control, and data leaving the country — these risks all hide beneath the low price.
Experts at the National Information Center have already called for: unifying the token measurement standard, standardizing billing disclosure, and introducing third-party verification. Tokens from different models cannot be directly price-compared, and rules on quota validity, refunds, and privacy protection also need improvement.
So how should ordinary people and businesses count this compute bill?
iFlytek gave a division forecast at the 1024 Developer Festival: In the future, 60% of AI tasks can be completed locally, 30% via device-cloud collaboration, and only 10% truly needs the cloud. In other words, the bulk of the bill can stay in your own hands.
This is exactly the approach we are practicing at Codigger:
The Boby computer uses local inference for high-frequency, privacy-sensitive daily tasks — this part does not consume cloud Tokens.
The cloud workstation handles heavy compute and large-scale collaboration — this part is pay-as-you-go, on-demand.
Bought compute handles daily use, metered compute handles peaks — the bill becomes predictable.
Token will increasingly become as routine as electricity bills, but remember: 30%-off "electricity" may not save money — first ask which wire it comes from. And an AI computer that knows how to count can help you find the most cost-effective path between "buying" (local compute) and "renting" (cloud Tokens).