321, Go Token: Who's Grabbing the "Recharge Entry Point" of the AI Era

11/18 2026 381

On the evening of November 2, a 1-cent AI membership monthly card sold nearly 100,000 units in half an hour in Li Jiaqi's livestream, corresponding to about 2 trillion tokens, with actual transaction value of about 1,000 yuan. This is a beautiful enlightenment, but not a viable business. What the platform gives up is profit, in exchange for whose page users open when they pay for AI for the first time; what vendors cut is price, in exchange for the default entry point; and what the industry truly lacks is a ruler that can calculate tokens clearly. When "water, electricity, and coal" are put on the shelf, being cheap is only the first step; being measurable, comparable, and repeat-purchasable is what determines how far this road can go.

Original Xinshang AI New Tech Team

Over the past two years, tokens only appeared in developers' bills and large-model service subscription pages, priced as "a few yuan per million tokens." This November, their trading venue switched to livestreams and shopping carts, and the unit of measurement became "units." From B-end backends to C-end shelves, AI commercialization has taken a path no one pre-designed, and this path was paved together by platforms, vendors, and streamers over two months.

Selling computing power as a commodity was never easy to begin with: it is invisible, intangible, has no packaging, no shelf life, yet is almost the common fuel behind all productivity tools in this era. It took twenty years for phone bills and data plans to make ordinary people accustomed to paying monthly for communications. But there is an easily overlooked difference here—data plans are standard products; 1GB is 1GB everywhere; tokens are not. For tokens to walk the same road, the first step is to get the most people to complete their first order, and the last step is to let everyone clearly calculate what they have bought.

An invisible commodity: 1 cent, 2 trillion, 1,000 yuan

On the evening of November 2, the Qianwen Office personal membership monthly card entered Li Jiaqi's livestream. The original price was 28 yuan, and the livestream price was 0.01 yuan. It sold 50,000 units in 10 minutes and nearly 100,000 units cumulatively in half an hour, corresponding to about 2 trillion tokens of invocation quota. Chen Yusen, vice president of Alibaba Group and CEO of Qianwen Office, demonstrated on site using a few sentences of dialogue to generate a global fishing map accessible to multiple people; Li Jiaqi translated 500 points into a task volume ordinary people could understand—50 PPTs, 160 Word documents, and 250 Excel spreadsheets—and revealed that his livestream was already using AI to generate shopping rankings in real time.

This was the first time a top livestream room listed AI computing power as an independent commodity. The process was no different from other products: explain the selling points, give the price, count down and put up the link. The only difference was that consumers were not buying a lipstick, but a piece of instantly credited computing capability.

But there are two details more worth watching than "2 trillion."

First, this order was designed not to make money. Each Taobao ID was limited to one order, and only new users could activate it. 100,000 units times 1 cent equals actual transaction value of about 1,000 yuan. This is not a pricing mistake, but a standard user-acquisition move: what 0.01 yuan buys is not computing power, but the first binding, the first opening, the first payment.

Second, the buyers may not necessarily be able to use it. Cui Lili, vice dean of the Institute of Digital Economy at Shanghai University of Finance and Economics, said directly that the audience in Li Jiaqi's livestream is mainly white-collar workers and mothers, highly overlapping with groups that previously had almost no connection to AI computing power; "most people may just be taking advantage of the freebie." Token use still has technical thresholds, and long-term effects depend on activation rate. Whether they are willing to renew at 28 yuan after a month is the hard indicator. Zhang Yi of iiMedia Research judged that the "enlightenment significance far exceeds actual sales," while Chen Duan of the Central University of Finance and Economics positioned it as a landmark event in AI commercialization extending from the B-end to the C-end.

In summary: what it sold was not 2 trillion tokens, but a mass-level cognitive popularization; what it recovered was not revenue, but 100,000 accounts awaiting verification. One cent sent tokens into 100,000 shopping carts, but left the most expensive question to vendors—among these people, how many will actually open it?

A two-month positioning battle: what platforms want is not transaction value, but the default entry point

This battle for the entry point was compressed into an extremely short time window.

On September 2, Zhipu's first official flagship store landed on Tmall; on September 3, Tmall's "AI Space Station" Token recharge center went online, connecting vendors such as Alibaba Cloud, Zhipu, Kimi, MiniMax, and DeepSeek, followed by Qianwen Office, StepFun, and Qoder; JD.com launched an "AI Software Recharge Hall" and "AI Token Service Station" in the same period, packing model selection, quota recharge, and membership subscriptions onto the same page. On October 20, the first day of Double 11 sales, Tmall's Token category transaction value increased 450% month-on-month; on November 2, Li Jiaqi shouted "321, Go Token"; on November 11, vendors such as Zhipu, MiniMax, Qianwen AI, and Qoder offered packages at 50% or even 30% of the price.

The supply side moved faster than the demand side. On November 9, Alibaba Cloud's Token Plan personal edition kept monthly fees unchanged across three tiers while raising quotas by about 74% to 78%; the Pro tier monthly fee dropped from 499 yuan to 312 yuan, and the unit quota cost of all four tiers fell by more than 40%. Baidu Qianfan offered new users 50% off on first purchase, starting as low as 4.9 yuan, with some packages adding up to 74% bonus points and designated models as low as 20% of the price at night; Tencent Cloud TokenHub offered 50% off pay-as-you-go calls and a limited-time 25% price on Hy4 for Token Plan personal edition; Volcano Ark Coding Plan was about 35% of the original price after stacking; the Qianwen AI Lite monthly card was priced at 39 yuan, more than 30% cheaper than usual. In the past month, at least eight domestic and foreign vendors adjusted prices, quotas, or usage restrictions.

Looking further back, this was not even a Double 11 impulse: in March 2026, Alibaba established the Alibaba Token Hub business group, directly led by the group CEO, with a mission stated plainly—"create tokens, deliver tokens, apply tokens"; in the same period, Tencent Cloud upgraded its MaaS platform to Token Hub, and in May the three major telecom operators collectively launched Token packages. The stakes for the entry point had long been laid.

It would be too shallow to see only promotions. On the surface, the platform added a new category; in the middle layer, vendors are seeking stable cash flow—reliance on government and enterprise projects makes revenue structure volatile, and C-end subscriptions are another leg; further down, it is a defense. The ultimate form of AI Agents is to check information, compare prices, and place orders for users, without going through browsing and search at all. Once AI assistants become the new transaction entry point, the shelf value that traditional e-commerce prides itself on will be hollowed out entirely. Si Dengyu, general manager of Tmall 3C Digital, put it plainly: the goal is to "let consumers complete consumption like recharging phone bills or buying memberships," and the next step is to graft Token onto AI PCs and AI phones—"without Token, these are just concepts."

Therefore, moving tokens into one's own recharge system and placing AI office memberships in one's own livestream is essentially keeping the Agent's consumption behavior within one's own venue. Behind more volume at the same price, platforms are using concessions to buy time for user habit formation.

What the platform gives up is profit, in exchange for habits; what vendors cut is price, in exchange for defaults; both are fighting for the same thing—whose page users open when they pay for AI for the first time.

Three accounts being calculated simultaneously: C-end enlightenment, B-end volume growth, and rising unit prices

Beyond the bustle, three accounts are being calculated at the same time, and their directions are not consistent.

The first is the enlightenment account for ordinary people. 1,000 yuan in transaction value, an activation rate awaiting verification, and a renewal rate after one month. Computing power is not a data package, and stockpiling tokens is not stockpiling tissues: unused data can be carried over, but tokens stockpiled and unused are sunk costs, and their value depends entirely on how you use them next. Promotions can push sales numbers high, but they cannot answer the three questions of retention, paid conversion, and real usage scenarios.

The second is the production account calculated for enterprises. Qiao Zhongliang, founder of Xiaoyu Zhizao, calculated: at the beginning of the year, the whole company consumed about 4 billion tokens per month; now he alone consumes 150 billion per month, running four or five agents autonomously at the same time, with computing power costs up 10 times, and about 80% of the company's spending going to AI computing power and data. Closer to the ground is another sample: a small company in Chongqing spends more than 2,000 yuan per month on Tokens and this time bought 12 monthly cards at once. The enterprise side has already taken tokens as means of production—as long as the math works, consumption will grow on its own. Digital economy scholar Liu Xingliang's reminder is sober: spending more does not equal good application, and a large call volume does not equal high profit; what really matters is how much it costs to complete a qualified task, how much time is saved, and how much revenue is created.

The third is the account most easily obscured by "price cuts." Data from the China Academy of Information and Communications Technology shows that in June 2026, domestic average daily token consumption was about 500 trillion, nearly four times that of December 2025; while the normalized average price rose from 6.8 yuan to 13.46 yuan over the same period, nearly doubling.

Put these two numbers together, and the conclusion reverses: what is cheap is the entry package, not tokens themselves; what falls is the customer acquisition price, and what rises is the unit value of real usage. The so-called "Tokens are becoming less and less valuable" is an illusion created by the shelf. Precisely for this reason, the draft's statement that "marginal cost has been compressed to negligible" must be retracted—every invocation consumes real computing power, and computing power has marginal costs; it has only been pushed low enough to support 1-cent user acquisition, and has never gone to zero.

Beyond the three accounts, there is also a supporting hand. On September 18, the Beijing Municipal Bureau of Economy and Information Technology and the Municipal Development and Reform Commission jointly issued the "Action Plan for Accelerating the Development of the Token Economy in Beijing (2026–2028)," the country's first provincial-level special policy for the token economy, launching ten support policies around six dimensions: token production, pricing and circulation, scenario application, international cooperation, data factor utilization, and factor guarantees. The official forecast is that the entire industrial chain is expected to exceed a scale of one trillion yuan during the "15th Five-Year Plan" period. The cost-reduction path is also specific: model adaptation to domestic chips, lightweighting and inference optimization to raise token output per unit of computing power; unified interfaces and public distribution platforms to save enterprises repeated adaptation expenses; special financial products and medium- and long-term credit to ease funding pressure from early investment. The supply side is expanding simultaneously—Beijing Yizhuang has built the city's largest 5,000P public computing power platform, and the 10,000P domestic computing power center has listed its first batch of computing power. The Ministry of Industry and Information Technology is exploring "computing power banks" and "computing power supermarkets," allowing idle computing power to be deposited and withdrawn and scheduled across regions and cycles, pushing supply from "subsidy-style transfusion" toward "market-based water-and-electricity-style provision."

One is the enlightenment account, one is the production account, and one is the industrial backstop cost-reduction account. Only when all three are calculated simultaneously can the token economy truly move from concept into the financial statements.

A ruler not yet unified: from invocation volume to gross profit per unit token

Capital has already begun pricing this chain, and the positions of three companies happen to form a complete puzzle.

Xingyun Technology entered "cross-border digital trade + AI computing power infrastructure," directly writing "token revenue sharing" clauses into computing power contracts, and signed a framework cooperation agreement with Volcano Engine: revenue in the first three quarters of 2026 was 1.175 billion yuan, up 1,896.66% year-on-year, and net profit attributable to parent was 269 million yuan, turning losses into profits; in the third quarter alone, revenue was 921 million yuan, up 5,556.94% year-on-year, and net profit attributable to parent was 257 million yuan. Xiechuang Data launched a token sharing platform, focusing on token factories and computing power scheduling, with net profit attributable to parent in the first three quarters up more than 338% year-on-year. iFlytek, meanwhile, saw continued volume growth in C-end smart office products, and expanding personal user scale drove steady increases in token invocations; third-quarter revenue was 6.564 billion yuan, up 7.99% year-on-year, and net profit attributable to parent was 226 million yuan, up 31.05% year-on-year.

Xingyun Technology is a computing power provider, Xiechuang Data is a computing power distributor, and iFlytek is a scenario consumer. Tokens are no longer just a concept in PPTs; they have become revenue-sharing clauses in contracts and net profit on financial statements.

But the real watershed is not in these numbers, but in a more fundamental question: Tokens are not standard products.

Cui Lili put it directly: "Tokens are obviously not standard products." Statistical standards differ markedly among vendors—input, output, context caching, and image parsing each have their own algorithms; some bill by characters, some convert by points; benefit scopes are also inconsistent. For the same 39-yuan monthly card, the actual usable service volume differs considerably, and price alone makes it impossible to compare cost-effectiveness. The "Price Behavior Rules for Internet Platforms" require operators to indicate service items, service content, and prices or pricing methods; digital goods have no physical form, so transparency is the core foundation of trust. After-sales is an obvious pitfall: how to refund unused recharge, how to compare usage across different models, what to do when customer service cannot answer, and even split resale may emerge.

The significance of policy must also be stated accurately. What it aims to do is turn tokens from a "money-burning cost item" into a production factor that can be priced, measured, and circulated—this is providing a practicable path, not completed rights confirmation, and certainly not equivalent to placing it on the same established footing as land and capital. Saying one sentence less is safer than saying one sentence more.

Thus, the metrics for measuring this business are switching. Chen Duan's reminder is worth copying down: to judge fundamentals, one should not only watch token invocation volume, but also track customer retention and gross profit per unit token, and distinguish concept-driven market moves from real performance. From invocation volume to activation rate, renewal rate, cost per unit task, and gross profit per unit token—traffic gained through money-burning subsidies is only the starting point, and between concept-driven market moves and real performance lies a river called "commercialization verification."

Listing solves "can buy," the entry point solves "can find," renewal solves "can retain," and the ruler solves "can calculate clearly." If any one of the four is missing, 2 trillion is just a pretty number.

Looking past the hype, the chain of events is actually quite clear: 1 cent completed the enlightenment, two months completed the positioning, three sets of accounts were calculated simultaneously, and in the end it got stuck on a ruler that still has no markings. Platforms win at the entry point, manufacturers win by default, users win on cheapness—but only the party that solidifies the measurement standards, retention, and unit costs wins in the end.

The first step toward popularizing AI computing power is not making it affordable for everyone, but making it calculable for everyone.

Next Double 11, there will probably still be Tokens on the shelves. Only by then, what is more worth looking at than "how much of a discount" is whether that ruler has markings.

References:

China Fund News: C-end subscriptions and stable cash flow

China Economic Net: For the first time placed on the "Double 11" shelves

Guancha.cn: Double 11 is quiet, selling tokens is lively

The Paper: The new commercial war of large models has reached Double 11

WeChat: Alibaba has cut prices again, are Tokens becoming less and less valuable?

China Newsweek: Top streamer sold "2 trillion" in half an hour

Dazhong Net/Red Star Capital Bureau: "Stockpiling Tokens" becomes a trending phrase for "Double 11"

Beijing Municipal People's Government: Computing power bills remain high; how can the token economy reduce costs

China Economic Net: Double 11 livestream token hot sales; listed companies increase bets on the computing power economy

Securities Times: Tmall Double 11 includes Token as an independent category in the major promotion for the first time

CNR: Li Jiaqi sells AI computing power for the first time, selling 2 trillion Tokens in half an hour

Jiemian News: Double 11 Token rush: AI relies on discounts, but cannot trade for a commercial future

Alibaba Cloud Token Plan significantly raises quotas, with three tiers increased by over 70% while monthly fees remain unchanged

Toutiao: The 0.01 yuan trial package sold 100,000 units in half an hour! "Stockpiling AI computing power" becomes a new consumer trend for "Double 11"

- END -

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.