At 7:30 p.m. on November 2, 2026, Chen Yusen, Vice President of Alibaba Group and CEO of QwenWork, stepped into Li Jiaqi’s Double 11 shopping livestream room. The featured product of the evening was a monthly Personal Basic membership card for QwenWork—officially priced at 28 yuan, but offered to new users in the livestream for just 0.01 yuan (1 fen), limited to 1 per account. Sina Tech’s livestream sales report and Beijing Business Today’s on-the-ground report cited data provided by the QwenWork team showing that nearly 100,000 monthly cards were sold within half an hour of going on sale. Roughly converted by the team based on the theoretical upper limit, this batch of monthly cards amounted to a cumulative total of approximately 2 trillion tokens—the word-count-based billing unit used by large language models (LLMs)—with each card containing a compute quota of roughly 20 million tokens on average.
On the surface, this seemed to suggest that LLM compute had, much like the discounted phone credit and mobile data packs promoted on Taobao over a decade ago, become an everyday consumer good affordable to ordinary people. Yet the moment you cut to the livestream feed, an unusual phenomenon immediately emerged: in front of the camera, Li Jiaqi and Chen Yusen barely discussed any technical details. Instead, they spent a great deal of time demonstrating to viewers watching behind their screens how to snap a photo of a test paper so the software could tutor a child on homework, and how to generate a year-end performance review report with a single click. They also set out laptop computers in the livestream room as physical merchandise, promoting Lenovo, Lecoo, and Machenike.
Looking back two months ago, in the imagination of Alibaba Cloud and various vendors, tokens were supposed to be a standardized commodity aimed at the general public. In the eyes of cloud vendors, LLM compute—just like mobile data packs or utilities such as water, electricity, and gas in earlier years—is backed by asset-heavy data centers on the backend and sold as a standardized, pay-as-you-go digital good on the front end, while society-wide demand for API calls is exploding. Statistics from the China Academy of Information and Communications Technology (CAICT) cited by 36Kr showed that in June of this year, China’s domestic average daily token consumption reached approximately 500 trillion—nearly 4 times the level in December of last year—while the normalized average domestic token price also rose from 6.8 yuan to 13.46 yuan. In an interview, Chen Yusen also drew an analogy to the Jevons paradox in economics, pointing out that once the price per call drops and processing efficiency improves, enterprises and merchants actually multiply their total consumption.
Over a decade ago, when the mobile internet was just taking off, Taobao successfully cultivated the habit of buying virtual goods online among hundreds of millions of Chinese consumers by selling discounted phone top-up cards and mobile data packs. This autumn, Alibaba attempted to transplant this proven retail playbook intact onto the LLM track. On September 3, a report on Tmall launching its Compute Top-Up Center in early September confirmed that Taobao’s top-up section had officially opened a dedicated compute shelf, onboarding an initial batch of subscription plans and top-up cards from vendors including Alibaba Cloud Bailian, QwenWork, Zhipu AI, Moonshot AI, DeepSeek, and StepFun. The platform’s promotional slogan directly called on users to buy LLM compute just like topping up phone credit or buying mobile data packs—adding items to their shopping carts to compare prices before placing an order. At the opening ceremony of the Apsara Conference on September 22, Alibaba Group CEO Eddie Wu (Wu Yongming) delivered a speech predicting that the total volume of machine-supplied thinking in the future would exceed human thinking by more than 1,000 times. Likening tokens to electricity in 1882, he proposed that the industry’s next core questions are who will produce tokens, who will transmit tokens, and which devices and applications will ultimately consume them.
By mid-October, Liu Bo, Senior Vice President of Alibaba Group and President of Tmall, announced that Tmall Double 11 would establish Artificial Intelligence Services as a first-tier category for the first time. Data from the first day of sales brought the team a huge surprise: gross merchandise volume (GMV) for the category surged 4.5 times period-over-period, as large numbers of e-commerce sellers, one-person companies, and academic or research users with existing practical needs flocked to the shelves to stock up on compute. Extrapolating along this steep growth curve, packaging 20 million tokens into a 0.01-yuan blockbuster monthly card and placing it into Li Jiaqi’s livestream room should have—just like the Double 11 rush for discounted data packs years ago—instantly ignited a stockpiling frenzy among tens of millions of ordinary consumers.
However, once this phone-top-up business logic stepped outside the small circle of programmers and e-commerce sellers and was placed directly in front of ordinary consumers in Li Jiaqi’s livestream room, it immediately ran into three real-world obstacles. Those rushing to buy compute on the first day of Tmall Double 11 were mostly mature users like women’s shoe seller Liu Xianglong. In his daily operations, Liu Xianglong was already using the E-Commerce Edition of QwenWork to process store backend data and have the software automatically generate business dashboards for hundreds of women’s shoe styles; he knew full well how much manual labor time each unit of compute saved him. Li Jiaqi’s livestream room, by contrast, caters to tens of millions of ordinary consumers with no programming or prompting experience. The reason everyone buys mobile data is that data serves as the toll fee for everyday entertainment consumption. When you open a short-video app to scroll through funny skits or open a video streaming site to binge-watch dramas, the content itself delivers instant gratification without requiring any user skill; data runs out imperceptibly in the background on your phone, and everyone knows clearly how many days of binge-watching 10GB of data will last them. LLM tokens, however, are raw flour meant for doing work, not freshly baked bread served on the table. If you hand an ordinary office worker 20 million tokens, facing a blank chat input box, they neither know what instructions to issue nor how to break down their daily work into steps a machine can understand—leaving that compute quota sitting in their account without even the first cent being spent.
Beyond the barrier to use, another problem with tokens is the lack of a measurement standard that ordinary people can understand. Wang Jiandong, Deputy Director of the Big Data Development Department at the State Information Center, pointed out that when people buy tissues at the supermarket, they look at how much 1 sheet costs; when buying milk, they look at how much 1 milliliter costs; and when buying mobile data, they look at how much 1GB costs. In the LLM market, however, different models consume vastly different numbers of tokens for the exact same document, every vendor has its own set of credit rules, and some plans include expiration-and-reset clauses—so when ordinary consumers see “20 million tokens” written on a poster, they have no way of figuring out how many resumes it can revise or how many presentation slides it can make. From the supply side, the cost structure of LLM tokens is also very different from that of telecom carriers. Ding Ding, Deputy General Manager of China Unicom Data Intelligence Co., Ltd., analyzed that the marginal cost of mobile data is nearly zero, and overselling plans at most causes fluctuations in network speed; yet every actual consumption of LLM tokens corresponds to GPU compute, electricity, and data center depreciation—meaning that once plans are oversold and redeemed en masse, the system faces queuing and rate limiting.
Precisely because raw tokens cannot be delivered directly as a mass-market commodity, the livestream product details verified by Phoenix New Media (Ifeng) reporters revealed the true nature of this 0.01-yuan monthly card. What buyers got for 0.01 yuan was not general-purpose spot compute that could be freely circulated in the developer market, exported via API keys, or resold. What they actually received were consumer-grade usage credits encapsulated within QwenWork’s membership system. This Personal Basic monthly card originally includes 500 base credits; after stacking livestream bonus perks and first-time registration and activation benefits, eligible new users could receive up to 900 credits, along with access to AI office courses and unlimited use of mainstream foundation models—though the backend restricts execution to a maximum of 5 concurrent tasks at a time.
As for the high-profile claim in the sales report of nearly 100,000 cards sold in half an hour—equivalent to a cumulative total of roughly 2 trillion tokens—that was a self-reported promotional figure calculated by the QwenWork team based on the theoretical upper limit. Calculated at the actual out-of-pocket price of 0.01 yuan, the direct cash flow generated by 100,000 monthly cards was only about 1,000 yuan. From the perspective of a cloud vendor’s financial model, this deal works on the exact same principle as a gym selling off cheap trial passes at the beginning of the year. Most viewers who place impulse orders will never come close to burning through all 20 million tokens, and unactivated quotas do not trigger GPU inference computation, meaning their marginal cost is zero; when the occasional user opens the app in the evening to try it out, their usage conveniently fills the off-peak idle compute capacity in Alibaba Cloud’s data centers at night.
Since ordinary consumers have no intuitive feel for 20 million tokens, why were nearly 100,000 orders still sold within half an hour that night? The answer lies in the stark contrast in narrative inside and outside the livestream room. External PR releases and sales posters continued to loudly trumpet “2 trillion in compute hitting the shopping links” to echo the group leadership’s compute narrative; yet once seated in front of the livestream cameras, Chen Yusen and Li Jiaqi—both born in 1992—immediately switched to a down-to-earth sales pitch. Throughout the broadcast, QwenWork was positioned directly as viewers’ “AI office buddy.” On camera, Chen Yusen used just a single sentence to liken tokens to the gasoline a car needs and the electricity household appliances run on, after which the two set aside all technical jargon and turned the livestream room into a live demo session on hiring an all-around tutor and personal secretary for 0.01 yuan. For the audience watching behind their screens, they demonstrated live how to take a photo of a child’s test paper for elementary, middle, and high school homework tutoring, how to generate year-end performance reviews and slide presentations with a single click, and even how to create travel itineraries, draw fishing maps, analyze livestream replay and bullet-comment data, and help viewers pick out PC builds based on their budgets. When consumers paid that 0.01 yuan, what they were buying was a hassle-free outcome: having their performance reviews written and their children’s homework graded for them.
Why would a group vice president in charge of enterprise-grade office software personally go to a mass-entertainment livestream room to pitch so hard? Opening the records of Alibaba’s organizational restructuring this summer reveals the growth pressure on Chen Yusen’s shoulders. Between June and July of this year, Chen Yusen, having just taken over DingTalk, swiftly ended Alibaba’s internal “horse-racing” friction. According to Sina Finance’s breakdown of the three-in-one integration of QwenWork, he merged and reorganized 3 previously siloed internal Alibaba agent products into a single unified brand: QwenWork. These 3 predecessor products were QoderWork, a desktop-grade local agent tool incubated by Alibaba Cloud; Wukong, a DingTalk enterprise workflow agent championed by former DingTalk head Wuzhao; and MuleRun, a cross-platform cloud agent execution engine already serving 43 countries and regions worldwide. On August 3, the integrated QwenWork launched its public beta, paired with the flagship model Qwen3.8-Max, which features 2.4 trillion total parameters and 95 billion active parameters.
Among those 3 predecessor products, Wukong—which focused on being embedded within DingTalk—had been shunned by internal employees due to issues such as accidentally deleting images when editing documents or displaying loading errors; instead, the desktop tool QoderWork delivered the most solid performance. When attending a team meeting in June of this year, Eddie Wu revealed that QoderWork ranked first among all AI tools across Alibaba Group in daily active users (DAU), weekly active users (WAU), and token consumption. People familiar with the matter disclosed that group executives including Eddie Wu still personally attend key meetings for QwenWork to this day. Although QwenWork surpassed 30 million users in the first month of its August public beta thanks to tech early adopters and e-commerce sellers, the existing user pool within that core circle quickly hit a ceiling. Chen Yusen urgently needed to leverage the Double 11 shopping festival and the nationwide traffic of Li Jiaqi’s livestream room to break out of that circle—using 0.01-yuan homework tutoring and performance reviews as a hook to pull ordinary office workers who don’t normally follow tech reviews into QwenWork’s growth pool.
Merely attracting users with a 0.01-yuan monthly card is far from completing the commercial loop. If users only try out homework tutoring twice in a mobile app or web chat box, then 30 days later—when faced with regular bills of 28 yuan per month for the Personal Basic plan, 78 to 158 yuan per month for the Advanced Personal plan, or 198 yuan per person per month for the Enterprise Standard plan—retention and renewal conversion rates will still plunge to freezing point. The core moat that truly makes office workers dependent on QwenWork comes from the desktop client capabilities it inherited from QoderWork. According to Alibaba Cloud’s official documentation on QwenWork Computer Use, once a user installs the QwenWork desktop client on their computer, the software can directly read and modify documents, spreadsheets, presentations, images, and video files on the local disk—eliminating the tedious steps of repeatedly uploading and downloading files manually on the web interface, and saving files back to their original paths with a single click once processing is complete. The desktop client also exclusively supports browser automation and computer-use features, enabling it to directly recognize and operate the user’s computer screen, click software buttons, fill out web forms, and organize data across local applications on the user’s behalf. However, a web chat box is constrained by the browser sandbox and cannot handle such heavy lifting; and if you expect an ordinary consumer to visit the official website on their own, download an installer, install it on their computer, and then grant local disk read/write and screen-control permissions step by step, the vast majority will drop off along the way.
Beyond the lengthy conversion funnel for the desktop client, sibling rivalry within Alibaba also kept Chen Yusen on edge. TMTPost’s investigation into the internal competition between the Qwen App and QwenWork revealed that barely 1 week after QwenWork launched its public beta on August 3, the Qwen App—operated under another consumer-facing business line within Alibaba—suddenly rolled out an office assistant feature between August 7 and August 10, along with 3 paid Pro membership tiers priced at 19 yuan, 49 yuan, and 128 yuan for recurring monthly subscriptions, colliding head-on with QwenWork in the office productivity arena. Statistics from third-party research firm QuestMobile showed that although the Qwen App’s monthly active users (MAU) reached as high as 167 million in June of this year, average monthly sessions per user stood at only 17.4, and average monthly time spent per user was just 22.9 minutes—far below ByteDance’s Doubao at 92.7 sessions and 183 minutes per user over the same period. Meanwhile, on the PC desktop front, Tencent’s WorkBuddy ranked first with 20.97 million monthly visits, flanked by encirclement from ByteDance’s Feishu (Lark). Squeezed at close quarters on mobile by both formidable external rivals and its own sibling product, QwenWork had to bypass the cumbersome download-and-install process and nail its client directly onto users’ computer desktops if it wanted to hold its ground.
How do you get ordinary consumers past the high hurdle of downloading software and granting local permissions? The approach chosen by Chen Yusen’s team was to go straight to PC manufacturers for factory pre-installation—just as Microsoft bundled its office software into every new PC in the early days. Just half a month before stepping into Li Jiaqi’s livestream room, Tmall Digital Business Head Cheng Yuan and QwenWork jointly launched the Tmall QwenWork AI PC Initiative on October 15 and 16, bringing on board Lenovo’s Zhaoyang government-and-enterprise series along with brands including Lecoo, ASUS, MSI, Machenike, and ThundeRobot to intensively roll out more than 70 laptop models factory-preinstalled with the QwenWork desktop client during Double 11—featuring the software built in right out of the box and bundled with up to 3 months of complimentary membership benefits. Traditional PC manufacturers urgently needed to clear finished-machine inventory during the Double 11 shopping festival, while also struggling with the fact that the “AI PCs” marketed across the industry lacked software selling points that consumers could intuitively perceive. Seizing the opportunity, Alibaba brought out Tmall Double 11’s traffic resources and prime product slots in Li Jiaqi’s livestream room, listing and promoting 3 laptops in the very same November 2 evening broadcast—the Lecoo Air, the Machenike Dawn 16S Youth Edition, and the Lenovo Zhaoyang Yue Plus QwenWork Edition—alongside QwenWork audio-transcription cards priced between 999 yuan and 1,299 yuan, plus desktop robots. Alibaba helped PC manufacturers sell merchandise and clear inventory, and in return, the manufacturers handed over factory desktop pre-installation rights across more than 70 new models.
When consumers buy these computers factory-preinstalled with QwenWork and boot them up, QwenWork automatically appears on the computer desktop, seamlessly taking over permissions for local file access and screen operation. Looking back at that Double 11 night and the livestream experiment that sold nearly 100,000 orders in a mere 30 minutes, that 0.01-yuan monthly card finally revealed its true commercial identity. LLM tokens may look like the mobile data packs of 10 years ago, but ultimately, raw compute cannot be sold directly to the general public. If cloud vendors want ordinary people to consume compute, on one hand they need to package it into visible, done-for-you deliverables like tutoring homework and writing performance reviews; on the other hand, they must use the opportunity of driving hardware sales for PC manufacturers to place the desktop client directly onto the computers users rely on every day.