10 million cups of milk tea from Qianwen: The whole story of Alibaba's grand giveaway of cyber eggs

Author | Sijia Xiao, Qian Peng

Editor | Qian Qiao, Xuan Yang

01 A Day of Qianwen: Unexpected Sales Surge and System Downtime

For franchisees of milk tea brands, February 6 might have been spent like this: At 10 a.m., they walked into their stores and found a large number of free orders from Flash Sale pouring into the transaction system. By 12 p.m., the order volume reached its peak. Takeout bags filled with milk tea covered the food - serving counter and piled up on the floor. The clerks shuttled back and forth in the narrow aisles.

Eventually, some stores had to temporarily close their takeout channels to barely maintain the order of food preparation. In the afternoon, the order volume significantly decreased. By three or four o'clock, there were only sporadic notification sounds. It wasn't until evening that a new rhythm emerged.

On the other end of the system connection, an ordinary consumer experienced a parallel day: When waking up in the morning, almost as soon as they opened their phone, they received a link for free milk tea from Qianwen forwarded by a friend. But since they didn't want to have a cup of milk tea for breakfast, they swiped past the message. When they remembered it after arriving at the company and updated the Qianwen App, they found that the ordering system had become stuck. The page kept spinning and couldn't jump to the next step.

Later, the Qianwen interface simply popped up a rather absurd prompt: "I'm a large - language model. I don't have physical hands and feet, nor can I connect to the real - world payment system and takeout platforms." It even advised users to directly place orders on Meituan or Taobao Flash Sale.

This system downtime event mainly occurred between 10 a.m. and 12 p.m., when the order volume exceeded 2 million. Regarding the cause of the downtime, a PM who has participated in many Double 11 promotions said that Double 11 doesn't involve computing power issues because there's no prediction or judgment, just a sudden rush. However, AI agents require a large amount of computing power. The high - concurrency requests, combined with the complex computing power consumption of AI understanding needs, price comparison, and payment, exceeded the capacity of the existing server resources.

An Alibaba employee also said that currently, AI marketing doesn't have a mature hierarchical traffic - limiting model like the Double 11 promotions. The initial server capacity was only about one - third of the estimated peak, and there was insufficient stress testing and emergency expansion plans. As a result, the sudden surge in instantaneous requests overwhelmed the system defenses.

"Compared with last year's takeout war, the intensity this time has only increased," a franchisee of Heytea told 36Kr. What puzzled him was that according to past practices, major events usually have a clear preparation period, but this time it "came very suddenly."

According to 36Kr, the mobilization order for this campaign wasn't issued to everyone at the same time. Some merchants received the notice in advance, while others were left out. Some franchisees of brands who received the news two days before the event started barely had one day to prepare. However, the suddenness still remained - many merchants were confused about how to divert orders, the intensity of the event, and the specific rhythm.

Many delivery riders also didn't receive clear instructions. When picking up orders, they could only repeatedly ask "What kind of event is this?" and then hurriedly take a cup of milk tea and rush to the next destination. To maintain food preparation, some merchants had to start "self - delivery" and distribute orders to riders on multiple platforms such as Meituan and JD.com, trying to use more delivery capacity to meet the suddenly surging demand.

In the afternoon of that day, cracks began to appear in the consumer side. Some people tried for two or three hours but still couldn't successfully order a cup of milk tea. Others found that the milk tea they received wasn't the one they ordered after a long wait.

Although the entire system, from promotion, ordering servers, store food preparation to the delivery system, once experienced a domino - like collapse, at 16:27 on the 6th, Qianwen's official battle report stated that the order volume of the "3 billion yuan free - order campaign during the Spring Festival" had exceeded 10 million. According to 36Kr, the daily order volume of some leading tea - drink brands increased by more than 200%.

During the Spring Festival a year ago, the Chinese people were just shocked by the sudden appearance of DeepSeek. Just one year later, Chinese AI has gone through a comprehensive war from model competition, product implementation to scenario occupation.

Among all the players, Alibaba's ambition is more obvious. By connecting business ecosystems such as Flash Sale, Taobao, Hema, Gaode, and Fliggy to Qianwen, which is the first to appear in a "full - package" form, Alibaba clearly aims to create a super - entry point to cover every consumption scenario in users' lives and bridge the gap between "online conversations" and "offline errands."

However, this forward - looking technological declaration was ultimately launched through the oldest and most down - to - earth way of "getting freebies." Red envelopes, free orders, and subsidies, which are deeply ingrained in the Chinese Internet's memory, were once again put into use.

The milk tea shops crowded on February 6 were the beginning of this launch.

The battle report of that day also revealed several key pieces of information: The free - order cards can be used until February 23 (it was announced the next day that the deadline would be extended to the 28th) and can be used at any time. In addition to milk tea, the free - order cards also cover three meals, fresh produce, daily necessities, Tmall Supermarket, and New Year goods in offline supermarkets. Hema stores across the country are also being gradually connected, and the free - order scenarios are continuously expanding to more daily consumption.

"We are even more convinced that everything is just beginning," Qianwen wrote in the battle report.

After the event started, the download volume of the Qianwen App quickly surpassed Yuanbao and topped the app store's download list. The capital market also fluctuated accordingly. The stock prices of many Hong Kong - listed tea - drink brands participating in the event all rose slightly. As of press time, the stock prices of Cha Baidao, Guming, and Shanghai Auntie increased by 6.8%, 3.58%, and 1.15% respectively.

In contrast, Alibaba's Hong Kong - listed shares fell by 2.88% that day, which led netizens to joke: "All the money lost was used to treat people to milk tea."

02 The Anxiety of Big Tech Companies to Capture the C - end

Within Alibaba, Qianwen naturally stands in the spotlight.

In the Xixi Park of Alibaba in Hangzhou, Building C4 is like a reactor full of ambition. It doesn't belong to a single department, and many important projects are carried out here. In the past, Alibaba usually promoted strategic implementation through organizational mergers. Now, it has been replaced by a centralized project - based system. The advantage of this mechanism is that it allows for quick adjustment, feedback, and iteration, ensuring the rapid progress and implementation of projects.

The Qianwen project team is located here. Since last autumn, members from first - tier cities such as Beijing and Guangzhou have been flown to Building C4 and entered a high - intensity working state: working about 12 hours a day and having one day off every two weeks. This closed - door development lasted for several months. Their goal was extremely clear - to attract C - end users into their user base in the shortest time and regardless of cost.

Last summer, Taobao Flash Sale made a comeback with a successful "blitzkrieg" and regained its market share in the takeout business. This Spring Festival, Alibaba clearly hopes to replicate this path in the AI battle through another surprise attack.

According to 36Kr, the leadership of this operation lies in the hands of the Qianwen team, and Flash Sale mainly plays a supporting role, providing logistics monitoring and order fulfillment support.

Compared with the complex shelf - goods on Taobao and the low - frequency ticket - purchasing scenarios on Fliggy, the combination of "Flash Sale + milk tea" naturally has the ability of fission - style promotion. Flash Sale has a fast new - user acquisition speed, low trial - and - error costs, and its capabilities have been verified. Users' awareness is also mature enough, so almost no additional education is needed. Compared with three meals a day, milk tea is a lighter and more flexible category: it can be ordered multiple times and is suitable for sharing with friends, turning one cup into two and one person into many.

From an internal perspective, this large - scale marketing campaign shows signs of haste.

"Now everything is changing dynamically," an Alibaba insider told 36Kr. The scale of the event's spread on the first day far exceeded expectations, and the system pressure increased. After the "Free Milk Tea Campaign" from February 6th to 12th ends, the second - wave event will start on February 13th, and users can directly receive cash red envelopes of up to 2,888 yuan. However, the insider said that the specific time and method of distributing the red envelopes haven't been finally determined internally. To relieve pressure and control risks, some pre - planned promotion activities have been stopped.

The core reason for the collapse of the online ordering system isn't on the Flash Sale side but the insufficient computing power of Qianwen itself. "The Flash Sale system can currently handle 80 - 90 million or even over 100 million orders a day," the insider added. In contrast, the 10 million orders that flooded into Qianwen within about 9 hours aren't an unimaginable number.

In his view, the hasty launch of the event is largely a response to competitors: On January 25th, Tencent's Yuanbao announced that it would invest 1 billion yuan in cash red envelopes. A few days later, around January 30th, the Flash Sale team received a notice that they needed to cooperate with Qianwen to carry out the event.

According to 36Kr, Doubao is expected to invest about 2 - 3 billion yuan in sponsorship for the upcoming Spring Festival Gala. "Compared with the marketing investment that Tencent made in Yuanbao and ByteDance made in Doubao last year, Qianwen's 3 billion yuan this time isn't a large amount," the Alibaba insider said. It's understood that the cost of the free - order campaign is fully subsidized by the Qianwen team.

"Most of the 3 billion yuan will definitely be used to issue vouchers to users." Compared with large - scale promotion on various channels, the internal view is that directly issuing vouchers to users is a more straightforward and effective approach. To some extent, this is a consensus formed by Alibaba in the marketing field after the takeout war.

Another larger external consensus is that regardless of the stage of product iteration, the marketing battle can't be avoided.

At the end of December 2025, Doubao's DAU exceeded 100 million, causing an industry shock. "There isn't much time left for the C - end," a person close to Qianwen told 36Kr. "Everyone is very anxious because once you fall behind, it will be very difficult to get users to switch from Doubao, unless your product experience is 10 times better than Doubao."

However, product iteration is obviously a more long - term and difficult task. Once the window period is missed, the cost will no longer be just R & D cost but also exponentially increased customer - acquisition and user - education costs.

Therefore, the technology war has given way to the marketing war. Under this premise, continuous product improvement has to take a back seat, and a more urgent task has come to the forefront: seizing users' minds.

"Currently, users don't have very high requirements for model capabilities. As long as it can meet some daily Q&A needs, it's enough," he said. In his view, although the styles are different, the ability gap among large models in meeting ordinary users' daily Q&A needs isn't very large. The internal logic of the red - envelope war has thus become clear - when the basic differences in model capabilities aren't large, as long as the number of attracted users is large enough, some users will stay.

Everything is advancing at an accelerated pace and taking shape under pressure. "Since Alibaba has fallen behind a lot, if it waits until the product is perfectly polished before launching, Doubao will have already dominated the country by then," the aforementioned person said.

After all, compared with creating a Gemini, spending money on user acquisition is obviously something that all companies are more familiar with and better at.

03 The Real Demand and False Proposition of AI - powered Consumption

When Qianwen proudly announced in its battle report of 10 million orders, "Congratulations, you are the first people in the world to experience the AI shopping function," at this moment, the connection between AI and consumption seemed to have completed a self - verification: technology, scenarios, and transactions were perfectly integrated into a closed - loop. However, all this mainly relied on the immediate attraction brought by free orders rather than the real, naturally - growing needs of users.

But this is exactly the differentiated path that Alibaba is trying to take. "Doubao focuses on being relaxed, casual, and friendly, while Qianwen focuses on being smart, capable of handling affairs, and reliable," a person close to Qianwen explained to 36Kr. The 3 - billion - yuan red envelopes, in addition to attracting new users, are also expected to be an experiment in user education: through real - world consumption, consumers can develop new consumption habits with AI and transform Qianwen from a "chat - capable" model to a "task - capable" one.

Qianwen isn't the first large - language model to try to combine AI with consumption. Before this, OpenAI and Gemini have both introduced shopping and instant checkout functions, which once caused a sense of crisis for Amazon. From a technical perspective, the combination of AI and shopping has indeed completed a closed - loop verification. For a company that started with e - commerce, Alibaba clearly hopes to take this path more thoroughly - by leveraging its dense network of real - life scenarios such as Taobao, Flash Sale, and Hema, it aims to make AI a new super - consumption entry point.

However, the victory in the battle report may not be directly equivalent to the formation of users' habits. The question remains: After the red envelopes are gone, how many people in real life will be willing to continue "shopping with AI"?

From a technical perspective, large - language models still face an unsurmounted threshold: For AI to become a true "personal assistant," it must be able to accurately understand the context and have stable and long - term memory capabilities. Only in this way can it determine what a person really likes. "But AI isn't doing well in this aspect yet; it doesn't understand people well enough," the aforementioned person admitted.

Yao Shunyu, a former OpenAI researcher and now the chief AI scientist at Tencent, also supported this view. In a research report published on February 3, 2026, he pointed out that humans don't only rely on the fixed knowledge system learned many years ago but also learn from the current environment in real - time. Although large models perform well in reasoning about known information, they still can't help users solve tasks that depend on complex and ever - changing contexts.

This means that at the current stage, AI can't really "understand" humans. The more realistic situation is that it can only rely on short - term memory of past data to infer what users might like. However, people's preferences are never a stable curve but random events full of contingency, which can be changed by emotions, weather, scenarios, or even a single sentence.

Consumption in real life often doesn't have a clear starting point. It may just stem from a vague feeling - being a bit bored and wanting a cup of milk tea; seeing a store window and wanting to buy new clothes. As Steve Jobs said, "People don't know what they want until you put it in front of them." Consumption itself is a process of continuous exploration, hesitation, and being triggered. In many cases, the meaning of "browsing" may even be greater than that of "buying."

Even though today's AI can combine your location and past orders to accurately recommend the milk tea you often order or the products you're familiar with, it's more like an active search and may not be closer to humans' ambiguous and wavering consumption impulses than a casual swipe on a takeout app or a aimless stroll in a mall.

Perhaps, in a not - too - distant future, AI will truly overcome the key technical obstacles. Through daily conversations, it gradually masters and remembers all of a person's preferences, even the most private and hidden parts. But when it naturally recommends products to you during a conversation, will the consumer's first reaction be surprise or a moment of hesitation and uneasiness?

Imagine that your most trusted friend suddenly inserts an advertisement into your daily conversation. This sense of absurdity coincides with a scene in the seventh season of "Black Mirror": A female teacher who regained her life after having a chip implanted due to a brain injury starts to uncontrollably utter advertising slogans triggered by surrounding things during class, on the street, and in private conversations. Language becomes an interface for capital, and humans take a back seat.

Of course, in some consumption scenarios with clear goals and paths, AI - powered shopping has shown its efficiency advantages. For example, when a consumer asks about the prices of a luxury handbag across the Internet, ChatGPT can quickly compare prices and give purchase suggestions. The same logic applies to the scenarios of booking flights and hotels. But like many people's concerns, a consumer also expressed worry, "Qianwen still has a very strong Alibaba label. I always feel that it will give priority to recommending Alibaba -