Alibaba "stocks up", ByteDance "trains skills"
In the last week of May, two adjacent events in the AI industry laid out two stances of Chinese tech giants on the same table.
On May 11th, Alibaba's Qianwen App was fully integrated with Taobao. Users can chat and shop in Qianwen, and directly use AI for virtual try - ons, price comparison, and coupon grabbing in Taobao. Qianwen accessed Taobao's 4 billion product library and 20 - year shopping scenario data at once.
Nine days later, from May 20th to 21st, Alibaba Cloud held a summit at Xizi Hotel in Hangzhou. Wu Yongming upgraded the five - layer full - stack of "chip - cloud - model - inference - application". The self - developed chip Zhenwu M890, the flagship model Qwen3.7 - Max, the new entrance Qianwen Cloud, and Agentic Cloud were all launched in one go. At the end of the meeting, he said, "The capital expenditure in the next five years will far exceed the previous plan of 380 billion yuan in the past three years."
Three months ago, ByteDance's Jimeng AI released Seedance 2.0, which topped the Artificial Analysis Video Arena with an Elo score of 1269, surpassing Google Veo 3, OpenAI Sora 2, and Runway Gen - 4.5. Feng Ji (the author of "Black Myth: Wukong") publicly said it was the "strongest AI video model". Later, on May 27th, foreign media reported that ByteDance's capex ceiling in 2026 is 470 billion yuan (about 7 billion US dollars), and it could reach 10 billion US dollars if the environment is ideal. This is nearly three times the 2.5 billion US dollars in 2024.
Alibaba is building the "utilities" and "retail cash register" in the AI era, while ByteDance is creating the "Nobel - style laboratory" in the AI era.
One goes live in 0 days, and the other starts with a five - year plan.
Both are called AI strategies, but the routes are completely different.
Alibaba Puts AI into the Cash Register One by One
The biggest change in Alibaba this year is not in chips or models, but in the organization.
In March 2024, Han Xinyi, the CFO of Ant Group, became the president of Ant Group. On March 1st, 2025, he officially took over the position of CEO from Jing Xiandong, and Jing Xiandong focused on the chairman's work. After taking over, Han Xinyi launched three strategies: "AI First + Alipay Dual - Flywheel + Accelerated Globalization". Half a year later, Ant Group was split into four parts. Ant International, OceanBase, and Ant Digital Technology each established independent boards of directors and operated independently in the market. A few months later, on February 2nd, 2026, Han Xinyi sent an email to all employees, launching the "AI Credit" special incentive program. Teams and individuals with pioneering contributions in the AI field will receive an additional bonus on top of the original performance incentives.
The meaning of this series of actions is very straightforward: break the organization into manageable parts, aim the incentives at AI, and then start adding products.
What are the products?
Ant's AI payment: By the Spring Festival in February 2026, the number of payment transactions exceeded 120 million, and the number of users exceeded 100 million. It became the world's first AI - native payment product with both the number of payment transactions and the number of users exceeding 100 million. Ant's health assistant "Afu" has a monthly active user base of 30 million. Taotian connected with Qianwen App, and AI virtual try - ons, AI price comparison, and AI coupon grabbing have become consumers' shopping actions. It is rumored that medium - sized Taobao merchants in the internal test reported that they voluntarily lowered the prices of three SKUs within a week after the AI price comparison feature was launched. AI is not for merchants, but for consumers to get better deals from merchants.
What's more worth noting is the ACT protocol.
On January 16th, 2026, six business units including Alipay, Qianwen App, Taobao Flash Sale, Rokid, Damai, and Alibaba Cloud Bailian jointly released the "Agentic Commerce Trust Protocol", which is the trust infrastructure for "AI paying on behalf of users". It is rare in Alibaba's history for six BUs to jointly release a protocol. Two years ago, during Zhang Yong's tenure, Taotian and Alibaba Cloud even fought over data sharing. Now they can stand in the same press release for an AI protocol. This is the organizational transformation completed by Wu Yongming in one year.
The rewards of this organizational transformation are reflected in the financial reports.
Alibaba's Q4 revenue increased by 3%, and the external cloud revenue increased by 40%. The figure of external cloud revenue is crucial. It doesn't represent Alibaba's own consumption of cloud computing power, but others paying for Alibaba's computing power. A 40% increase in this curve means that Alibaba's infrastructure investment has a cash - flow return channel. Liu Weiguang, a senior vice - president of Alibaba Cloud, said at the summit, "We are building China's largest AI factory." The core customers of this factory are Yuezhianmian, MiniMax, Kimi, Zhipu, and also DeepSeek.
On which cloud do domestic large - scale models run in China? A significant portion runs on Alibaba Cloud.
MaaS revenue is about to replace ECS as Alibaba Cloud's largest product line, which means that Alibaba Cloud's growth engine has shifted from traditional cloud computing to AI services.
Wu Yongming's exact words: "Now, there is hardly an empty card in Alibaba's servers."
This statement is very powerful. It's not just a CEO's bold claim, but a promise from a listed company to the capital market: every card burned with capex is generating revenue.
But the cost also needs to be mentioned.
Alibaba can "add products" like this on the premise that the model is just good enough. It doesn't have to be the strongest in the world, as long as it can be used for business monetization. Qwen3.7 - Max is close to the capabilities of DeepSeek and Kimi, but there is no significant gap. In the academic influence of international AI basic model original research, Alibaba is relatively low - key. Although the open - source version of Qwen has a high download volume on HuggingFace, Alibaba's contribution to papers on "how the next - generation architecture should develop" is far less than that of ByteDance. If one day the gap in basic models is widened by five times by ByteDance, OpenAI, or Anthropic, all the AI put into the cash register today will become old hardware that needs to be upgraded.
Alibaba is betting that the gap in basic model capabilities will not widen by five times in five years.
ByteDance Keeps AI in the Seed Department
ByteDance takes a different approach.
There are two parallel lines in the Seed department. One is Zhu Wenjia, who is in charge of model applications, including products like Doubao, Jimeng, and Kouzi. The other is Wu Yonghui, who is responsible for AI basic research and exploration, specifically the AGI route. When they first appeared on the same stage at the all - employee meeting, the only goal they set was: "The most important goal of the Seed department is to explore the upper limit of intelligence."
They also showed a rarer attitude: "Considering promoting open - source."
Among domestic tech giants, the word "open - source" is usually only mentioned repeatedly in the Linux era. ByteDance's willingness to mention open - source in the AGI era means that it no longer expects to charge for basic models. What it wants to do is to make the basic model the global technological infrastructure itself.
The external evidence is Seedance 2.0.
This video - generation model, released on February 10th, 2026, topped the Artificial Analysis Video Arena with an Elo score of 1269, surpassing Google Veo 3, OpenAI Sora 2, and Runway Gen - 4.5. It uses a dual - branch diffusion transformer architecture to achieve native multi - modality, unifying the processing of four types of inputs: text, images, audio, and video. It can generate a movie - level multi - shot video with native audio in 60 seconds, and the 2K video generation speed is 30% faster than similar models. Feng Ji's public evaluation was not a PR article but a post on his Weibo. It was the judgment of a game producer after using the model.
The internal evidence is even stronger.
The Top Seed talent program was launched in May 2024, targeting fresh doctoral graduates. In July of the same year, it was extended to research interns among doctoral students. ByteDance offers a daily salary of 2000 yuan to recruit talented young people, competing openly with DeepSeek in Silicon Valley and Tsinghua University. It is rumored that a former ByteDance Seed intern once mentioned at the dinner table that on his first day in the group, his KPI was not set as DAU or revenue, but "to rank in the top three on a certain international list by the end of the year". He had never seen such a KPI in other companies he had worked for.
There is also the eight - month paper.
The Doubao large - model team spent eight months conducting a systematic experiment titled "How Far Are Video - Generation Models from World Models?" The conclusion is quite modest: "Video - generation models can remember training cases, but they cannot truly understand physical laws for the time being." This kind of paper has no commercial conversion and is a pure academic answer to a question that will take 5 - 7 years to materialize.
How much does ByteDance spend on these things every year?
In December 2025, the Financial Times reported that ByteDance's capex plan for 2026 was 160 billion yuan. On May 9th, the South China Morning Post reported over 200 billion yuan (a 25% increase). On May 27th, Bloomberg reported a maximum of 470 billion yuan (7 billion US dollars). The figure was revised upwards three times within five months, each time with a significant jump. The latest figure is 2.8 times that of 2024. According to Bloomberg, the funds come from ByteDance's profit of about 50 billion US dollars in 2025. ByteDance internally has reservations about the accuracy of this figure, which means that after using up this year's profit, it will need to borrow an additional 20 billion US dollars. If the environment is ideal, it could reach 100 billion US dollars (about 678.1 billion yuan).
Does ByteDance have enough money to burn?
Yes. It can afford it because it is not a listed company and doesn't need to report ROI to the capital market every quarter.
But does it still have enough patience?
Not necessarily. Doubao has just started testing paid services. The title of a Titanium Media article is "ByteDance Hits the Brakes on Doubao's Free Service". Doubao's DAU has exceeded 100 million, making it the product with the least promotion cost in ByteDance's history to reach this milestone. It should be the least pressured to monetize, but the company is still considering inserting ads. This shows that even for a non - listed company, after three years of burning money, the financial pressure will start to mount.
It is rumored that a VC that invested in Doubao in the early stage privately commented that the biggest change in ByteDance in the past two years is not that the model has become stronger, but that it really starts to believe that "technological leadership can attract money on its own". This kind of belief was not in the company's dictionary before.
But belief is one thing, and it can't replace practical results. It's one thing for Seedance 2.0 to top the list, and another thing to turn Seedance 2.0 into the next Douyin. ByteDance has not proven the latter yet.
"Selling Goods" and "Making Products" Are Not Philosophies, but Origins
At this point, a counter - consensus is needed.
The mainstream narrative is that Alibaba is short - term practical, while ByteDance is long - term idealistic; Alibaba is more shrewd, and ByteDance has a longer - term vision. This view has some truth. Alibaba's 40% increase in external cloud revenue is real, and ByteDance's Seedance 2