In the last week of May, two adjacent events in the AI industry laid out the two different postures of Chinese tech giants on the same card table.
On May 11th, Alibaba's Qianwen App and Taobao were fully integrated—users can chat and shop within Qianwen, or directly invoke AI features for virtual try-on, price comparison, and coupon hunting within Taobao. Qianwen gained access to Taobao's 4-billion-item product library and 20 years of e-commerce scenario data.
Nine days later, from May 20th to 21st, Alibaba Cloud held a summit at Hangzhou's Xizi Hotel. Wu Yongming upgraded the entire "chip-cloud-model-inference-application" five-layer stack in one go, launching the self-developed Zhenwu M890 chip, the flagship model Qwen3.7-Max, the new entry point Qianwen Cloud, and the Agentic Cloud. At the end of the conference, he said, "Capital expenditure over the next five years will far exceed the previous plan of 380 billion yuan."
Rewinding to 3 months earlier, ByteDance's Jiemeng AI released Seedance 2.0, which topped the Artificial Analysis Video Arena benchmark with an Elo score of 1269, surpassing Google Veo 3, OpenAI Sora 2, and Runway Gen-4.5. Feng Ji (creator of Black Myth: Wukong) publicly called it the "strongest AI video model." Looking further ahead, on May 27th, foreign media reported that ByteDance's capex ceiling for 2026 is 470 billion RMB (approx. $70 billion USD), potentially reaching $100 billion USD under ideal conditions—nearly 3 times the $25 billion USD in 2024.
Alibaba is building the "water, electricity, and gas" and the "retail checkout counter" of the AI era, while ByteDance is building the "Nobel Prize laboratory" of the AI era.
One aims for immediate deployment, the other is built on a 5+ year horizon.
Both are called AI strategies, but their paths are completely different.
Alibaba Loaded AI, Piece by Piece, into the Checkout Counter
The biggest change at Alibaba this year isn't in chips or models, but in organization.
In March 2024, Ant Group's CFO Han Xinyi became Ant Group's President; on March 1st, 2025, he officially took over the CEO role from Jing Xiandong, who focused on his Chairman duties. After taking over, Han Xinyi launched three strategies: "AI First + Alipay Dual Flywheel + Accelerated Globalization." Half a year later, Ant Group split into four independent entities—Ant International, OceanBase, and Ant Digital Technology—each with its own board and operating independently in the market. A few months later, on February 2nd, 2026, Han Xinyi sent a company-wide email announcing the "AI Credit" special incentive scheme—teams and individuals making groundbreaking contributions in AI would receive an additional bonus on top of their regular performance incentives.
The meaning of this series of moves is very clear: break down the organization until it can run fast, align incentives with AI, and then start stocking the shelves.
What exactly is being stocked?
Ant's AI Payment—by the Spring Festival of February 2026, transaction volume exceeded 120 million, and user numbers surpassed 100 million, making it the world's first AI-native payment product to achieve both milestones. Ant's health assistant "Afu" reached 30 million MAU. Taotian integrated with the Qianwen App, turning AI try-on, price comparison, and coupon hunting into consumer shopping actions. Industry rumors suggest that mid-sized Taobao merchants in internal testing reported that after AI price comparison went live for a week, they proactively lowered prices on three SKUs—AI isn't for merchants; it's for consumers to get the best deals from merchants.
Even more noteworthy is the ACT Protocol.
On January 16th, 2026, six business units—Alipay, Qianwen App, Taobao Flash Sales, Rokid, Damai, and Alibaba Cloud Bailian—jointly released the "Agentic Commerce Trust Protocol," building trust infrastructure for "AI spending money on behalf of users." It's rare in Alibaba's history for six BUs to jointly release a protocol. Two years ago, during Zhang Yong's era, Taotian and Alibaba Cloud fought even over data sharing; now they stand together in the same press release for an AI protocol—this is the organizational surgery Wu Yongming completed in one year.
The return on this organizational surgery is in the financial reports.
Alibaba's Q4 revenue grew +3%, while Cloud external revenue grew +40%. This Cloud external revenue number is key—it represents not Alibaba consuming its own cloud compute, but others paying for Alibaba's compute power. A +40% curve means Alibaba's infrastructure investment has a cash flow channel for payback. Alibaba Cloud SVP Liu Weiguang said at the summit "building China's largest AI factory," whose core customers are Moonshot, MiniMax, Kimi, Zhipu—and also include DeepSeek.
Which cloud do domestic large language models run on in China? A significant portion run on Alibaba Cloud.
MaaS revenue is about to replace ECS as Alibaba Cloud's largest product line—this means Alibaba Cloud's growth engine has already switched from traditional cloud computing to AI services.
Wu Yongming's exact words: "Currently, there is almost not a single empty GPU card in Alibaba's servers."
That statement is fierce. It's fierce because it's not just a CEO's bold claim; it's a public company's promise to the capital markets: every card purchased with capex is generating revenue.
But the cost must also be stated.
The prerequisite for Alibaba's ability to "stock up" like this is that the models just need to be good enough—not necessarily the best globally, just capable of handling business and monetization. Qwen3.7-Max closely follows the capability line of DeepSeek and Kimi but hasn't created a generational gap. In terms of academic influence in international AI foundational model original research, Alibaba is relatively low-key—Qwen's open-source version has high download counts on HuggingFace, but in terms of paper weight on "where the next-generation architecture should go," Alibaba contributes far less than ByteDance. If one day the generational gap in foundational models is widened 5x by ByteDance, OpenAI, or Anthropic, all the AI loaded into today's checkout counters will become outdated hardware needing upgrade and replacement.
Alibaba's bet is: within 5 years, foundational model capabilities won't widen to a 5x generational gap.
ByteDance Locked AI Inside the Seed Department
ByteDance takes another posture.
There are two parallel lines within the Seed department. One is Zhu Wenjia, responsible for model applications—products like Doubao, Jiemeng, and Kouzi fall under him. The other is Wu Yonghui, responsible for AI foundational research exploration—the AGI roadmap belongs to him. When they first shared the stage at a company-wide meeting, the goal set was just one sentence: "The Seed department's most important goal is to explore the upper limit of intelligence."
They also projected an even rarer stance: "Considering promoting open source."
Among domestic tech giants, the word "open source" is usually only spoken repeatedly in the Linux era. ByteDance daring to mention open source in the AGI era means it no longer expects to charge for foundational models—it wants to turn foundational models into the global technological foundation itself.
The external evidence is Seedance 2.0.
Released on February 10th, 2026, this video generation model 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 multimodal capabilities—processing text, images, audio, and video inputs uniformly, generating 60-second movie-quality multi-shot videos with native audio, with 2K video generation speed 30% faster than peers. Feng Ji's public comment wasn't a PR piece; he posted it on his own Weibo—a judgment from a game creator after using it.
The internal evidence is even harder.
The Top Seed talent program, launched in May 2024, targets fresh PhD graduates; expanded in July of the same year to research interns among current PhD students. A daily salary of 2000 RMB to attract genius youths, openly competing with DeepSeek in Silicon Valley and Tsinghua campuses. Industry rumors say a former ByteDance Seed intern mentioned at a dinner that on his first day, his KPI wasn't about DAU or revenue, but "to rank in the top three on a certain international benchmark by year-end"—this kind of KPI was something he'd never seen in other companies he'd been at.
And then there's the 8-month paper.
The Doubao large model team spent 8 months on a systematic experiment titled "How Far Are Video Generation Models From World Models?". The conclusion was humble: "Video generation models can memorize training cases but cannot yet truly understand physical laws." This paper carries no commercial conversion, purely academically answering a question that might take 5-7 years to materialize.
How much does it cost ByteDance to do these things every year?
In December 2025, the Financial Times reported ByteDance's 2026 capex plan at 160 billion; on May 9th, the South China Morning Post reported 200+ billion (+25%); on May 27th, Bloomberg reported a maximum of 470 billion ($70 billion USD). Three upward revisions within 5 months, each a major jump, the latest number being 2.8x that of 2024. According to Bloomberg, funding comes from ByteDance's estimated 2025 profit of about $50 billion USD—ByteDance internally has reservations about the accuracy of that figure—meaning after spending this year's profit, it would need to borrow another $20 billion USD. Under ideal conditions, it could reach $100 billion USD (approx. 6,781 billion RMB).
Does ByteDance have enough money to burn?
Yes. It can withstand it because it's not publicly listed and doesn't have to justify ROI to the capital markets every quarter.
But does it have enough patience?
Not necessarily. Doubao just started testing paid features; that Titanium Media headline was "ByteDance Puts the Brakes on Doubao's Free Model." Doubao's DAU surpassed 100 million, becoming the product in ByteDance's history to reach that milestone with the least promotion spend, meaning logically it should have the least pressure to monetize—yet internally they've started considering inserting ads. This shows that even for a non-listed company, after burning for three years, the balance sheet starts to exert pressure.
Industry rumors suggest that VCs who invested in Doubao early on later privately commented that ByteDance's biggest change in the past two years isn't that its models got stronger, but that it truly started believing "technological leadership will make money come on its own." This kind of belief wasn't in this company's past dictionary.
But belief is one thing; belief can't be eaten. Seedance 2.0 topping the charts is one thing; turning Seedance 2.0 into the next Douyin is another—the latter is something ByteDance hasn't yet proven.
"Selling Goods" and "Making Products" Are Not Philosophies; They're Origins
Writing this far requires a counter-consensus.
The mainstream narrative is: Alibaba is pragmatic short-term, ByteDance is idealistic long-term; Alibaba is shrewder, ByteDance is more visionary. There's truth to this side—Alibaba's +40% Cloud external revenue is real, and ByteDance's Seedance 2.0 topping the charts is also real. The two are indeed on different paths.
But the real issue isn't strategic philosophy.
Alibaba is a public company. Every quarter's financial reports, stock price, buybacks, dividends must pass scrutiny in front of the capital markets. Wu Yongming says "far exceeding 380 billion over the next five years" but gives no specific number—this is the manifestation of being held hostage by the stock price. It doesn't have the luxury of "burn money for 5 years first, then see the results." If any quarter its Cloud external revenue growth falls to single digits, the next day's stock price will teach it a lesson.
ByteDance is not. It can let the Seed team spend 8 months writing a paper on world models with no commercial conversion, can let capex triple in 5 months, can let Wu Yonghui publish only academic papers and not write product requirement documents—because it doesn't have to explain to anyone the reason for these people's existence.
Therefore: If Alibaba weren't listed, it would most likely also bet on foundational models. Reference its early posture of investing in DAMO Academy, T-Head, Luohan Academy—that was Alibaba before Zhang Yong, when Jack Ma truly believed "tech companies must nurture their own scientists."
If ByteDance were listed, it would most likely also have to stick close to monetization. Reference the reaction curve where every time ByteDance faces IPO rumors, Doubao suddenly ramps up commercialization—market expectations would immediately back-program strategic choices.
What truly determines the path of Chinese AI strategy isn't the CEO's vision, but whether the company is publicly listed.
This means that for the next 5 years, BAT and other publicly listed companies starting at 4 trillion RMB market cap cannot possibly "do AI" like ByteDance does; they can only "sell AI." Conversely, non-listed companies like DeepSeek and Moonshot have the luxury to "do AI."
And conversely—if one day ByteDance truly initiates an IPO, the long-term research budget of its Seed team will be the first to face pressure. This point is more worth watching than any judgment about "Doubao benchmarking GPT."
Within one week, Wu Yongming stood on stage at the Alibaba Cloud summit shouting "5 years far exceeding 380 billion," while Zhu Wenjia and Wu Yonghui stood at the Seed all-hands meeting shouting "explore the upper limit of intelligence." One is explaining the checkout counter to shareholders of a 4 trillion RMB market cap company, the other is explaining the laboratory to their own engineers backed by non-public equity.
Alibaba loaded AI, piece by piece, into the storefront; ByteDance wrote AI, line by line, into papers.
When, one day in 2027, ByteDance truly files its S-1, we will see—for how many pages of the prospectus can the words "training" hold out.
This article is from the WeChat public account "AI Sings the Opposite Tune," author: Joshua
