Author: Dongcha Beating
Tencent still has dreams
On August 12, 2026, Tencent released its second-quarter financial results.
Quarterly capital expenditure: RMB 52.78 billion. Three months ago, this figure was RMB 31.9 billion. A year ago, the total for the entire year was only RMB 79.2 billion.
The company has always been known for its frugality. At one point, its pace of purchasing GPUs was so slow that the market doubted whether it truly wanted to join the AI table. Now, it has ramped up its spending to this level within just one year.
On the same day’s earnings call, Ma Huateng said Tencent is “building a brand-new, AI-empowered Tencent.” Hunyuan is now called HY, and Hy4 is set to be released soon. The last time this company described itself as “brand-new” was during the era of WeChat’s creation.
Tencent still has dreams. Its dream is not just about AI—it wants to relearn how to be an uneasy company.
In 2018, Pan Luan said in "Tencent Has No Dreams" that Tencent is a company like water—water benefits all things without striving, flowing wherever there is a channel. Water has no personality, so water has no dreams.
Water flows downhill—that’s its nature; going against the current is the domain of those who choose to defy it. In 2026, this 28-year-old company did something contrary to nature: it admitted that the article from eight years ago was right—that living like water is no longer sustainable.
Water
At 9 PM on May 5, "Tencent Has No Dreams" was published. Thirteen thousand words.
At 2 a.m., Liu Chiping and Tencent’s PR Director Zhang Jun responded on Moments. Liu Chiping said Tencent is a larger organization and ecosystem than outsiders imagine: “Reducing Tencent to the success or failure of a single product, a single strategy, or one person’s will is too narrow.”
At 2:19 PM, a screenshot allegedly showing Ma Huateng's response began circulating in Moments.
At 2:39, the real Ma Huateng spoke up, saying to a friend who cared about him, "It's good to have criticism."
Then he said:
Since writing my first line of code, my goal has always been to create the best possible product, not to make as much money as possible.
During the day, national media nearly quoted the screenshot of the alleged Ma Huateng response in full, and even Zhang Yiming defended Tencent on his Moments, calling it a “Don Quixote-like imagination,” stating that Tencent is not only powerful but continuously evolving across all dimensions.
Zhang Jun had been on a long flight that day, and after landing, he said: "We are certainly not as bad as outsiders think, but the criticism has made us realize that we are not as good as we thought we were."
Of course, there were also differing voices regarding that article. Hong Bo said that many of the issues mentioned in the article were real, but can such a large company have only one correct answer? “Perhaps the author believes Zhang Yiming is the only correct answer. He somewhat idolizes Zhang Yiming.”
The article also recorded an earlier story. In early 2011, shortly after the 3Q war ended, Tencent held a senior management meeting to discuss what Tencent’s open capabilities were. Ma Huateng asked the 16 executives present to write down on paper what they believed to be Tencent’s core capabilities, resulting in a total of 21 responses. Ultimately, two were selected.
Capital, traffic.
The term "capital" was advocated by Liu Chiping. Openness means releasing traffic so it becomes investment. Open traffic, open capital—“I’m no longer doing it myself.”
These two words governed Tencent for a decade. WeChat controlled the entry point of traffic, while investments controlled the exit, with games and advertising continuously generating cash in between. JD.com’s e-commerce entry point entered WeChat’s nine-grid, Sogou took over search, and Meituan took over local life services. Traffic was exchanged for equity, and equity was exchanged for allies. Over the course of a decade, Tencent’s market value increased tenfold, surpassing Facebook. At the time the article was published, it still appeared invincible.
Eight years later, rereading that article reveals that it almost predicted every fall Tencent would later experience.
On December 23, 2021, ten years later, Tencent distributed 14.7% of its JD.com shares to its shareholders, with a market value of approximately HK$100 billion. In January 2022, it reduced its stake in Sea, realizing proceeds of $3.2 billion. In November 2022, it further distributed 9.6% of its Meituan shares, worth approximately HK$159.4 billion.
The capital once designated as a "core capability" was dismantled by Tencent. After flowing for decades, it began to flow backward for the first time.
The First AI Dream
There was a passage in that article that received little attention at the time. It mentioned Tencent’s AI, specifically the AI Lab’s Go program “Jueyi,” which lost to two amateur projects—one a personal hobby of a Toutiao vice president, and the other the side project of several engineers from the WeChat translation team.
Few realize that Tencent's first AI dream began two years before that article was published.
In the spring of 2016, AlphaGo defeated Lee Sedol, and the world took artificial intelligence seriously for the first time. Chinese internet companies felt as if they had heard the starting gun. That year, Yao Xing, vice president of Tencent, established the AI Lab. The following October, Alibaba's Damo Academy was founded, announcing an investment of 100 billion yuan over three years.
On March 23, 2017, Tencent announced the appointment of Zhang Tong as Director of the AI Lab. Zhang Tong is one of the foremost Chinese scholars in the field of machine learning and previously served as Deputy Director of Baidu Research. Upon joining, he said: "Chinese diaspora talent around the world holds strong technical advantages in AI, which presents an opportunity for China's AI development."
At the time, the AI Lab had more than 50 scientists, over 90% of whom held PhDs in AI and had studied abroad. Tencent still felt this wasn’t enough, so in May 2017, it appointed Yu Dong, a principal researcher at Microsoft Research, as deputy director of the AI Lab, overseeing the newly established Seattle lab. The lab was located in Seattle because many Microsoft employees were unwilling to leave the area, so Tencent built its lab right next to Microsoft.
The first dream began with great fanfare. In March 2017, Yixin won the UEC Cup in Japan, and by year’s end, it also claimed victory in the Dragon Star Tournament. Go was the flagship of AI during those two years. After AlphaGo, every AI team had to be able to play Go—it was the pinnacle of Tencent AI’s glory.
Then, the dream began to lose steam.
The issue lies in Tencent’s own approach. On September 30, 2018, Tencent launched the 930 Reform and established CSIG. Three days after the reform, Caixin published a report whose title directly pointed out the problem:
Three AI teams still operate in silos—how to break the horse-racing mechanism in B2B business?
AI Lab, YouTu, and WeChat AI operate independently, with no data sharing or talent exchange among them.
On January 3, 2019, Zhang Tong stepped down from his management position and returned to academia. He later joined the joint laboratory between HKUST and Innovation Works. More than fifty scientists left, and Zhang Zhengyou succeeded as director.
In 2021, Yao Xing also left. He didn’t join any major tech company but founded his own company, Yuanxi, focused on “metaverse internet.” The angel round investors included Tencent and Hillhouse.
In December 2025, Yu Dong, then head of the Seattle lab, left the AI Lab to join Capital One in the United States. Three months later, on March 21, 2026, Tencent officially disbanded the AI Lab, with some personnel integrated into the Hunyuan large model team.
From Zhang Tong’s appointment to the lab’s dissolution, nine years. The complete lifecycle of a dream.
Tencent's first AI dream didn't fail to begin, nor was it neglected. It hired the most expensive talent in the field and built its lab right next to Microsoft. But it died over a simpler problem: the company didn't know how to accommodate a dream that couldn't be monetized immediately.
Research doesn't translate into execution, business doesn't provide support, and the three teams are racing against each other—no one loses, so no one wins.
Withdraw
On September 30, 2018, five months after the publication of "Tencent Has No Dreams," Tencent launched the 930 Reform. MIG was dissolved, and PCG and CSIG were newly established, forming six business groups. In his letter to all employees, Ma Huateng wrote:
Rooted in consumer internet, embracing industrial internet.
No one can prove that the article drove this transformation, but what matters is the resonance. The issues raised in the article are now being acknowledged by Tencent itself. This is Tencent’s biggest transformation in six years, and Ma Huateng said it makes him “feel like walking on thin ice every day.”
After the restructuring, Tencent's newly established CSIG, led by Tang DaoSheng, announced that CSIG will no longer use the horse-racing mechanism.
Racing is Tencent's family heirloom, and WeChat is a product of that racing. In 2010, several teams were developing mobile instant messaging simultaneously, and Zhang Xiaolong’s Guangzhou team won. This mechanism produced Tencent’s last great product.
On November 11, 2019, during Tencent’s 21st anniversary, Ma Huateng sent a company-wide letter announcing the new mission and vision:
User-focused, technology for good.
In the past, Tencent’s old vision was “the most respected internet company,” which was very broad and abstract, with no one clearly explaining how to achieve it. A year and a half later, Tencent changed it.
There's also Tencent Cloud. The industrial internet, which was highly anticipated at 930, has faced challenges. In 2021, Tencent Cloud shifted away from pursuing scale and focused on reducing losses. In 2023, Tang Daosheng clearly stated the goal of transitioning from an integrator to a被集成商, scaling back unprofitable total integration projects. The gross margin for fintech and enterprise services gradually rose to 51% by 2025.
The cost is that the ambition of this transformation has been scaled back from "recreating Tencent" to "achieving profitability."
Beyond reducing losses, these five years have also borne fruit. Tang Daosheng defined industrial internet as C2B. Industrial internet is not only ToB and ToG—it is also ToC; ultimately, B-side and G-side are both C-side. Products like Tencent Meeting and Enterprise WeChat emerged from this path. They didn’t recreate Tencent, but they enabled Tencent to learn for the first time how to build software on the cloud, and how to stop directly exchanging traffic for money.
At the 930 anniversary, Yang Anguo said Tencent began to move away from its "trust fund kid" dependency on social and gaming revenue. But this statement is only half correct—the path has indeed changed, yet the dependency remains. Years later, Tang Daosheng would say that large models, like cloud computing, are hard to monopolize. That confidence stems from the five years of reduced losses.
In May 2023, Liu Chiping stepped down as an executive director. He remains president and continues to oversee the investment committee, but the symbolic meaning is clear: those who championed "capital as a core competency" have departed from the board. Media outlets calculated that his total compensation over 16 years exceeded 2.3 billion yuan.
From 930 to 2023, over five years, Tencent did almost nothing but "dismantle"—dismantling its organization, its investments, its vision, its horse racing ventures. It knew it was sick, so it slowly took off its old clothes, one by one. But the hardest thing for a company is not changing direction—it’s admitting that even after changing direction, it still doesn’t have the answers. This second transformation into industrial internet ended with reduced losses.
Water is still water, even if it flows through a different riverbed.
Tide
On May 19, 2023, Ma Huateng was asked about AI at a shareholder meeting in Hong Kong, China, and he said:
AI is a once-in-a-century opportunity for the internet.
They added that Tencent won't release unfinished products just to boost its stock price.
This statement contains both clarity and delay. The clarity lies in his recognition of the weight of this revolution. The delay lies in the fact that when these words were spoken, Baidu’s Wenxin Yiyan had already been released for two months, Alibaba’s Tongyi Qianwen for one month, and Tencent’s model was still four months away. On September 7, 2023, the Hunyuan large model was officially launched, touting its ability to “solve incoherent output.”
Security and restraint, consistent with Tencent's approach.
After holding back for nearly a year, Tencent open-sourced Hunyuan-Large in November 2024—a 389-billion-parameter model, the largest open-source MoE model in the world at the time. A company long indifferent to open source, it opened its doors to models for the first time.
On May 30, 2024, Tencent Yuanbao was launched. The company publicly stated: "Large models do not compete for short-term advantage."
A year later, it overthrew itself.
On February 13, 2025, Yuanbao integrated the full-power version of DeepSeek-R1 for free. Three days later, WeChat’s “Search” feature launched AI search in beta, also integrating DeepSeek. Tencent specifically stated that this feature does not access users’ chat records or Moments. On March 4, Yuanbao reached the top of the Free Download chart on the Apple App Store in China.
This is the last time the old traffic trick will work. Integrate the opponent’s model, pump in your own traffic, and rise to the top in one month.
Then, Tencent wanted to perform a miracle once again.
At the employee conference on January 26, 2026, Ma Huateng announced a new initiative: "Yuanbao Pay," distributing 1 billion red packets during the Spring Festival. He said he hoped to recreate the moment of WeChat red packets eleven years ago.
On New Year's Eve in 2014, WeChat Red Packets launched. With 8 million users and 40 million red packets, Jack Ma called it a "surprise attack on Pearl Harbor." A year later on New Year's Eve, WeChat partnered with CCTV's Spring Festival Gala, as people across the country shook their phones in front of their TVs. In a single night, red packets were sent and received 1.01 billion times. The landscape of mobile payments was rewritten between those two New Year's Eves.
That was the last time Tencent achieved a surprise victory through products and traffic. In the eleven years since, it has never had another night like that.
In the 2026 Spring Festival, Tencent wanted to buy back this night with one billion yuan.
It wasn't bought back. Shortly after the red packet campaign launched, WeChat took action. Around February 4, WeChat restricted access to the Yuanbao red packet links, citing "inducing users to frequently share links" and "harassing users."
Our own product is being blocked outside by our own access control.
Yuanbao responded that the sharing mechanism is being urgently adjusted. WeChat's PR director provided four words:
Treated equally.
To be fair, WeChat was not entirely closed off. Yuanbao’s invitations were once allowed to spread through WeChat friends, group chats, and Moments. This was Tencent’s most core entry point for social connections, and WeChat had cracked open a small door. But in the end, Yuanbao couldn’t squeeze through that crack.
At its peak, Yuanbao had over 50 million daily active users and 114 million monthly active users. Then, the tide receded. Yuanbao fell out of the top ten free apps on the Apple App Store. Subsequent data from QuestMobile showed that after the red packet campaign ended, Yuanbao’s daily active users dropped back to pre-Chinese New Year levels.
By March 2026, the monthly active users of Yuanbao's standalone app reached 57.35 million, and only three AI apps in China had over 100 million monthly active users: Doubao, Qwen, and DeepSeek.
Three months later, at the shareholders' meeting on May 13, 2026, Ma Huateng said something heartfelt:
A year ago, we thought we had boarded the ship, but later we found out it was leaking. We started looking for another ship, and now we feel like we’ve stepped onto it, but we still can’t sit down comfortably—we still hope the ship can move faster.
In 2014, WeChat red packets transformed the payment landscape in a single Lunar New Year's Eve. In 2026, Tencent spent one billion yuan to relearn the same lesson. Traffic is not users; water is still water—it can wash anything to the top, but it can’t hold onto anything.
In June 2026, Tencent invested 10 billion yuan in DeepSeek’s seed round. Some interpreted this as Tencent admitting it couldn’t build it itself, so it chose to pay for it instead.
A 27-year-old chief scientist cannot change the organizational inertia of a company with 100,000 employees. The lessons learned over nine years of AI Lab funding are right there before us. It’s not that there’s a lack of talent—it’s that this company has always used its most familiar methods to consume both talent and dreams.
Reverse flow
2026, against the current.
Look at the money first. In 2024, Tencent's capital expenditures amounted to 76.7 billion yuan, a 221% year-over-year increase. In 2025, it reached 79.2 billion yuan—an increase of just 3% during the most intense year of the AI arms race. Meanwhile, ByteDance's capital expenditures are estimated to exceed 160 billion yuan, and Alibaba has announced a three-year investment of 380 billion yuan.
Tencent used the money not spent on cards to conduct buybacks.
In 2025, it repurchased approximately HK$80 billion worth of shares, and together with dividends, returned over HK$240 billion to shareholders over two years. On the earnings call on March 18, Liu Chi-ping explained that in 2025, GPU supply constraints prevented the company from purchasing cards. In reality, it wasn’t that they couldn’t buy them—they simply didn’t want to buy at premium prices. The day after the earnings report, Tencent’s stock dropped more than 6%.
Then came 2026. In the first quarter, capital expenditures reached 31.9 billion. In the second quarter, they rose to 52.78 billion. For the first half of the year, total spending amounted to 84.68 billion, surpassing the entire year of 2024. During the earnings call, Liu Chiping also hinted that the company might reduce buybacks in the future and redirect funds toward AI. The abacus that had calculated a ten-year ROI was put aside in 2026.
Now consider the organization. In December 2025, Tencent upgraded its large model development architecture, establishing new departments: AI Infra, AI Data, and the Data Computing Platform. In March 2026, the AI Lab was dissolved. The vessel of the first dream was dismantled, with personnel and assets integrated into Hunyuan. A company dismantling its own nine-year-old research institution is like holding a funeral and a handover ceremony on the same day.
Then there are people. In December 2025, Tencent announced that 27-year-old Yao Shunyu has been appointed Chief AI Scientist in the CEO/President’s Office, reporting to President Liu Chiping.
Yao Shunyu, born in 1998, is from Hefei, Anhui. He attended Hefei No.1 High School and ranked third in the province in the science category of the 2014 college entrance exam, gaining admission to the Yao Class at Tsinghua University. At Tsinghua, he organized a rap society. His high school homeroom teacher remembers that this top-ranked science student loved singing, dancing, rap, and basketball.
He earned his Ph.D. from Princeton University, where he developed the ReAct framework and Tree of Thought, and led the SWE-bench project. Today, half of the world’s agent research stands on the scaffolding he built. In August 2024, he joined OpenAI to work on Operator and Deep Research, leaving in September 2025. Three months later, at age 27, he became Tencent’s Chief AI Scientist.
The media asked him why he chose Tencent. His answer was:
Most importantly, everyone is very honest.
On January 10, 2026, after joining, he made his first public statement, saying, "Claude Code is reshaping the computer industry." On June 5, he had a conversation with Tang Daosheng and was asked whether Tencent's AI was falling behind. He said:
AI is a long-term game; the second half has just begun.
Tang Daosheng continued: "This is a marathon, and Tencent still has very rich scenarios."
In Tencent's AI landscape, two armies stand tall.
At TEG, Hunyuan. After Yao Shunyu took over, he gradually brought together model development, data, and training infrastructure that were previously scattered across different teams. The dissolved AI Lab saw some researchers transferred to the Large Language Model team, while others moved to the Industry-Academia-Research Collaboration Center.
Another team within WXG, WeChat, developed a model called WeLM. In 2022, WeChat released this Chinese pre-trained model with 10 billion parameters, outperforming other models of similar scale on 18 Chinese tasks. In August 2025, WeChat unveiled WeChat-YATT, a large model reinforcement learning training framework. By 2026, the next-generation WeLM transitioned to a sparse MoE architecture: the base version has 80 billion parameters with only 3 billion activated per inference, while the expanded version has 130 billion parameters with 4.9 billion activated. On July 9, the WeChat AI team published research on Hidden Decoding, introducing two new models in a paper.
People are also moving between both sides. By the end of 2025, Xu Can, a senior researcher at HunYuan, transferred to WeLM. Xu Can is the creator of WizardLM and, in 2023, introduced Evol-Instruct at Microsoft, enabling large models to generate challenges for themselves. After his departure, HunYuan’s post-training team was split into several smaller groups, each forging ahead on their own path.
Two armies, two logics. Hunyuan holds centralized computing power, data, and model resources. WeLM holds WeChat’s native entry points, mini-program services, and real-time feedback from over a billion users. They collaborate, and they compete. Tencent’s battle for AI entry points begins at home.
In 2019, Zhang Xiaolong said, "Every day, a hundred million people teach me how to build products." On January 9, 2020, at the WeChat Public Lecture, Zhang Xiaolong did not attend in person. Instead, he released a 13-minute video outlining seven thoughts on information connectivity, acknowledged two mistakes with the public platform, and announced plans to focus on short-form content.
A few weeks later, Video Channel was born. This was WeChat’s only direct counterattack during TikTok’s most aggressive years. That battle later proved one thing: Tencent’s product capabilities were never dead—they’ve always lived inside WeChat’s building.
In 2021, he summarized WeChat's first decade with two words: connection and simplicity.
He said he hopes WeChat will always remain a small yet beautiful product, with its own soul, its own aesthetics, its own creativity, and its own philosophy. Pan Luan once called him "a fool among the smart ones."
In 2026, during the Spring Festival, it was this fool’s product obsession that blocked Yuanbao’s 1 billion red packets outside WeChat. The most restrained part of Tencent turned out to be the most like a product company. In 2026, WeChat was put on the table. Search One was integrated with AI, WeChat Agent began testing, and the AI assistant “Xiao Wei” entered gradual rollout. Some said this time, Zhang Xiaolong had bet everything.
Large models have long been integrated into WeChat's infrastructure. In the ad retrieval process for Moments and Video Channels, a system called LEADRE is deployed, built on top of the HunYuan model with 1 billion parameters.
Xiao Wei began a limited beta release in June 2026. Users can directly ask it to perform native WeChat functions or trigger mini-programs to complete tasks. Internal testers say Xiao Wei’s primary model is WeLM, and some responses invoke DeepSeek.
In June, WeChat Open Platform took a more thorough step: mini-programs can integrate with WeChat's AI ecosystem. In automatic mode, the platform reads the mini-program's source code, analyzes the page structure, and enables WeChat AI to operate directly; in development mode, developers package their business capabilities as skills, which, after approval, can be invoked by WeChat AI. WeChat wants AI to become its own hand.
Finally, there was the 10-billion-yuan investment. In June 2026, DeepSeek completed its first funding round, raising over 50 billion yuan. Liang Wenhong contributed approximately 20 billion yuan, CATL around 5 billion yuan, and Tencent about 10 billion yuan. It is said that the investors were personally selected by Liang Wenhong.
Investment is what Tencent does best. But this investment can also be read as a new posture. Eight years ago, Tencent invested to "contain Alibaba"—it was defensive. Today, it invests in a company it cannot yet match, and has opened its most valuable traffic channel, WeChat, to that company’s model. The very companies it once "absorbed" through investment are now its teachers, acknowledged through investment.
Water flows downhill—that’s its nature; going against the current is the concern of those who choose to do so.
In 2026, Tencent continues to flow—maintaining its traffic, investments, and ecosystem—while also swimming upstream, pouring real capital into initiatives before seeing any return, betting its most core assets on uncertainty. Its dream has returned.
Epilogue
The ending of "Tencent Has No Dreams" writes:
I hope Tencent doesn't have to wait four and a half years to reflect and review.
Tencent waited for more than four and a half years; it took eight.
Over these eight years, it distributed "capital" to shareholders, opened "traffic" to competitors, and consigned the horse-racing model to history. The old vision was replaced by a new one, and the AI Lab it built itself was dismantled. In the end, a 27-year-old young man took the seat of Chief Scientist. All the falls it was once predicted to suffer had already happened—the ebb of Yuanbao, the cloud’s reduced losses, the funeral of the first AI dream. After falling, it still kept moving in that same direction.
The Great Gatsby was written in 1925. Gatsby is a man who bets his entire life on the past. He buys a mansion across the bay in Long Island, throws lavish parties every night with lights blazing, all for the green light across the water. Behind that light lives the girl he loved five years ago. In the final pages of the novel, Fitzgerald writes these words on behalf of Gatsby—and everyone who wishes to rewrite time:
So we beat on, boats against the current, borne back ceaselessly into the past.
Gatsby fought against the current to return to the past. He believed that if he rowed hard enough, time would reverse and old dreams would reignite. He failed. The bullet pierced the pool, the party ended, and that green light remained across the bay.
Tencent is also rowing, paddling upstream to never go back—to never return to the version of itself that knew only capital and traffic, only certainty, only a life as fluid as water.
Tencent still has dreams.
Twitter: https://twitter.com/BitpushNewsCN
BitPush Telegram community: https://t.me/BitPushCommunity
BitPush TG subscription: https://t.me/bitpush