After disbanding AI Lab and spending 84.6 billion dollars to buy cards in half a year, Tencent is going against the current
Author: Beating
Original title: Tencent still has dreams
On August 12, 2026, Tencent released its second quarter earnings report.
Capital expenditure for a single quarter, $52.78 billion. Three months ago, that number was 31.9 billion. Moving forward a year, the total for the whole year would be less than 79.2 billion.
This company has always been known for spending money with restraint. The speed at which it bought cards was once slow enough to make the market wonder if it actually wanted AI at the table. Now, it has brought the speed of spending money to this level within a year.
At the earnings conference on the same day, Ma Huateng said that Tencent is “building a brand new, AI-enabled Tencent.” The hybrid was renamed HY, and Hy4 will be released soon. The last time this company described itself as “brand new” was in the era when WeChat was born.
Tencent still has dreams. Its dream is not just AI; it needs to relearn to be an unstable company.
In 2018, Pan Ran said in “Tencent Has No Dreams” that Tencent is a company like water. Water is good for all things, and there is no dispute; wherever there is a channel, it flows. Water has no personality, so water doesn't have dreams.
It is natural for water to flow to a low place; backflow is for those who have reflux. In 2026, the 28-year-old company did something against nature. It admits that the article eight years ago was right,Admit that you can't live like water anymore.
水
On May 5 at 9 p.m., “Tencent Has No Dreams” was published. 13,000 words.
At 2 o'clock at night, Liu Chiping and Tencent PR director Zhang Jun responded in the circle of friends. Liu Chiping said that Tencent is a larger organization and ecosystem than the outside world can imagine. “It's too narrow to reduce Tencent to the gains and losses of a product, a kind of strategic deployment, and one person's will.”
At 2:19, a screenshot suspected of Ma Huateng's response began circulating in the circle of friends.
At 2:39 the real Ma Huateng spoke up, saying “It's nice to have criticism” to a friend who cares about him.
Afterwards, he said:
“From writing the first line of code, my dream was how to make the best product, not how much money to make.”
During the day, the national media quoted almost the full screenshot of Ma Huateng's response. Even Zhang Yiming spoke for Tencent in his circle of friends, saying this was a “Don Quixote imagination.” Tencent is not only powerful, but it is also constantly evolving in every dimension.
Zhang Jun was on the long-haul flight that day. After landing, he said, “We certainly weren't as bad as the outside world thought, but the criticism also made us realize that we weren't as good as we thought.”
Of course, there were a few different voices about that article at the time. Hong Bo said that many of the questions mentioned in the article are real questions, but is there only one correct answer for such a large company? “Perhaps the author thinks Zhang Yiming is the only correct answer. He is a bit superstitious about Zhang Yiming.”
That article also recorded an earlier story. At the beginning of 2011, just after the 3Q war ended, Tencent held a general meeting to discuss what Tencent's ability to open up is. Ma Huateng asked the 16 executives who attended each to write down what they thought Tencent's core competencies were on paper, and came up with a total of 21 answers. Finally, decide on two.
Capital, flow.
The term capital was advocated by Liu Chiping. Opening up means releasing traffic and turning it into an investment. Traffic is open, capital is open, “I don't do it myself anymore.”
These two terms have governed Tencent for ten years. The entrance to WeChat traffic and the exit of investment traffic is in the middle is a steady stream of cash generated by games and advertisements. JD's e-commerce portal entered the WeChat Jiugong grid. Sogou picked up the search, and Meituan took over the local life. Traffic is exchanged for shares, and shares are exchanged for allies. In ten years, Tencent's market capitalization has increased tenfold, surpassing Facebook's. When that article was published, it still looked invincible.
If you look back and reread it eight years later, you'll find that the article predicted almost every time Tencent fell since then.
Ten years later, on December 23, 2021, Tencent distributed 14.7% of JD shares to its shareholders, with a market value of about HK$100 billion. In January 2022, Sea holdings were reduced and $3.2 billion was cashed out. In November 2022, 9.6% of Meituan was split, or approximately HK$159.4 billion.
The capital, which was designated as a “core competency” back then, was personally destroyed by Tencent. The water has flowed back and forth for the first time in decades.
The first AI dream
There is a paragraph in that article that not many people paid attention back then. It's written in Tencent's AI. The Go program “Amazing Art” created by AI Lab successively lost to two amateur games. One is the personal hobby of Headline's vice president, and the other is an amateur work by several engineers on the WeChat translation team.
Few people realise that Tencent's first AI dream began two years before that article was published.
In the spring of 2016, AlphaGo defeated Lee Se-seok for the first time in the world to take artificial intelligence seriously. The Chinese internet company seemed to have heard a firing gun. This year, Yao Xing, vice president of Tencent, is preparing to build an AI lab. In October of the following year, the Alidamo Academy was established and announced a three-year investment of 100 billion dollars.
On March 23, 2017, Tencent announced the appointment of Zhang Tong as AI Lab Director. Zhang Tong is one of the top Chinese scholars in the field of machine learning. Previously, he was the vice president of Baidu Research Institute. When he joined, he said, “Chinese talents from around the world have strong technical advantages in the field of AI. This is an opportunity for China to develop AI.”
At the time, the AI Lab had more than 50 scientists, and more than 90% had AI-related doctorates and overseas study backgrounds. Tencent didn't think it was enough. In May 2017, it appointed Yu Dong, the chief researcher at Microsoft Research, as the deputy director of the AI Lab to oversee the newly established Seattle lab. The reason it was set up in Seattle was because many Microsoft people didn't want to leave there, so Tencent built a lab next to Microsoft.
At the beginning of the first dream, there was plenty of momentum. In March 2017, Zhuyi won the UEC Cup in Japan, and reached the top in the Dragon Star game at the end of the year. Go is the business card of those two years of AI. After AlphaGo, everyone's AI will have to play chess, and that was also Tencent's AI moment at its best.
Then the dream started to get discouraged.
The problem was Tencent's own approach. On September 30, 2018, Tencent launched the 930 revolution and established CSIG. On the third day after the change, Caixin published a report. The title directly solved the problem:
“The three AI teams are still divided, how can the 2B business horse racing mechanism be broken?”
AI Lab, YouTuo, and WeChat AI each have their own three lines; data is not shared, and talent is not shared.
On January 3, 2019, Zhang Tong no longer held a management position in the company and returned to academia. He later went to the joint laboratory of the Hong Kong University of Science and Technology and Innovation Factory. More than 50 scientists were scattered, and Zhang Zhengyou took over as director.
In 2021, Yao Xing also left. He didn't go to any major factory and founded his own company Yuanxiang. The direction is “Quanzhen Internet.” Tencent and Gao Yu sit on the Angel Round's list of investors.
In December 2025, Yu Dong, the former head of the Seattle Lab, left AI Lab and joined US Capital One. Three months later, on March 21, 2026, Tencent officially abolished the AI Lab, and some personnel were merged into the mixed yuan big model team.
It's been nine years since Zhang Tong took office until the lab was abolished. The complete life cycle of a dream.
Tencent's first AI dream wasn't without beginning, nor was it that it wasn't taken seriously. It dug up the most expensive people in the field and built the lab next to Microsoft. It died on a simpler problem. The company didn't know how to set up a dream that couldn't be realized right away.
Research is not implemented, there is no blood supply in the business, the three lines are racing against each other, and no one has lost, so no one has won.
拆
On September 30, 2018, five months after “Tencent Has No Dreams” was published, Tencent launched the 930 revolution. MIG was abolished, PCG and CSIG were newly established, and six major business groups were formed. Ma Huateng wrote in a letter to all the staff:
“Take root in the consumer Internet and embrace the industrial Internet.”
No one can prove that article drove this change, but the important thing is resonance. Tencent itself has begun to acknowledge the problems mentioned in the article. This is Tencent's biggest change in six years. Ma Huateng said this incident made him “walk on thin ice every day.”
After the change, Tencent's newly established CSIG was headed by Tang Daosheng. He announced that CSIG will no longer use the horse racing mechanism.
Horse racing is Tencent's heirloom; WeChat is the product of horse racing. In 2010, several teams were working on mobile instant messaging at the same time, and Bruce Zhang's Guangzhou team won. This mechanism produced Tencent's last great product.
On November 11, 2019, Tencent's 21st anniversary, Ma Huateng sent a letter from all employees announcing the new mission vision:
“User-oriented, technology for good.”
Back then, Tencent's old vision was “the most respected internet company”. It was very big and very vague, and no one explained how to achieve it. After a year and a half, Tencent replaced it.
There's also Tencent Cloud. The industrial internet, which had high hopes at 930, was not going well. In 2021, Tencent Cloud stopped pursuing scale and switched to reducing losses. In 2023, Tang Daosheng made it clear that he wanted to switch from an integrator to an integrator and reduce overall integration projects that are not profitable. The gross margin for fintech and corporate services slowly climbed to 51% in 2025.
The cost is that the ambition of this transformation has shrunk from “reinventing Tencent” to “making profits right.”
In addition to reducing losses, these five years have also borne fruit. Tang Daosheng decided on a term for the industrial internet called C2B. The industrial Internet is not only ToB and ToG, but also ToC; at the end of the day, the B-side and G-end are still the C-side. A batch of products such as Tencent Conference and Enterprise WeChat came along this path. They didn't create another Tencent, but they taught Tencent how to make software on the cloud for the first time, and learned that they no longer use data directly in exchange for money.
On the 930th anniversary, Yang Guoan said that Tencent began to get rid of its dependence on the “rich second generation” style of eating with social networking and games. This statement is actually only half right. The path is indeed changing, but the dependency is still there. Many years later, Tang Daosheng will say that big models, like clouds, are difficult to be monopolized. The motivation behind this statement is that it was reduced by reducing losses in those five years.
In May 2023, Liu Chiping retired as an executive director. He is still the president and still manages the Investment Committee. But the symbolism is clear. Those who claim that “capital is a core competency” have been removed from the board of directors list. Some media calculated it for him. The total salary for 16 years was over 2.3 billion yuan.
In five years from 930 to 2023, almost everything Tencent did was “demolish.” Break up the organization, break up the investment, break up the vision, and dismantle the horse race. It knew it was sick, so it took off its old clothes one by one. But the hardest thing for a company is not to change direction; it is that after admitting that it has changed direction, it still doesn't know the answer. This second transformation of the industrial Internet ended with a loss reduction.
The water changed to a river, or water.
潮
Ma Huateng was asked about AI at a shareholders' meeting in Hong Kong, China on May 19, 2023. He said:
“AI is an opportunity the internet hasn't seen in 100 years.”
I added that Tencent will not take out semi-finished products in order to boost stock prices.
There is both sobriety and delay in this sentence. Being sober is that he can see the weight of this revolution. The delay is that at the time of saying this, Baidu's Wenxin 1 statement had already been posted for 2 months, Ali's Tongyi Qianwen was published for 1 month, and Tencent's model would have to wait another 4 months. On September 7, 2023, the mixed yuan big model was officially released, focusing on “solving the nonsense.”
Safety and restraint, Tencent's consistent attitude.
After almost a year of restraint, Tencent opened the Hunyuan-Large in November 2024. 389 billion parameters, the world's largest open source MoE model at the time. A company that has always had a cold attitude towards open source opened the door to models for the first time.
On May 30, 2024, Tencent Yuanbao went live. The company said to the outside world: “Big models are definitely ahead of the curve.”
After a year, it overthrew itself.
On February 13, 2025, Yuanbao will access the DeepSeek-R1 full-blooded version, free of charge. Three days later, WeChat Search Grayscale launched AI search, and also connected to DeepSeek. Tencent specifically stated that this feature will not read the user's chat history and circle of friends. On March 4, Yuanbao topped the Apple App Store free download list in China.
This is the last time that the old Traffic spell has appeared. Connect to your opponent's model, inject your own traffic, and reach the top in a month.
Then, Tencent wanted to show up again.
At the staff meeting on January 26, 2026, Ma Huateng announced a game method. The “ingot party” distributed 1 billion red envelopes during the Spring Festival. He said he hoped to recreate the moment of WeChat Red Envelopes 11 years ago.
On New Year's Eve 2014, WeChat red envelopes were launched. Eight million users and 40 million red envelopes, Ma Yun called it a “sneak attack on Pearl Harbor.” A year later, on New Year's Eve, WeChat and CCTV collaborated on the Spring Festival Gala, and people all over the country shook their phones on the TV. Red envelopes were sent and received 1.01 billion times overnight. The mobile payment landscape was rewritten between these two New Year's Eves.
It was Tencent's last surprise attack with products and traffic. Eleven years later, it hasn't had a night like that again.
During the 2026 Spring Festival, Tencent spent 1 billion dollars to buy back this night.
It hasn't been bought back. It's only been a few days since the red envelope campaign went live, so WeChat took the lead. Around February 4, WeChat restricted the opening of links to grab red envelopes for yuan on the grounds of “inducing users to share links frequently” and “causing harassment to users.”
Our products were blocked from the outside by our own access control.
Yuanbao responded that the sharing mechanism is being adjusted urgently. The WeChat PR director gave four words:
“Treat them all the same.”
To be sure, WeChat isn't completely cynical. Invitations from the Ingot Party were allowed to spread through WeChat friends, group chats, and friend circles. This is Tencent's core relationship chain entry, and WeChat has opened up a slit. But in the end, the ingot couldn't get through that gap.
At its peak, Yuanbao had more than 50 million daily active users and 114 million monthly active users. Then the tide recedes. YuanBao fell out of the top ten free listings on the Apple Store. QuestMobile's later data showed that as soon as the red envelope campaign ended, Yuanbao's daily active users fell back to the level before the Spring Festival.
By March 2026, Yuanbao's independent app had 57.35 million monthly active users, and there were only three domestic AI apps with monthly survival of over 100 million: Doubao, QianQ, and DeepSeek.
Three months later, at the shareholders' meeting on May 13, 2026, Ma Huateng said something out of his heart:
“We thought we were on a boat a year ago, but then we discovered that the boat was leaking. I've started changing boats again, and now I feel like I'm standing up. I still can't sit down. I still want the boat to be a little faster.”
In 2014, WeChat Red Envelopes changed the payment landscape with a New Year's Eve. In 2026, Tencent used 1 billion dollars to buy back the same lesson. Traffic isn't users; it's water or water; it can take anything to the top of the list, but it doesn't keep anything.
In June 2026, Tencent paid 10 billion dollars for the first round of DeepSeek. Some people interpreted it as Tencent admitting they couldn't do it, so they paid to buy it.
A 27-year-old chief scientist can't change the organizational inertia of a 100,000-person company. The lessons learned from AI Lab's nine years of burning are there. It's not that there aren't any talents; it's that this company used to digest talent and dreams in the way it was best.
reflux
2026, backflow.
Let's look at the money first. In 2024, Tencent's capital expenditure was 76.7 billion yuan, an increase of 221% over the previous year. In 2025, 79.2 billion. This figure only rose 3% in the year when the AI arms race was at its peak. In the same period, Byte's capital expenditure was estimated to be over 160 billion dollars, and Ali announced a three-year investment of 380 billion dollars.
Tencent made a repurchase without the money to buy the card.
In 2025, it repurchased approximately HK$80 billion worth of shares, plus dividends, and returned more than HK$240 billion to shareholders in two years. During the annual report conference call on March 18, Liu Chiping explained that due to GPU supply restrictions in 2025, it is impossible to buy a card. Actually, it's not that I can't buy it; I don't want to buy it at a high premium. On the day after the earnings report, Tencent's stock price fell by more than 6%.
Then 2026 came. In the first quarter, capital expenditure was $31.9 billion. In the second quarter, $52.78 billion. The total for the first half of the year was 84.68 billion, more than the full year of 2024. During the conference call, Liu Chiping also hinted that in the future, repurchases may be reduced and money invested in AI. That abacus, which calculated a 10-year ROI, was put down in 2026.
Let's look at the organization again. In December 2025, Tencent upgraded its large model R&D architecture and established a new AI Infra Department, AI Data Department, and Data Computing Platform Division. In March 2026, AI Lab was decommissioned. The carrier of the first dream was dismantled, and people and assets merged into the mix. A company personally demolished a research institution it had built for nine years. The funeral and handover ceremony were held on the same day.
Then there are people. In December 2025, Tencent announced that Yao Shunyu, 27, will be the chief AI scientist in the “CEO/President's Office” and will report to President Liu Chiping.
Yao Shunyu, born in 1998, is a native of Hefei, Anhui. The high school is in Hefei No. 1 Middle School, third in science in the province in the 2014 college entrance examination, and was admitted to the Yao Class of Tsinghua University. In Tsinghua, he organized a rap club. The high school class teacher remembers that this third-year science kid in the province likes to sing and dance rap basketball.
He studied for his PhD at Princeton University, created the ReACT framework and thought tree during his blog, and led Swe-bench. Today, half of the world's smart body research stands on the scaffolding he set up. In August 2024, he joined OpenAI to participate in Operator and Deep Research. It left in September 2025. Three months later, he became Tencent's chief AI scientist at age 27.
The media asked him why he chose Tencent. His answer was:
“The most important thing is that everyone is very honest.”
On January 10, 2026, he made his first public statement after joining the company, saying “Claude Code is reshaping the computer industry.” On June 5, he had a conversation with Tang Daosheng and was asked if Tencent AI was slow. He said:
“AI is a long-term game, and the second half is just beginning.”
Tang Daosheng went on to say, “This is a long-distance run. This is a marathon. Tencent still has a very rich scene.”
In Tencent's AI map, there are two armies standing.
One team at TEG, mixed. After Yao Shunyu took over, it took over the scattered model development, data, and training infrastructure one by one into its own hands. In the abolished AI Lab, some researchers were merged into the Big Language Model Department, and some were transferred to the Industry-University-Research Cooperation Center.
The other one is on WXG, WeChat. Its name is WelM. In 2022, WeChat released this Chinese pre-trained model, with 10 billion parameters, surpassing the model of the same scale in 18 Chinese tasks. In August 2025, WeChat also unveiled WeChat-yatt, a large model reinforcement learning training framework. By 2026, the new generation of WeLM will move to a sparse MoE architecture, with 80 billion parameters in the basic version, which only activates 3 billion inference each time; 130 billion with 4.9 billion activations in the extended version. On July 9, the WeChat AI team published the Hidden Decoding Study and wrote two new models into the paper.
People are also moving on both sides. At the end of 2025, Xu Can, a senior mixed-yuan researcher, was transferred to WeLM. Xu Can, the creator of WizardLM, proposed Evol-Instruct at Microsoft in 2023 to let big models solve their own problems. After he left, the mixed training team split into several groups to explore and move forward.
Two armies, two logics. Mixture has concentrated computing power, data, and model resources in its hands. WeLM is in the hands of WeChat's native portal, applet service, and real feedback from over a billion users. They collaborate and compete. Tencent's AI portal dispute, the first one took place in my own home.
In 2019, Xiaolong Zhang said, “Every day, 100 million people teach me how to make products.” On January 9, 2020, Bruce Zhang did not attend the WeChat Open Course. He recorded a 13-minute video talking about seven thoughts on the Internet of Information, admitting that the public platform made two mistakes, and announced that he wanted to write short content.
Ten days later, the video account was born. This is the only time that WeChat has fought back head-on in Douyin's most ferocious years. That battle later proved one thing: Tencent's product capabilities have not died out; it has always lived in the WeChat building.
In 2021, he summed up the decade of WeChat in two words: connected and simple.
He said he hoped WeChat would always be like a small but beautiful product, with its own soul, its own aesthetics, its own ideas, and its own ideas. Back then, Pan called him an “idiot among smart people.”
During the 2026 Spring Festival, it was this idiot's product hygiene addiction that blocked Yuanbao's 1 billion red envelopes out of WeChat. The part of Tencent that is most restrained is instead the most like a product company. In 2026, WeChat was taken to the table. Search One Search connects to the AI, WeChat Agent starts testing, and the AI Assistant has a “small” gray scale. Some people say that Bruce Zhang is betting on everything this time.
Big models have also entered the WeChat community for a long time. The advertising recall process for video accounts and friend circles has deployed a system called LEADRE. The bottom layer uses a mixed model with 1 billion parameters.
“Xiaowei” will start a small range of grayscale in June 2026. Users can directly call it to operate WeChat's native functions and call applets to complete tasks. Internal testers said that Xiaomiao's main model is WelM, and some answers will call DeepSeek.
In June, the WeChat Open Platform did something more thorough. The applet can be connected to the WeChat AI ecosystem. In automatic mode, the platform reads the applet source code, analyzes the page structure, and allows WeChat AI to directly operate; in development mode, developers encapsulate business capabilities into skills, which can be used by WeChat AI after approval. WeChat wants AI to grow into its own hands.
At the end, it was that $10 billion. In June 2026, DeepSeek completed the first round of financing, totaling over $50 billion. Liang Wenfeng personally spent about 20 billion dollars, during the Ningde era, about 5 billion dollars, and Tencent about 10 billion dollars. It is said that the investors were personally selected by Liang Wenfeng.
Investing is Tencent's most popular position. But this investment can also be read as a new gesture. Eight years ago, Tencent invested to “contain Ali”; it was a defense. Today, it invests in a company that it can't catch up with for a while, and has opened its most expensive traffic channel, WeChat, to its model. The kind of company that it used to invest back then to “take over” is now a teacher that it recognizes with its investment.
It is natural for water to flow to a low place; backflow is for those who have reflux.
Tencent in 2026 continued to make water, traffic, investment, and ecology without losing it.On the other hand, they began to reverse the current trend, and before they saw a return, they threw real money into it and bet their most core assets on uncertainty. Its dream came back.
Epilogue
At the end of “Tencent Has No Dreams”, it says:
“I hope Tencent doesn't have to wait four and a half years before reviewing and reflecting.”
Tencent waited more than four and a half years; it took eight years.
In these eight years, it distributed “capital” to shareholders, opened up “traffic” to rivals, and brought the horse racing system into history. The old vision was replaced with a new one, and the AI lab I built myself was torn down. Finally, a 27-year-old took the position of chief scientist. It has fallen over everything that was predicted, and it has fallen past everything. The decline of ingots, loss reduction in clouds, and the funeral of the first AI dream. After falling, it was still moving in that direction.
“The Great Gatsby” was written in 1925. Gatsby is a person who bets his whole life on the “past.” He bought a mansion on the other side of Long Island Bay and hosted banquets at night. The lights were bright, just for a green light on the other side. Behind the lamp lives the girl he loved five years ago. On the last few pages of the novel, Fitzgerald wrote a passage for Gatsby and for anyone who wanted to rewrite time:
“As a result, we struggled to move forward, and were constantly pushed back to the past.”
Gatsby went against the trend to go back in time. He believes that as long as he paddles the boat fast enough, time will go back and his old dreams will be revived. He failed. Bullets passed through the pool and the party broke up, and the green light was always across a bay.
Tencent is also boating; it's going against the tide so it won't go back. Don't go back to that self that only has capital and flow, only certainty, and only a way of living like water.
Tencent still has dreams.
Twitter:https://twitter.com/BitpushNewsCN
Compare the TG exchange group:https://t.me/BitPushCommunity
Compare TG subscriptions:https://t.me/bitpush
