---
title: "Tencent's AI Gamble: Backend Pretends to be Poor, Frontend Goes Crazy"
type: "News"
locale: "en"
url: "https://longbridge.com/en/news/279755722.md"
description: "Betting the same way"
datetime: "2026-03-19T09:14:40.000Z"
locales:
  - [zh-CN](https://longbridge.com/zh-CN/news/279755722.md)
  - [en](https://longbridge.com/en/news/279755722.md)
  - [zh-HK](https://longbridge.com/zh-HK/news/279755722.md)
---

> Supported Languages: [简体中文](https://longbridge.com/zh-CN/news/279755722.md) | [繁體中文](https://longbridge.com/zh-HK/news/279755722.md)


# Tencent's AI Gamble: Backend Pretends to be Poor, Frontend Goes Crazy

On March 18, 2026, Tencent released its 2025 annual report, with total revenue of 751.8 billion yuan, a year-on-year increase of 14%; Non-IFRS operating profit was 280.7 billion yuan, a year-on-year increase of 18%.

This is a report card that the market cannot find major faults with. But the truly interesting part is in the conference call.

President Liu Chiping said during the earnings call regarding the disappointing capital expenditure on computing power that, in essence: **In 2025, due to GPU supply constraints, the company couldn't buy the cards, and if it can buy more cards in 2026, it will definitely increase investment.**

Translated, it means **it's not that they don't want to spend money on cards, it's that they can't buy them.**

The day after the earnings report was released, Tencent's stock price fell more than 6%, and its market value once again dropped below 5 trillion Hong Kong dollars, seemingly a ruthless footnote to the market's rejection of this explanation.

At a time when major companies are all suffering from a fear of insufficient AI computing power, funds are casting a vote of no confidence in Tencent's conservative AI spending.

It's worth noting that if the narrative of "can't buy cards" had appeared over two years ago when high-end card imports were just restricted, it might have been understandable.

But now it's 2026. The same external environment, the same compliance pressures, and the market estimates that ByteDance's capital expenditure for 2025 will soar to over 160 billion yuan—of which about 90 billion yuan is directly used for AI computing power procurement.

Alibaba has even officially announced: over the next three years, it will invest at least 380 billion yuan in cloud and AI infrastructure construction. Actual spending in 2025 has already exceeded 100 billion yuan.

This perhaps raises a sharper question:

If ByteDance and Alibaba can buy them, why can't Tencent "buy them"?

Is it truly a physical inability to buy, or is it a reluctance to spend?

![Image](https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/ab4d6e6e-4768-40b6-800e-a4e6032fa152.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg)

## **The Human Affairs of Buying Cards**

First, let's look at the data.

Tencent's total capital expenditure for 2025 was 79.2 billion yuan, which was 76.8 billion yuan in 2024.

In the most intense year of the AI arms race, Tencent's capital expenditure only grew by 3%. During the same period, revenue grew by 14%, and the proportion of capital expenditure to revenue actually decreased from 12% to 10.5%. This is significantly lower than the earlier guidance of "low double-digit percentage."

While the absolute value is not small, in comparison:

Institutions estimate that ByteDance's capital expenditure for 2025 will soar to 160 billion yuan, of which about 90 billion yuan is for AI computing power, and an additional nearly 50 billion yuan will be invested in Southeast Asia. Alibaba's three-year investment plan of 380 billion yuan is in full swing, with over 100 billion yuan invested in a single year in 2025, even if it means using AMD alternatives like the MI308 to "downgrade consumption."

Looking back at Tencent. Liu Chiping also revealed another detail during the conference call: in 2025, the company supplemented resources through renting computing power while moderately reducing external sales to ensure internal demand.

On one hand, they say "supply is constrained" to explain the slowdown in expenditure, while on the other hand, they say "supplementing through rentals" indicates that they are not truly unable to use it The two sentences are not contradictory when placed together, as they refer to two different matters:

**The subtext of the first sentence is that, at this point in time and at this premium, it is impossible to buy cards that meet Tencent's ROI expectations.**

**The subtext of the second sentence is that the existing stock is sufficient; Tencent is not running bare.**

It's not that the cards cannot be bought; rather, due to Tencent's capital discipline, they are unwilling to make large purchases at high premiums.

Zhang Kun from E Fund has a good saying: the most important measure of a company's management level is to look at its capital allocation ability.

Liu Chiping's words sound more like an extremely clever public relations excuse. It accomplishes two tasks simultaneously:

First, it provides an explanation to the capital market that the slowdown in capital expenditure is not because Tencent does not value AI, but rather due to supply chain issues;

Second, it conceals Tencent's choice to "slow down and follow" in the underlying large model infrastructure competition.

## **Prudent Participation**

The question arises, since there are no card purchases, where did the saved money go?

**The answer is that it was returned to shareholders.**

In 2024, Tencent repurchased HKD 112 billion, retaining its title as the king of buybacks in Hong Kong stocks. In 2025, it continued to spend about HKD 80 billion, all of which will be canceled, further compressing the total share capital to around 9.15 billion shares, the lowest in its history since listing.

In addition, the year-end dividend per share in 2025 is HKD 5.30 (an 18% year-on-year increase), meaning Tencent has returned over HKD 240 billion in cash to shareholders in the past two years.

Liu Chiping also hinted: considering the return potential in the AI field, there may be a reduction in buybacks when the stock price is relatively low, allocating more resources to AI.

Let's do a simple calculation: Tencent's capital expenditure in 2025 is HKD 79.2 billion, with buybacks of about HKD 80 billion and dividends of about HKD 48 billion during the same period. R&D investment reached HKD 85.7 billion, a historical high.

The money Tencent earns is divided into two streams: **one flows to computing infrastructure, and the other flows into shareholders' pockets. In terms of amount, the total of buybacks and dividends has already exceeded capital expenditure.**

This is a very clear signal of capital allocation: between investment in AI infrastructure and shareholder returns, Tencent has chosen to balance, even slightly leaning towards the latter.

Liu Chiping candidly explained this logic in a conference call: **The primary principle of the company's capital allocation is to create returns for the company and shareholders.**

AI has enormous potential, so an investment strategy must be formulated that invests in the company's future while continuing to provide current returns to shareholders through dividends and buybacks.

**In other words, Tencent views the AI arms race as a marathon rather than a sprint.**

ByteDance and Alibaba have chosen a sprinting approach, investing heavily to seize the initiative, betting on first-mover advantages;

Tencent has chosen a different approach: controlling the pace, maintaining stamina, and waiting for competitors to make mistakes.

The vast territory of existing businesses may be the source of this prudent confidence.

WeChat has 1.418 billion monthly active users. The gaming segment's total revenue is HKD 241.6 billion, a year-on-year increase of 22%, with international market game revenue exceeding USD 10 billion for the first time. Advertising revenue is HKD 145 billion, a 19% increase. Gross margins continue to improve, with value-added service gross margins reaching 60% The monopoly position in social and gaming may allow Tencent to have the capital to "wait," waiting for computing power costs to decline with Moore's Law, waiting for Huawei's Ascend and other domestic alternatives to mature, and waiting for breakthroughs from open-source projects like DeepSeek to level the model gap.

In fact, a year ago, the emergence of DeepSeek partially validated this logic.

At that time, Tencent's management clearly stated in a conference call: With DeepSeek's technological breakthroughs, the entire industry no longer needs to increase GPU procurement as rapidly as previously expected. The training efficiency of existing resources has significantly improved.

The money saved has not disappeared; it has transformed into another form of value, possibly resulting in thicker earnings per share, higher dividends, and a lower total share capital.

For value investors, this is precisely what matters most.

## Nintendo's Strategy

To understand Tencent's strategic choices, it may be necessary to look back further into business competition for clues.

In 1989, a century-long battle was staged in the handheld gaming market, with three competitors: Atari Lynx, Sega Game Gear, and Nintendo Game Boy.

Looking solely at hardware specifications, the Game Boy seems like a joke, as the industry was extremely enthusiastic about the revolutionary experience of color backlit screens at that time.

The Lynx featured a 4096-color backlit screen, a hardware scaling and rotation engine, and even supported eight-player multiplayer, making it the "fully powered flagship" of its time; the Game Gear also came equipped with a color backlit screen and was based on the Sega Master System architecture, capable of running an entire library of mature platform games.

What about the Game Boy? It had a green monochrome LCD screen without backlighting. No color, no scaling, and it was even hard to see in sunlight.

![Image](https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/feeb385e-fad9-46c0-a2e5-654c8c6bde41.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg)

Clearly, the Game Boy was at an absolute disadvantage in terms of technical specifications.

However, Gunpei Yokoi, who led Nintendo's hardware division at the time, made a counterintuitive choice: **not to pursue the cutting-edge technology but to combine those low-cost mature technologies to create a brand new user experience.**

This strategic philosophy was later summarized in one phrase: **the level-headed thinking of withered technology.**

What did the Game Boy gain by giving up the color screen?

**First, battery life.** The Game Boy could run for 30 hours on four AA batteries, while the Lynx and Game Gear required six AA batteries, with battery lives of only 4-5 hours and 3-5 hours, respectively.

One could accompany players from Beijing to Shenzhen by train, while the other two couldn't even last to Shijiazhuang.

**Second, price.** The Game Boy was priced at $89.99, while the Game Gear was $149.99 and the Lynx was $179.95, nearly double the price difference.

**Finally, portability.** The Game Boy was compact and lightweight enough to fit in a pocket, while the Lynx and Game Gear were bulky "bricks." The final result is that the Game Boy series sold approximately 119 million units over its lifetime. Game Gear sold 14 million units. Lynx sold about 3 million units.

Nintendo won against its most advanced competitors with hardware that was a generation behind, nearly 10 times over.

The key here is that Gunpei Yokoi saw something that everyone chasing parameters could not see: **the core competitiveness of handheld devices is not the graphics, nor the parameter battles, but the ability to play anytime and anywhere.**

For every bit of performance piled on the hardware, there is a loss in battery life, portability, and price, and these three factors are crucial in determining whether users will actually put the device in their pockets.

Looking back at today's AI large model battle, it remains to be seen whether it will play out like the handheld console wars of the past.

ByteDance and Alibaba are pursuing extreme hardware-side offensives, just as Sega and Atari once sought the most extreme hardware configurations and graphical performance, which translates today into computing power and model benchmark scores.

Their belief is that the intelligence gap of the underlying models determines the life and death of the application layer. They are willing to burn through massive capital, just like Lynx consumed six batteries for a color backlit screen.

**The approach Tencent is currently imitating is akin to Nintendo in 1989; however, Tencent's current situation is undoubtedly much more perilous than that of Nintendo back then.**

The Hunyuan large model has performed well in some international benchmark tests, the open-source MoE model's performance rivals that of hundred-billion parameter models, and the download count for 3D generation models exceeds 3 million.

However, Hunyuan is hardly seen in rankings within the AI circle, and Tencent seems to have not invested much in climbing the charts.

Liu Chiping confirmed this judgment during the Q3 2025 earnings call: **"Currently in the Chinese market, we do not believe there is an absolutely leading model; all parties in the industry are in a fierce chase, and different models have their advantages in different usage scenarios."**

The implication here is that the gap between underlying models is being rapidly closed. At this stage, putting all chips on "training larger models" is leading to a sharp deterioration in return on investment.

Back in 1989, spending twice the price for a color screen only to find it runs out of battery in three hours, the marginal returns of performance are being consumed by the marginal increases in cost.

The Game Boy did not need a color screen to support Tetris; similarly, Tencent seems unwilling to compete for a few tenths of a percentage point in model rankings, especially in the current climate where global giants are hoarding graphics cards and memory prices are rising.

As long as the underlying model can reach the "usable" passing line, Tencent may want to win in its own way.

## Agent's Major Offensive

If one only looks at Tencent's hesitation in purchasing cards, one might think this company has fallen asleep in the AI era.

But if we shift our focus to its response since March following the Openclaw craze earlier this year, Tencent's organization of live installations under the office building in Nanshan and the waves of lobster special forces resemble an activated, fierce, and cost-ignoring giant beast.

During the Spring Festival of 2026, Tencent spent 1 billion yuan in cash to promote Yuanbao, marking a rare large-scale cash giveaway during the Spring Festival since WeChat Pay in 2015. The "Yuanbao faction" conducted a small-scale public test, and Yuanbao was also migrated into the WeChat contact list as a contact **Backend is stingy, frontend is generous. Behind this sense of disconnection is an extremely clear strategic choice: Tencent's battlefield is not in the underlying model, but in the application ecosystem.**

The reasons boil down to three points:

**First, applications are closest to users.** Agents are directly facing users and can immediately generate interaction and commercial value. Tencent's advertising revenue is expected to grow by 19% to 145 billion yuan by 2025, with AI-driven precise advertising placement technology as the core driver. Every penny invested in the frontend can quickly pay off through improved advertising conversion rates.

**Second, leverage home-field advantage.** WeChat has 1.4 billion monthly active users, covering all scenarios such as social networking, public accounts, video accounts, mini-programs, e-commerce, and payments. Others need to spend a lot of money to acquire traffic for Agents, while Tencent's Agents only need to be placed on WeChat to become a national-level entry point.

**Third, a little effort can yield great results.** A good Agent experience can largely compensate for the slight gap in intelligence of the underlying model. Just like the monochrome screen of the Game Boy does not hinder the addictive nature of Tetris—fun is the key. WeChat AI search does not require AGI, just as Pokémon does not need 4096 colors.

Here is a set of intriguing comparative figures:

By the end of 2025, Tencent's Yuanbao monthly active users are expected to exceed 100 million, and the AI workbench ima monthly active users will surpass 13 million. The intelligence level of the Hongyuan 3.0 large language model continues to improve, with AI tools like WorkBuddy and QClaw being launched one after another.

Liu Chiping stated that the total investment in AI new products will exceed 18 billion yuan in 2025, and will at least double in 2026.

If we switch the statistical caliber to "ecological reach," Yuanbao has already been embedded in WeChat friend lists, public accounts, video account comment sections, and WeChat search, meaning its actual potential reach theoretically encompasses all 1.418 billion monthly active users of WeChat.

One might ask, if a developer of a vertical application is willing to build an AI Agent, which platform would they choose: an independent app ecosystem that requires separate customer acquisition, or a super platform with 1.3 billion traffic?

The answer is self-evident.

## Heavy Stakes

In summary, if we were to summarize Tencent's overall actions, it could be:

**Be a "follower" in underlying computing power, and a "ruler" in the application ecosystem.**

The prudence of the backend releases a large amount of cash flow for shareholder returns and strategic reserves; the aggressiveness of the frontend, using the WeChat ecosystem as a fulcrum, leverages the traffic dividend of the Agent era.

The underlying logic of this strategy is quite similar to Nintendo's survival philosophy in the handheld console war—**not competing with opponents on their home turf with hard metrics, but redefining the dimensions of competition.**

However, this strategy also has its Achilles' heel.

If the gap in the underlying model becomes too large, to the extent that the application layer experience cannot compensate, then Tencent's "sufficient" strategy will turn into "insufficient."

Just like if the Game Boy's screen is so poor that the game graphics are completely unrecognizable, no matter how cheap the price or how long the battery life, it won't save it Ultimately, the outcome of this gamble will depend on two variables:

**First, is the performance gap of the underlying models continuing to converge, or is it widening again?**

If the open-source community and domestic chips continue to narrow the gap, Tencent's strategy is correct; if OpenAI suddenly makes a breakthrough by an order of magnitude, all "followers" will face the risk of being eliminated.

**Second, can truly revolutionary AI-native applications emerge from the WeChat ecosystem?**

If successful, this will be another classic case of "winning without fighting" in business history; if unsuccessful, every penny Tencent saves today will become a sunk cost of missed opportunity

### Related Stocks

- [Guotai CSI Animation Game ETF (516010.CN)](https://longbridge.com/en/quote/516010.CN.md)
- [iShares MSCI China ETF (MCHI.US)](https://longbridge.com/en/quote/MCHI.US.md)
- [Tencent Holdings Limited (TCTZF.US)](https://longbridge.com/en/quote/TCTZF.US.md)
- [ChinaAMC CSI Animation Game ETF (159869.CN)](https://longbridge.com/en/quote/159869.CN.md)
- [First Trust S-Network Strmng & Gmng ETF (BNGE.US)](https://longbridge.com/en/quote/BNGE.US.md)
- [TENCENT (00700.HK)](https://longbridge.com/en/quote/00700.HK.md)
- [VanEck Vdo Gaming and eSprts ETF (ESPO.US)](https://longbridge.com/en/quote/ESPO.US.md)
- [Amplify Video Game Leaders ETF (GAMR.US)](https://longbridge.com/en/quote/GAMR.US.md)
- [Tencent Holdings Limited (TCEHY.US)](https://longbridge.com/en/quote/TCEHY.US.md)
- [Global X Video Games & Esports ETF (HERO.US)](https://longbridge.com/en/quote/HERO.US.md)
- [KraneShares CSI China Internet ETF (KWEB.US)](https://longbridge.com/en/quote/KWEB.US.md)

## Related News & Research

- [Pony AI Taps WeChat's Billion Users To Supercharge Robotaxi Bookings](https://longbridge.com/en/news/279027140.md)
- [OpenClaw Founder Accuses Tencent of “Copying,” Company Responds Citing Support for the Ecosystem](https://longbridge.com/en/news/278848478.md)
- [Tencent to Double AI Investment, President Says After Chinese Tech Giant Delivers Solid Earnings Growth](https://longbridge.com/en/news/279717961.md)
- [Alibaba Unveils 'Wukong' AI To Automate Office Work, Replace Daily Chores](https://longbridge.com/en/news/279427888.md)
- [Baidu joins China's OpenClaw frenzy with new AI agents](https://longbridge.com/en/news/279465870.md)