Tencent's AI Gamble: Playing Poor in the Backend, Going Wild on the Frontend

```

On March 18, 2026, Tencent released its 2025 annual report: total revenue for the year reached 751.8 billion yuan, up 14% year-on-year; Non-IFRS operating profit was 280.7 billion yuan, up 18% year-on-year.

This is a performance sheet that the market can hardly fault. But the real interesting stuff happened on the earnings call.

President Liu Chiping addressed the below-expectation capex for computing power on the earnings call, essentially saying: In 2025, due to GPU supply constraints, the company couldn’t buy cards; if we can buy more cards in 2026, we will definitely invest more.

Translated, that means it’s not that we don’t want to spend money to buy cards—it’s that we can’t get them.

The day after the financial report was released, Tencent’s share price instantly dropped more than 6%, and its market capitalization once again fell below HK$5 trillion, seemingly a ruthless footnote of the market’s rejection of this explanation.

At a time when all tech giants are suffering from firepower insufficiency in AI computing, capital is voting “no confidence” with its feet against Tencent’s conservative AI spending.

It’s worth noting that the “can’t buy cards” narrative would have been understandable more than two years ago, when high-end card imports were just being restricted.

But now it’s 2026. With the same external environment and compliance pressures, it’s estimated that ByteDance’s 2025 capex soared to more than 160 billion—of which about 90 billion was directly spent on AI computing.

Alibaba has even officially announced: at least 380 billion will be invested for cloud and AI infrastructure in the next three years, with actual spending in 2025 already exceeding 100 billion.

This perhaps points to an even sharper question:

If ByteDance and Alibaba can buy, why is it that only Tencent “can’t buy”?

Is it physically can’t buy, or mathematically don’t want to buy?

Buying Cards—The Art of Relationships

First, let’s look at the numbers.

Tencent’s full-year capital expenditure in 2025 was 79.2 billion yuan, up from 76.8 billion in 2024.

In the most intense year of the AI arms race, Tencent’s capex grew only 3%. Meanwhile, revenue grew by 14%, so capex as a proportion of revenue dropped from 12% to 10.5%. That’s much lower than the “low double-digit percentage” guidance at the start of the year.

The absolute value isn’t small, but compared with peers:

Estimates show ByteDance’s 2025 capex surged to 160 billion, with about 90 billion for AI computing, and an additional nearly 50 billion in Southeast Asia. Alibaba’s 380 billion three-year investment plan is at full throttle, with over 100 billion invested in 2025 alone—even mixing in AMD’s MI308 as substitutes for “downgraded consumption.”

Looking back at Tencent, Liu Chiping also revealed another detail on the call: In 2025, the company supplemented resources via leasing computing power, while moderately reducing external sales to assure internal needs.

He first attributed spending slowdown to “supply constraints,” then mentioned “supplementing via leasing,” signaling it’s not that they really can’t use it.

These two sentences aren’t contradictory—they’re about two different things:

The subtext of the first is: at this moment, at this premium, they can’t buy cards meeting Tencent’s ROI expectations.

The subtext of the second is: the current inventory is sufficient, Tencent is not running “bare.”

The cards aren’t truly unavailable, but Tencent’s capital discipline means they won’t buy at high premiums.

Zhang Kun from E Fund Management once said: to measure a company’s management, the most important thing is its ability to allocate capital.

Liu Chiping’s remarks sound more like a highly sophisticated PR pretext. They accomplish two things:

One, to explain to the capital market that the capex slowdown isn’t because Tencent doesn’t value AI, but because of supply chain issues;

Two, to mask Tencent’s choice to “slow down and follow” in the foundational LLM infrastructure race.

Cautious Engagement

So now, the question: since they didn’t buy cards, where did the money go?

The answer: It was returned to shareholders.

In 2024, Tencent repurchased HK$112 billion, continuing as Hong Kong’s repurchase king. In 2025 another HK$80 billion repurchased—all canceled, reducing total shares to around 9.15 billion, the lowest since IPO.

Add a year-end dividend of HK$5.30 per share (up 18% YoY), Tencent returned over HK$240 billion to shareholders in the past two years.

Liu Chiping indeed hinted: Considering AI’s potential returns, repurchases might decrease during lower share prices in the future, with more resources invested in AI.

A simple calculation: Tencent’s 2025 capex was 79.2 billion yuan, buybacks about HK$80 billion, dividends about HK$48 billion, and R&D spending at a record high of 85.7 billion yuan.

Tencent’s profits go down two pipelines: One towards computing infrastructure, one towards shareholders’ pockets. In terms of amount, buybacks + dividends already surpass capex.

This is a clear capital allocation signal: between AI infrastructure and shareholder returns, Tencent chooses balance—and leans slightly towards the latter.

Liu Chiping admitted this logic at the conference: The primary principle of capital allocation is to generate real returns for the company and its shareholders.

AI’s potential is massive, so they must formulate an investment strategy: invest in the future, and continue offering present returns via buybacks and dividends.

In other words, Tencent sees the AI arms race as a marathon, not a 100-meter dash.

ByteDance and Alibaba opt for a sprint, pouring in huge amounts to seize the first-mover advantage;

Tencent takes another approach: controlling pace, conserving strength, waiting for opponents to stumble.

The grand scale of Tencent’s existing businesses likely fuels this cautious confidence.

WeChat has 1.418 billion monthly active users. Games segment revenue reached 241.6 billion yuan, up 22%; international games revenue broke US$1 billion for the first time. Advertising revenue reached 145 billion yuan, up 19%. Gross margin continues rising; value-added services gross margin is at 60%.

Its monopolistic position in social and games may give Tencent the capital to “wait”—wait for computing costs to drop per Moore’s Law, wait for domestic alternatives like Huawei Ascend to mature, wait for open-source breakthroughs like DeepSeek to close the model gap.

Indeed, DeepSeek’s emergence a year ago has validated this logic.

At the time, management said on a call: with DeepSeek’s breakthroughs, the industry no longer needs to rapidly increase GPU purchases as previously expected. Existing resources’ training efficiency is greatly improved.

The money saved hasn’t disappeared—it becomes another form of value: thicker EPS, higher dividends, lower total shares.

For value investors, this is exactly what matters most.

Nintendo Techniques

To understand Tencent’s strategic choice, we might need to look back to earlier business competitions.

In 1989, Japan’s handheld gaming consoles staged a battle of the century, with three competitors: Atari Lynx, Sega Game Gear, and Nintendo Game Boy.

On hardware specs alone, Game Boy was almost a joke, as the industry was obsessed with investing in color backlit screens for revolutionary experiences.

Lynx had a 4096-color backlit screen, hardware scaling/rotation engine, and even supported 8-player multiplayer—it was the “full-blooded flagship” of the year; Game Gear also had a color backlit screen, hardware based on Sega’s Master System architecture able to directly run its game library.

Game Boy? A green monochrome LCD with no backlight, no color, no scaling, almost unreadable under sunlight.

Clearly, Game Boy was at an absolute hardware disadvantage.

But Nintendo hardware lead Gunpei Yokoi made a counterintuitive choice: Don’t chase cutting-edge tech; instead, use cheap, mature tech to assemble a novel user experience.

This strategic philosophy became known as: Lateral Thinking with Withered Technology.

What did Game Boy gain by sacrificing color?

First, endurance, Game Boy could run for 30 hours on four AA batteries, while Lynx and Game Gear needed six AA batteries and lasted just 4–5 hours and 3–5 hours respectively.

One could keep a player entertained all the way from Beijing to Shenzhen by train; the other two couldn’t even last to Shijiazhuang.

Second, price, Game Boy cost $89.99, Game Gear $149.99, Lynx $179.95—a near twofold difference.

Lastly, portability, Game Boy was small and pocketable, while Lynx and Game Gear were both bulky “bricks.”

The end results: Game Boy sold around 119 million units; Game Gear sold 14 million; Lynx sold about 3 million.

Nintendo used hardware a generation behind to beat the most advanced rivals almost tenfold.

The key is: Gunpei Yokoi saw what spec-chasers didn’t: Handhelds core competitiveness wasn’t graphics or specs—it was play-anywhere capability.

For every bit of performance piled on hardware, you lose endurance, portability, and price—and those three decide whether users actually put the device in their pocket.

Turning back to today’s AI LLM wars: will history repeat itself as the handheld console battles did? It remains to be seen.

ByteDance and Alibaba pursue extreme hardware offense, just as Sega and Atari pursued ultimate hardware and graphics—today, that means computing power and model benchmarks.

Their belief: intelligence gaps at the model layer decide death and survival at the application layer. They’ll burn huge capital if needed, like Lynx devouring 6 batteries for a color screen.

Tencent’s current logic mirrors Nintendo in 1989—but the challenges Tencent faces today are far harsher than Nintendo then.

The Hunyuan model performs well in some international benchmarks; its open-source MoE model matches trillion-parameter models; the 3D generation model has over 3 million downloads.

But Hunyuan is almost never seen on leaderboards in the AI world; Tencent doesn’t seem to care much about ranking.

Liu Chiping confirmed this in the 2025 Q3 earnings call: "Currently, in the Chinese market, we don’t think there is an absolutely leading model; all players are fiercely catching up, and different models have their own advantages in different scenarios."

The subtext: gaps at the model layer are closing fast. Betting everything on “training larger models” now is a rapidly deteriorating ROI prospect.

Back in 1989, paying twice as much for a color screen that died in three hours, the marginal benefits of performance were eaten up by the marginal cost increases.

Game Boy didn’t need a color screen for Tetris; similarly, Tencent doesn’t want to chase tenths of a percent in model rankings, especially when global tech giants are hoarding GPUs and RAM prices are surging.

As long as the foundational model hits the “good enough” bar, Tencent may want win with its own approach.

Agent Goes on the Offensive

If you only look at Tencent’s hesitance to buy cards, you’d think the company is “asleep” in the AI era.

But if you shift focus to its reaction since Openclaw’s hype at the start of the year, with in-office assembly lines and “lobster agent teams” under its Nanshan office, it looks activated and fiercely cost-irrelevant.

Spring Festival 2026: Tencent spent 1 billion yuan to promote Yuanbao, a move almost unseen since WeChat Pay in 2015. “Yuanbao Party” saw small-scale pilot and Yuanbao entered WeChat contacts.

Frugal on the backend, generous on the frontend. This split highlights a clear strategic decision: Tencent’s battlefield isn’t at the model layer, but in application ecosystems.

There are three main reasons:

First, applications are closest to users. Agents directly interface with users and convert business value immediately. Tencent ad revenue grew 19% in 2025 to 145 billion yuan; AI-driven ad targeting is the main driver. Every yuan spent upfront recoups quickly via ad conversion rates uplift.

Second, leverage home-field advantage. WeChat has 1.4 billion MAU, covering social, Official Accounts, Channels, mini-programs, e-commerce, payments—full-scenario coverage. Others need big spending to get traffic for Agents; Tencent’s agents on WeChat are already super apps.

Third, leverage efficiency. A good agent UX can compensate for small deficits in model IQ. Just as Game Boy’s monochrome didn’t hinder Tetris addiction—fun is king. WeChat AI search doesn’t need AGI, just as Pokémon didn’t need 4096 colors.

Consider these interesting numbers:

By the end of 2025, Yuanbao MAU had passed 100 million, AI Workbench ima MAU over 13 million. Hunyuan 3.0’s language intelligence keeps rising, WorkBuddy, QClaw and other AI agent tools launched.

Liu Chiping says Tencent invested over 18 billion yuan in new AI products in 2025, and this will at least double in 2026.

Switching to “ecosystem reach” metrics: Yuanbao is embedded in WeChat friend lists, Official Account comments, Channels, WeChat Search—its potential reach is theoretically WeChat’s entire 1.418 billion MAU.

Suppose a vertical app developer—would you build your AI Agent on a standalone app needing its own user acquisition, or a super platform with 1.3+ billion traffic?

The answer is obvious.

Heavy Bet

In summary, Tencent’s overall moves can be summed up as:

Be a “follower” at the foundational computing layer; be a “dominator” in app ecosystem.

Backend caution frees cash flow for shareholder returns and strategic reserves; frontend aggression leverages WeChat ecosystem to seize Agent era’s traffic dividends.

This logic echoes Nintendo’s handheld survival philosophy: Don’t fight competitors on their home field with hard metrics—redefine the competition’s dimensions.

Yet, there’s a fatal flaw.

If model layer gaps grow so large that application experiences can no longer compensate, Tencent’s “good enough” strategy turns into “not enough.”

It’s like if Game Boy’s screen became so bad users couldn’t see gameplay, neither price nor endurance would save it.

Ultimately, victory or defeat in this bet hinges on two variables:

First: Will model layer performance gaps keep converging, or re-expand?

If the open-source community and domestic chips keep closing the gap, Tencent’s strategy is right; if OpenAI makes a leap, all “followers” risk being eliminated.

Second: Can the WeChat ecosystem sprout truly revolutionary AI-native applications?

If yes, this is another “win without fighting” in business history; if not, every yuan saved now by Tencent will become sunk costs for a missed era.

Risk Warning and DisclaimerThe market involves risks, investment should be cautious. This article does not constitute personal investment advice and does not take into account individual users’ specific investment objectives, financial situation, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their circumstances. Investments based on this are at your own risk. ```