Recently, Tencent handed over a "steady and boring" financial report: annual revenue of 751.8 billion yuan, a year-on-year increase of 14%; adjusted operating profit was 280.7 billion yuan, an increase of 18%.

Looking at the data alone, this report alone cannot single out a big problem. But the real drama was in the subsequent conference call.

In the face of investors' doubts about the "not strong enough" investment in AI computing power, President Martin Lau explained: In 2025, it is not that I don't want to buy more GPUs (AI chips), but that I can't buy them at all. He also said that if more chips can be bought in 2026, Tencent will definitely increase investment.

To put it bluntly: it's not that you don't burn money, it's that the card has been robbed by others.

However, the market does not buy it. The day after the earnings report, Tencent's stock price plummeted by more than 6%, and its market value fell below HK$5 trillion again - as if responding with real money: "We don't believe in this reason." ”

Nowadays, major technology companies are frantically increasing their AI computing power, for fear of falling behind. Tencent's relatively conservative AI investment is being regarded as a "lagging signal" by investors, and funds are quietly voting with their feet.

It's not that I can't buy it, it's that I don't want to spend money indiscriminately

Tencent's capital expenditure for the full year of 2025 will be 79.2 billion yuan, only 3% more than the 76.8 billion yuan in 2024.

But this year was the most intense time for the AI "arms race" - and Tencent's revenue increased by 14%. As a result, capital expenditure as a percentage of revenue fell from 12% to 10.5%, far below the expectation of "low double-digit growth" at the beginning of the year.

The money is not spent, but compared with peers, the gap comes out:

According to institutional estimates, ByteDance's capital expenditure will soar to 160 billion yuan in 2025, of which about 90 billion will be smashed into AI computing power, and nearly 50 billion yuan will be added in Southeast Asia; Alibaba is promoting a three-year investment plan of 380 billion yuan, and the investment in a single year in 2025 has exceeded 100 billion, even if it is "downgraded and replaced" with more cost-effective AMD chips (such as MI308).

Let's look at Tencent again: President Martin Lau explained on the conference call that in 2025, due to GPU supply constraints, "you can't buy a card". But then he mentioned that the company actually supplemented resources by renting computing power, etc., and also reduced external sales and prioritized internal use.

These two sentences seem contradictory, but they are not in conflict: "I can't buy a card" really means: it's not that I can't buy it at all, but that I don't want to spend a lot of money when the price is high and the return is uncertain. "Renting computing power" shows that Tencent did not break the computing power, but chose a more cautious strategy.

Ma Huateng talks about "shrimp farming": "decentralization" like a mini program

At the communication meeting before the earnings call, Ma Huateng publicly talked for the first time about the "shrimp farming" (that is, the development of AI agents, internally codenamed "lobster") that Tencent is exploring.

He said that the biggest value of "lobster" is that AI is no longer limited to chatbots, but can be embedded in various real-world scenarios - such as e-commerce, customer service, content creation, etc. In this way, Tencent's various business lines can work together to truly use AI.

He also specifically mentioned that WeChat Mini Programs have always adhered to the concept of "decentralization" - not to engage in a unified entrance, but to allow every merchant and developer to have their own users and traffic. This idea can also be used in future "lobster" applications.

Why is it important? Because many service providers are most afraid of being "short-circuited" by the platform - working hard to provide services, as a result, all users are taken away by the platform and become tool people. Therefore, when Tencent designs the AI ecosystem, it will respect the independence of partners.

Ma Huateng emphasized: The future "shrimp farming" strategy should not only have centralized coordination, but also retain the freedom of decentralization. AI agent partners want to get traffic and ingress without becoming a "plug-in" that is called at will.

"This is a long-term project," he said, "and everyone has to be a little patient, not in a hurry." ”

When will AI investment take effect?

In response to analysts' questions about AI investment returns and profit margins in 2026, Tencent's management gave an answer. CFO Luo Shuohan said frankly: In the future, there may be a situation of "revenue rising quickly and profits rising slowly", but the company is not worried - because this shows that new businesses are expanding rapidly and are laying out in advance for greater opportunities.

James Mitchell, Chief Strategy Officer, further explained: The use of AI in existing businesses has seen obvious results. If you don't count the investment in new AI products, the profitability of the old business is actually very strong.

However, new products such as AI agents and subscription services are still in their early stages - user payment habits have not yet formed, and enterprises are still trying out, so they need to invest first and then harvest.

He used Tencent Cloud as an analogy: he also lost money for several years before gradually making a profit. "The same goes for AI, there is a time lag between investment and return, but it's worth the wait."

Tencent does not develop its own AI chips for the time being

Some investors asked Tencent if it would develop its own AI chips like Nvidia or Alibaba, and the management clearly responded: This is not the point at present.

They explained that AI chips are divided into two types - training chips and inference chips, and the strategies are completely different: the technical threshold for training chips (used to "teach" large models) is extremely high, and only one or two companies such as NVIDIA can provide them in the world. Tencent's strategy is to do its best to get the most advanced training chips and iterate the model quickly. For example, the newly released Hunyuan 3.0 is just the beginning, and the model update will be faster in the future.

Inference chips (used to "use" models, such as answering user questions) are more cost-conscious. The good news is that there are many inference chip choices in the Chinese market, fierce competition, and more reasonable prices, unlike training chips.

Therefore, Tencent's current priority is very clear: first use the best computing power to train the model, make the product, and make users like it. Wait until the product runs through and scales up, and then consider how to optimize the inference cost - including whether to develop its own chips.

Jingtai point of view|Short-term pressure, long-term ecological realization

For Hong Kong stock investors (0700. HK）：

Short-term fluctuations are inevitable: the market is sensitive to AI narratives, and if there is no blockbuster product release in Q2, valuations may continue to be under pressure; But the long-term value has not been lost: the cash bull of WeChat + games + advertising is still strong, and the restraint of AI investment has ensured free cash flow.

Key observation points: the implementation progress of the "lobster" agent on WeChat/Enterprise WeChat; whether Tencent Cloud's AI services can rebound; After the easing of GPU supply in 2026, will capital expenditure increase significantly?

Mapping of the AI industry chain:

Good for domestic inference chip manufacturers (such as Cambrian and Haiguang): Tencent has made it clear that the inference end is willing to use domestic solutions; Bearish pure computing power leasing platform: Tencent tends to control resources and has limited external procurement; Pay attention to AI SaaS service providers: If Tencent opens up the "lobster" ecosystem, vertical applications may welcome new opportunities.

Risk warning:

If AI products cannot be realized in 2026, the market's patience may run out. If Nvidia further tightens the supply of high-end GPUs, model iterations may be restricted.