Tencent Admits AI Lag, Accelerates Investments and Product Development

Computing power has indeed been Tencent's most tangible weakness over the past year.

Article author, source: 0x9999in1, ME News



• What’s most noteworthy about Tang Daosheng’s remarks this time isn’t “AI is a marathon,” but rather that Tencent, for the first time, explicitly grouped together issues of insufficient computing power, stalled model training, product rollout missteps, and near-cancellation of internal projects for review. For a large company accustomed to being known for its caution, this level of candor itself indicates that the pressure has become significant.
• Computing power has indeed been Tencent's most tangible weakness over the past year. In Q1 2026, Tencent's capital expenditure was RMB 31.9 billion, rising sharply to RMB 52.8 billion in Q2, bringing the total for the first half of the year to RMB 84.7 billion—exceeding the full-year capital expenditure of approximately RMB 79 billion in 2025. Tencent is not unwilling to spend; rather, it is clearly accelerating its efforts to catch up.
• But “insufficient computing power” cannot explain all the issues. In June 2026, according to QuestMobile, DouBao had 382 million monthly active users, Qwen had 167 million, DeepSeek had 130 million, and Yuanbao had 49.84 million. Tencent’s gap in C-end AI entry points is real—and it is no longer just a model issue, but also one of product timing and user perception.
• The real card worth reevaluating for Tencent is not the standalone HunYuan, but the ecosystem network composed of 1.439 billion monthly active WeChat and WeChat accounts, Enterprise WeChat, Tencent Meeting, Tencent Docs, cloud services, and the developer ecosystem. The issue is that ecosystems are merely resources—they don’t automatically equate to a moat.
• The significance of WorkBuddy and CodeBuddy may be greater than that of Yuanbao. They demonstrate that Tencent has established a closed loop where products drive model improvements and models enhance products in the areas of AI agents, office, and development tools, indicating that AI is transforming Tencent’s own R&D organization.
• Therefore, Tencent AI is no longer in a position of “responding after the fact”; a more accurate description is that it has moved past its period of hesitation and entered an expensive catching-up phase. Whether Hy4 can deliver, whether computing power can be converted into revenue, and whether use cases can be transformed into high-frequency Agent applications are the three critical questions that will determine the outcome going forward.

On August 26, Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of the Cloud and Smart Industries Group, published a lengthy article in Tencent’s internal publication. It naturally included Tencent’s familiar long-termism: “AI may have only completed the first kilometer of a marathon; enduring longer matters more than starting earlier.” But if you only focus on this sentence, you’re likely to miss the truly significant parts of this article.

What’s truly noteworthy is that Tencent has begun to openly address the AI-related issues from the past few years.

Tang Daosheng acknowledged that the perception that Tencent is moving slowly on AI is not entirely unfounded; there must be some anxiety within the company, as competitors have been more vocal and aggressive in AI over the past few years. More directly, he explicitly stated that Tencent’s overall computing power is “severely insufficient,” affecting not only the training of HunYuan but also slowing product development and “having a significant impact.”

This statement is heavy.

For large model companies, computing power is not an ordinary production factor. The scale of training clusters determines the frequency of experiments, which in turn determines how many architectures, data approaches, and post-training pathways the model team can explore. After entering the Agent era, inference demand will rapidly increase as task chains grow longer. A model company lacking sufficient computing power is essentially like a manufacturing company with insufficient production capacity and inadequate R&D experimentation lines—you not only sell less, but also iterate and test slower than your competitors.

Therefore, the issues Tencent AI has faced over the past two years can no longer be simply understood as "low-key."

Part of it is objectively slow.

More值得关注的是, Tencent has already begun to correct this issue with financial figures.

In 2025, Tencent's annual capital expenditure was approximately RMB 79 billion, and the company's management had previously acknowledged that actual spending fell short of internal targets due to factors such as supply constraints for high-end chips. Entering 2026, capital expenditure reached RMB 31.9 billion in the first quarter and further surged to RMB 52.8 billion in the second quarter, a year-over-year increase of 176%. In just six months, Tencent's capital expenditure reached RMB 84.7 billion, surpassing the entire year of 2025.

This change is very clear: Tencent, which previously maintained a relatively cautious approach to AI infrastructure, is stepping off the stage.

In the second quarter, Tencent's R&D expenditure reached RMB 27.28 billion, a 35% year-over-year increase. More interestingly, regarding profit metrics: Tencent disclosed that its Non-IFRS operating profit for the quarter was RMB 75.6 billion; however, if the revenue, costs, and expenses related to new AI products such as Hy, Yuanbao, CodeBuddy, WorkBuddy, and Xiao Wei are excluded, the operating profit would be approximately RMB 86.1 billion—a difference of about RMB 10.5 billion between the two figures.

These 10.5 billion yuan cannot be simply interpreted as “AI losses of 10.5 billion yuan,” but they are sufficient to illustrate one fact: Tencent’s AI today remains in a clear phase of strategic investment, and this investment has already been genuinely reflected on the income statement.

This is the most important context for Tang Daosheng's article.

Tencent is not telling the market "we aren't actually slow."

It's essentially saying: slowed down, but now ready to pay to catch up.

“Staying the course is more important than starting early,” and historically, this has held true in business.

The ultimate winner in the search era wasn't necessarily the first to enter search, and the final landscape of smartphones wasn't determined by the earliest PDA makers. Microsoft wasn't the first company to make smartphones, yet it has regained strategic initiative in the era of cloud computing and generative AI. The tech industry has numerous examples of latecomers turning the tables.

But the problem is that this AI marathon is different from a regular long-distance race.

It’s a race where you run, build the road, and increase the cost of participation all at the same time.

Model training is becoming increasingly expensive, inference calls are growing larger, data centers are becoming more capital-intensive, top research talent is commanding higher prices, and users are developing habits around their own AI entry points. Therefore, entering a year later isn’t just “running one kilometer less”—it could mean others have already gained more real user data, more mature inference infrastructure, a more stable developer ecosystem, and lower unit costs.

Tencent is facing exactly this pressure.

The AI application data for the first half of 2026 released by QuestMobile clearly illustrates the gap on the consumer side: in June, Doubao had approximately 382 million monthly active users, Qwen had about 167 million, DeepSeek had around 130 million, while Yuanbao had roughly 49.84 million, ranking fourth.

Forty million monthly active users is certainly not a failure, but for Tencent, which has the super entry point of WeChat, this is hardly leading.

More harshly, competitors did not wait for Tencent.

Alibaba has announced plans to invest RMB 380 billion over the coming years to build cloud and AI infrastructure, and a significant portion of this has already been executed as of this August; in the second quarter of this year, Alibaba’s capital expenditures reached approximately RMB 67.7 billion. On August 24, Alibaba completed a share placement of approximately HKD 80 billion, explicitly allocating the funds toward AI development and infrastructure.

This is where the real danger of the "marathon" lies: Tencent does have a chance to catch up, but the leaders ahead haven't slowed down.

So I agree with Tang Daosheng’s judgment on long-termism, but I don’t believe that being an early starter is unimportant.

In the AI industry, the correct statement is: Waking up early doesn't guarantee victory, but running slowly always comes at a cost.

And Tencent is now paying this cost.

So what does Tencent rely on to catch up?

The answer may very well not be to continue focusing all attention on "whether our model can rank first."

As of the end of June 2026, the combined monthly active accounts of WeChat and WeChat reached 1.439 billion. Looking at the enterprise side, Tencent also holds enterprise WeChat, Tencent Meeting, Tencent Docs, Tencent Cloud, knowledge management tools, a developer platform, and a large number of long-standing enterprise customers.

This is the fundamental difference between Tencent and the vast majority of large model startups.

The first challenge for a large model startup is: now that we have the model, where are the users?

Tencent faces a different challenge: users are already there—how should AI be integrated?

This may sound much easier, but it’s not necessarily so.

Because the scenario is not a natural moat.

WeChat has a billion-plus accounts, but that doesn't mean a billion people will use AI within WeChat; Tencent Docs has office users, but that doesn't mean these users are willing to hand over their workflows to Tencent Agent; enterprise customers using Tencent Cloud don't automatically become buyers of Tencent's next AI order.

Scenarios only generate value when truly integrated with models, data, tool calls, permission systems, and workflows.

Tang DaoSheng's statement that "algorithms determine the upper limit, while engineering determines the speed of reaching that limit" is, in my view, more noteworthy than the "AI marathon," as it actually highlights where Tencent may truly establish a competitive advantage.

The Agent era requires not just a conversational model, but an execution environment capable of accessing files, databases, browsers, internal systems, enterprise knowledge bases, and third-party tools. Enterprises must also consider permissions, security, auditing, token costs, and system stability.

These things are not sexy at all, but they're very Tencent.

Over the past two decades, Tencent's core strength has never been just algorithms, but rather reliably connecting massive user bases, complex systems, and business ecosystems.

If AI in the future truly evolves from "answering questions" to "completing tasks," Tencent's previously perceived heavy engineering system may regain its value.

But as I said before: this is one card, not a game already won.

Tang Dao-sheng’s review of Yuanbao in the article is particularly worth reading.

Over the past year, under significant user growth pressure, Yuanbao invested substantial resources in promotion and traffic acquisition, but at the time, “the model and product weren’t ready yet,” resulting in unsatisfactory outcomes.

This is almost the most classic mistake in the internet industry: pouring in traffic before the product has developed real retention capabilities.

Tencent was once one of the companies most qualified to do so, given its one of the strongest user distribution systems on China’s internet. But AI has now taught Tencent a lesson—in the era of large models, traffic does not automatically translate into product capability.

Users may come for the first time due to red packets, referral links, or social sharing—why do they return the second time?

The answer must be based on model performance, answer quality, task completion, and usage habits, not on the channel.

This also explains why, despite significant investment in Yuanbao promotion over the past year, Yuanbao’s monthly active users in June 2026 remain clearly lower than those of DouBao, Qwen, and DeepSeek. It’s not that Tencent doesn’t know how to drive growth, but rather that the growth mechanics of AI products have evolved beyond those of traditional internet products.

The model itself is part of the product.

The cost of reasoning is also part of the product.

Search data quality, answer reliability, tool invocation success rate, and response speed are all part of the product.

Therefore, the true value accumulation of Yuanbao may lie not just in its existing users, but in the evaluation system, experimental platform, search capabilities, and model-product Co-Design mechanism established by Tencent.

If these capabilities are eventually reused in WorkBuddy, CodeBuddy, and even the WeChat AI entry points, then the early tuition paid by Yuanbao won't have been in vain.

But the precondition remains that Tencent cannot use its distribution capabilities to mask product issues.

One of the biggest pitfalls of the AI era is that a company with a billion users may easily assume that all billion users are also its AI users.

Between them, there is still a gap in product capabilities.

Compared to Yuanbao, I’m more interested in CodeBuddy and WorkBuddy.

Because their development process is very unlike a traditional Tencent strategic project.

Tang DaoSheng himself admitted that these two products were not the result of a strategic plan laid out years ago by management in anticipation of the Agent wave.

Tencent Cloud initially focused on developer tools and even acquired Coding.net, gradually building capabilities in code hosting, CI/CD, web IDEs, and sandboxes. However, the developer tools business model has not been profitable, with related operations suffering long-term losses. During Tencent Cloud’s cost-reduction and efficiency-enhancement phase, the business once faced potential budget cuts.

Fortunately, it wasn't cut.

In the era of large models, these previously considered "unprofitable" foundational capabilities have suddenly begun to connect. In 2024, Tencent launched CodeBuddy, after which the model's coding capabilities continued to improve; in early 2026, the Agent boom erupted, and the team leveraged CodeBuddy’s underlying capabilities to launch WorkBuddy, designed for broader office scenarios.

The real highlight isn't "Tencent has launched another AI product," but that this team is beginning to use AI to transform their own development approach.

Tang Daosheng revealed that after WorkBuddy launched, it underwent more than 40 iterations within three months. The team no longer strictly follows the traditional workflow of requirement documentation, review, scheduling, development, and testing; instead, they first use AI to quickly generate runnable prototypes, which are then evaluated, debugged, and refined by humans. Much of the code is generated directly by AI, allowing humans to focus more on decision-making and oversight.

The importance of this matter exceeds even that of WorkBuddy itself.

In the past, when all companies discussed AI-driven efficiency gains, the easiest point to make was that "employee efficiency will increase by 30%." The truly difficult question, however, is: If efficiency truly improves, what should the organization look like?

When product managers can create prototypes directly, and developers can have agents handle large amounts of coding and testing, the previously clear boundaries between testing, product, and development will naturally be challenged.

A function team that previously required a dozen people may only need a few in the future.

This is not simply a “layoff logic,” but a change in organizational granularity.

In the past, the advantage of large companies was their ability to coordinate complex collaboration among hundreds or thousands of people; after the emergence of AI, small teams can now achieve the same execution capabilities that were once only possible for large organizations, by leveraging models and agents.

Conversely, this is also a challenge for large companies like Tencent with massive organizations.

If AI truly enhances individual leverage, large teams must also become leaner.

Market results indicate that WorkBuddy has already generated some early signals. According to data released by Analysys, in Q2 2026, the total monthly visits to 17 leading native AI desktop office agents reached over 60 million, with WorkBuddy accounting for approximately 20.97 million visits alone—ranking first.

Of course, this achievement does not yet prove that the business model is viable, but it does at least show one thing: Tencent is not falling behind on all its AI fronts.

In the category of Agent and productivity, it already has an opportunity to secure a relatively high position.

The first question is whether Hy4 can truly narrow the model gap.

As of August 26, Tencent has officially released and made Hy3 available globally, while Hy4 is still in the upcoming stage. Tang Daosheng said Hy4 will bring more breakthroughs, which is promising, but before its release, it remains an expectation, not a result.

What truly matters is not how many benchmarks outperformed whom at the launch, but three key metrics: completion rate in real Agent tasks, inference cost under equivalent performance, and whether it can directly improve user retention and task success rates after being integrated into Yuanbao, WorkBuddy, and CodeBuddy.

Only when model improvements translate into product improvements do model advancements truly belong to Tencent.

The second question is whether the massive increase in computing power this year can be turned into revenue.

This year's first half, Tencent's capital expenditures have already exceeded last year's full-year total, meaning that "insufficient computing power" is being addressed with real money. But capital expenditures are just the beginning—data centers and GPUs ultimately require utilization.

If WorkBuddy, CodeBuddy, cloud model invocation, and WeChat AI services continue to grow, the computing power purchased today will become productive assets for future revenue; if product growth cannot be sustained, substantial computing power purchases may instead turn into depreciation pressure.

Therefore, Tencent's management emphasized in the Q2 earnings report that they aim to convert model and application usage into revenue—a statement that is more important than any model ranking.

Tencent AI must ultimately return from a technology race to an economic model.

The third question, and the most critical one: Can Tencent truly turn its "scenario advantages" into an Agent network?

If WorkBuddy can someday access communication data from WeCom, business documents from Tencent Docs, meeting records from Tencent Meeting, and internal CRM and knowledge bases, and leverage different service Agents to handle permission controls and task execution, then Tencent will have more than just an office chatbot.

It could become a new layer of operational interface for enterprise workflows.

This is the most promising direction for Tencent.

Not building a Chinese version of ChatGPT, but making AI a smart operating system that connects Tencent’s existing services.

But this path is also the hardest.

Because it requires true integration across different business units and products within Tencent, it necessitates resolving issues of permissions, data boundaries, and利益分配, while also maintaining consistent product experiences over time. The most common problem in large companies in the past has not been lack of resources, but rather too many resources and overly thick organizational boundaries.

So, the ultimate question AI poses to Tencent may not be technological.

but rather an organization.

Tang DaoSheng said that AI may have only completed the first kilometer of a marathon.

I basically agree.

The large model architecture is still evolving, agents have just left the lab, enterprise business models are still being explored, and inference costs are rapidly declining—there is a high probability that new product forms will emerge over the next three years. Companies leading today have not secured the final ticket, and those falling behind have not been eliminated.

Tencent still has time.

Moreover, it has strong cash flow, a social network of 1.4 billion users, a large enterprise customer base, mature cloud infrastructure, and one of the most complex application scenarios in China’s internet industry. When it truly comes to the stage of sustained spending, Tencent may even have greater endurance than many companies that started earlier.

But long-termism never means ignoring today's gaps.

True long-termism requires acknowledging mistakes, reallocating resources, and being willing to pay the cost for previous delays.

From this perspective, the most valuable aspect of Tang Daosheng’s article today is not an explanation for why Tencent moved slowly in the past, but rather that Tencent is finally spending less time justifying “why there’s no problem” and beginning instead to directly acknowledge “where the problems lie.”

If you don’t have enough hashing power, buy more.

If the model lags, rebuild it.

If the promotion of Yuanbao is too early, refine the product again.

A developer tool that was nearly scrapped caught the agent wave and quickly doubled down.

These actions, of course, do not yet prove that Tencent has fully turned things around, but they at least signify that the Tencent, which appeared somewhat hesitant on AI issues over the past few years, is disappearing.

Tencent has now entered a more brutal yet more interesting phase: it has decided to chase.

There isn't much more grand narrative to tell.

Hy4 must speak through results, Yuanbao through retention, WorkBuddy through payments and productivity, and hundreds of billions in computing power investment must speak through revenue and profit.

AI is a marathon, and this statement is correct.

The true cruelty of a marathon has never been who runs out first.

But can you, after realizing you’ve fallen behind, use enough time, enough money, and enough determined self-change to gradually close that gap?

Tencent is now truly beginning to answer this question.

1. Tang Daosheng: "In the Second Half of AI, Endurance Matters More Than Getting Started Early," Tencent Internal Publication, August 26, 2026; Phoenix News reprinted "Tencent's Tang Daosheng: In the Second Half of AI, Endurance Matters More Than Getting Started Early!" (Phoenix Tech)
2. Tencent Holdings Limited,Tencent Announces 2026 Second Quarter Results, August 12, 2026.PR Newswire
3. Reuters, Tencent Q2 revenue climbs 11% on AI-driven ad gains, but profit falls short, August 12, 2026. (Reuters)
4. Reuters, Tencent pledges higher AI investment in 2026 after chip curbs hit capex plans, March 18, 2026. (Reuters)
5. QuestMobile: "Semi-Annual Report on the AI Application Market Development in 2026," July 14, 2026. (QuestMobile)
6. Analysys: "Insights into the Chinese Office Agent Platform Market, Q2 2026," July 2026. (Sohu)
7. Tencent: "Tencent Hunyuan Hy3 Opens Globally, Integrating Practical AI Capabilities Across Products, Workflows, and Cloud Services," August 5, 2026. (Tencent)
8. Reuters, Alibaba profit falls 75% after ramping up AI infrastructure spending；Alibaba shares fall 8% after $10 billion Hong Kong share sale，August 20, 2026, August 24. (Reuters)