Tencent has publicly acknowledged what many in the industry have suspected for over a year: its AI efforts have been hampered by severely insufficient computing power, stalled model training, and product rollout missteps. In a lengthy internal article published August 26, Tang Daosheng, Senior Executive Vice President and CEO of Tencent's Cloud and Smart Industries Group, admitted that the company's overall computing power is "severely insufficient," affecting not only the training of its HunYuan model but also slowing product development with "a significant impact."
This level of candor is unusual for Tencent, a company known for its cautious public posture. The admission signals that the pressure has become substantial - and that Tencent is now ready to pay to catch up.
The cost of catching up
Tencent's financial statements confirm the shift. In 2025, the company's annual capital expenditure was approximately RMB 79 billion. In the first half of 2026 alone, Tencent spent RMB 84.7 billion - exceeding its entire 2025 total. Q2 capital expenditure reached RMB 52.8 billion, a 176% year-over-year increase.
This spending is already visible on the income statement. Tencent disclosed that its Non-IFRS operating profit for Q2 was RMB 75.6 billion. Excluding revenue, costs, and expenses related to new AI products like Hy, Yuanbao, CodeBuddy, WorkBuddy, and Xiaowei, operating profit would be approximately RMB 86.1 billion - a difference of about RMB 10.5 billion. AI remains in a clear phase of strategic investment, and that investment is now genuinely reflected in the company's finances.
Tang's message to the market is not "we aren't actually slow." It's essentially: we slowed down, and now we're paying to catch up.
The consumer gap is real, but the enterprise story differs
On the consumer side, the gap is visible. According to QuestMobile data for June 2026, DouBao had approximately 382 million monthly active users, Qwen had 167 million, DeepSeek had 130 million, and Tencent's Yuanbao had roughly 49.84 million, ranking fourth. For a company with WeChat's 1.439 billion combined monthly active accounts, that position is hardly leading.
Tang acknowledged that Yuanbao invested substantial resources in promotion and traffic acquisition over the past year, but "the model and product weren't ready yet," resulting in unsatisfactory outcomes. It's a classic internet industry mistake: pouring in traffic before the product has developed real retention capabilities. In the era of large models, traffic does not automatically translate into product capability.
The more promising story is on the enterprise side. Tencent's WorkBuddy, an AI office agent, reached approximately 20.97 million monthly visits in Q2 2026, ranking first among 17 leading native AI desktop office agents tracked by Analysys. CodeBuddy, the developer tool, emerged from a business that was nearly cut during a cost-reduction phase before the Agent wave gave it new relevance.
Tang revealed that WorkBuddy has undergone more than 40 iterations within three months of launch. The team uses AI to quickly generate runnable prototypes, then evaluates, debugs, and secures them - with much of the code directly generated by AI. This represents a fundamental shift in how the organization operates, not just a new product launch.
Three questions will determine the outcome
Whether Tencent's catch-up succeeds depends on three factors. First, can the upcoming Hy4 model genuinely narrow the gap with competitors? Tang said Hy4 will bring more breakthroughs, but until release, it remains an expectation, not a result. What matters is not benchmark scores at launch, but completion rates in real Agent tasks and whether model improvements translate into product improvements.
Second, can this year's massive computing power investments be converted into revenue? Capital expenditure is only the beginning - data centers and GPUs require good utilization. Tencent's management emphasized in the Q2 earnings report that they aim to convert model and application usage into revenue, a statement more important than any model ranking.
Third, can Tencent turn its scenario advantages into a genuine Agent network? If WorkBuddy can access communication data, Tencent Docs, Tencent Meeting, and knowledge bases - with different Agents handling permission controls and task execution - Tencent could become a new operating layer for enterprise workflows. This is the most promising direction, but it requires coordination across different business units with historically separate interests.
Why this matters for product development
For product teams, Tencent's experience offers a concrete lesson: the growth mechanics of AI products have evolved beyond traditional internet tools. The model itself is part of the product. Reasoning costs, search data quality, answer reliability, tool invocation success rates, and response speed all determine retention. Red packets, referrals, and social sharing may drive first-time usage, but they don't create habits.
The organizational shift at Tencent is equally instructive. When product managers can create prototypes directly and developers handle larger amounts of testing and coding, the boundaries between testing, product, and development roles naturally blur. A function team that previously required a dozen people may only need a few. For teams building AI for Product Development, the question is no longer whether AI will change workflows, but how quickly your organization can adapt to smaller, AI-augmented teams.
Tencent's trajectory also highlights the strategic importance of AI Agents & Automation in enterprise software. The products that nearly got cut - developer tools, code hosting, CI/CD - became strategically valuable once the Agent wave arrived. For product teams, this suggests that foundational capabilities built today, even those without immediate profitability, may become critical as AI evolves from answering questions to completing tasks.
Tang's core judgment - that AI has completed only the first kilometer of a marathon - is probably correct. But as he also acknowledged, Tencent has moved past its period of hesitation and entered an expensive catching-up phase. Whether the company's spending translates into revenue, whether Hy4 delivers, and whether use cases become high-frequency Agent applications will determine if the marathon metaphor holds. Competitors haven't slowed down, and in the AI industry, running slowly always comes at a cost.
Your membership also unlocks:
