Third-quarter 2026 earnings call, November 18, 2026
WorkBuddy becomes a revenue line
Tencent's productivity agent WorkBuddy was the main new disclosure on the third-quarter call. President Martin Lau said it now has more than 10 million daily active users consuming over 100 trillion tokens a day. He called it the largest AI application in China by large language model token usage. Annualized gross revenue over the previous seven days exceeded $500 million, which Lau said is more than twice the level of Tencent Meeting, a product launched seven years ago. WorkBuddy's third-quarter revenue was more than 5 times the prior quarter's. Since the enterprise version launched in June, 60% of China's top 100 companies have become customers.
Lau said management has changed its view of the Chinese enterprise software market. "While we have historically been cautious about how quickly the enterprise software market will develop in China, we now believe that adding AI capabilities will allow productivity software spending to rapidly and substantially increase with WorkBuddy as a key driver and beneficiary," he said. He gave three reasons: AI services avoid the software piracy problem, WorkBuddy covers a broad set of enterprise use cases, and a flat entry price combined with token-based billing lets spending scale with value. He said some heavy users spend as much as RMB 40,000 a month on tokens. He contrasted this with earlier Chinese SaaS, which "could only charge a flat fee based on subscriptions."
Lau acknowledged that the product is still being funded. Tencent is offering temporary free services to certain groups, such as school teachers. Management expects to improve unit economics by raising the paying-user ratio, lifting token spend per paying user, and lowering token costs through inference optimization, model customization and compute efficiency. Asked about the falling blended price per token as cheaper flash models are added, Lau said the flash models are "fantastic customer acquisition models" for free users. Users who take on more complex work move to more capable models and spend more, so WorkBuddy "monetizes at all the different points on the demand curve." He added that WorkBuddy has become the biggest customer of almost all open-weight models in China, and that its token supply is diversified across Tencent's models and third-party models.
Xiaowei: a low-cost agent inside Weixin
Xiaowei, the personal assistant embedded in Weixin, has reached 30 million daily active users across more than 30 touch points, including Moments, Official Accounts and Video Accounts. Lau said it runs on a purpose-built vision-language model optimized for cost. Even if it were rolled out to all 1.4 billion Weixin users, he expects its quarterly cost to stay below what Yuanbao incurred earlier this year.
Lau said Xiaowei is deliberately less capable than consumer agents such as Meta's Muse. Tencent wants it available to every Weixin user rather than only a small number of paying users. He expects capability to improve several-fold with each version as compute efficiency multiplies, and said Xiaowei could match Muse-level power within 12 to 18 months while still covering all users at manageable cost. On monetization, he said Weixin's combined revenue streams run to hundreds of billions of RMB a year, and that adding 10% to 20% through advertising, transactions and eventually agent-to-agent services would make Xiaowei "hugely profitable." He expects this to become clear "in the next 6 to 12 months."
Hunyuan Hy4 and the economics of cheap inference
Tencent released the production version of Hunyuan 4 in early November. Lau said it ranked first among Chinese models on the Agent Arena leaderboard at launch, with, in his words, "zero targeted optimization." Pretraining and post-training were scaled up substantially, and vision and computer-use capabilities were added. The lightweight Hy4 flash is priced at one-twentieth of Hy4. Lau said its usage "way exceeded our expectation" and that it recorded the highest daily token volume on OpenRouter of any model launched this year. He said a good share of Hy4 usage is outside China. Multimodal development has been consolidated under Hunyuan, where Tencent says it leads in 3D generation.
Lau said the training spend has now been validated. Before Hy3 and Hy4, he said, "it's uncertain whether our training investments is generating valuable models." He described Yuanbao as less successful than expected. Spending on it was cut sharply, and WorkBuddy was the next product launched. On the timing of AI profits, he said "exactly when would that happen, no one knows," and added that this is "a generational technology revolution" in which building a business is unlike constructing a building and collecting rent.
Compute spending, prepayments and the new U.S. restrictions
CFO John Lo reported operating capex of RMB 63.6 billion, up 257% year on year. Free cash flow was negative RMB 69.1 billion, driven by large compute prepayments. Excluding those prepayments it would have been positive RMB 16.9 billion. Tencent moved from net cash of RMB 58.2 billion at the end of June to net debt of RMB 13.7 billion, partly offset by RMB 8.4 billion of net inflows from divestments in its investment portfolio.
Chief Strategy Officer James Mitchell said second- and third-quarter prepayments were "abnormally high" and that Tencent has now signed the long-term contracts it wanted. Prepayment outflows will decline, and some quarters may see reversals as previously prepaid contracts are cancelled following recent changes. He put cash capex at roughly RMB 65 billion against RMB 81 billion of accrual capex, and said the cash figure is at peak levels. It may fluctuate around that level through the rest of 2026 and 2027, then decline.
Lau gave three reasons for a better 2028: infrastructure built in 2026 and 2027 can keep training models and producing tokens, pricing in 2026 and 2027 is elevated and should fall significantly, and compute efficiency is multiplying each year.
On the tighter U.S. restrictions on advanced semiconductors, Lau called them "unfortunate" and said they "set everyone back in China." He said training will be affected, and the next Hunyuan version may take longer in the short run until the domestic supply chain comes on board. Management does not expect an inference or cloud impact. Mitchell said inference for efficient models such as Hy4 flash can now run "very effectively on a surprisingly limited number of domestic chips," and that Tencent can generate far more tokens on domestic chips than it imagined a year ago. Lau said Tencent can no longer offer high-end GPUs outside China directly and may work with partners abroad. Non-GPU cloud capacity can still be built overseas....