Tencent Reveals AI Compute Can Be Resold at 30% Profit as CapEx Triples to RMB 51.8 Billion
Q2 2026 earnings call, August 12, 2026
Tencent's second-quarter results confirmed what investors have suspected for months: the company is now running two businesses under one roof, a highly profitable legacy franchise and a capital-intensive AI build-out whose returns remain unproven but whose downside, management insists, is protected by an unusual insurance policy. Total revenue rose 11% year-on-year to RMB 205 billion, with gross profit up 13% to RMB 118 billion. But the more consequential disclosure came on capital expenditure, which surged 190% year-on-year and 66% quarter-on-quarter to RMB 51.8 billion, pushing free cash flow negative RMB 13.8 billion for the quarter, or a still-positive RMB 37.6 billion excluding AI-related prepayments.
The Compute Arbitrage Nobody Expected
The single most striking data point from the call was Chief Strategy Officer James Mitchell and President Martin Lau's disclosure that Tencent could already resell compute it committed to months ago at a markup. "For the prepayment and for some of the compute orders that we have made just a couple of months ago, today, we can actually sell it at more than 30% profit compared to the price that we paid just a few months ago," Lau said. This is a materially new data point for investors trying to underwrite the risk in Tencent's AI spending: it suggests GPU scarcity pricing in China has become severe enough that Tencent's CapEx book is, in effect, an appreciating asset rather than a sunk cost. Mitchell reinforced the point on depreciation risk, noting the company "could recover the depreciation almost immediately by renting the compute out to third parties as many neo-cloud businesses are doing." Management's stated strategy, however, is to prioritize model training and WorkBuddy inference over immediate GPU rental, betting that proprietary intelligence converts into better long-term economics than acting as a bare-metal reseller. Investors should note this is a hedge, not a plan. Tencent is explicit that the fallback option exists precisely because it isn't confident WorkBuddy and Xiaowei will monetize on the timeline needed to justify the spend outright.
WorkBuddy Emerging as the Real Swing Factor
Of the new AI product spending, which rose from roughly RMB 8.8 billion in the first quarter to RMB 10.5 billion in the second, management confirmed a significant internal reallocation toward WorkBuddy, the company's agentic productivity platform, at the expense of other AI initiatives. Mitchell was candid about the mechanics: "there was actually a very big change because we identified that WorkBuddy was breaking out. And therefore, we aggressively prioritized WorkBuddy while deprioritizing some of the other products in that new AI product portfolio." WorkBuddy now ranks first among AI productivity services in China by monthly interactions, and Lau described it not as enterprise software sitting alongside Tencent Meeting and Docs, but as a new platform layer: "It is indeed a new platform that it's a very flexible workspace for agentic AI... there will be many models serving the users through WorkBuddy... Hunyuan will be one of the models that will be provided by WorkBuddy, but it will not be the only model." That is a notable admission from a company that just spent two years building its own foundation model, and it signals Tencent sees itself as an orchestrator of third-party intelligence as much as a model owner. On unit economics, Mitchell added that paying-user gross margins for WorkBuddy and Model-as-a-Service already match Tencent Cloud's overall gross margin, though blended margins remain diluted by free-tier subsidization aimed at market share.
Cloud Pricing Environment Has Quietly Turned
Analysts pressed on the commoditization of token pricing in China, and Mitchell's response contained a genuinely underappreciated data point: input costs, particularly memory, have risen sharply enough that Tencent has been raising cloud prices. "We have increased prices across the board in May for Tencent Cloud. And beyond those headline price increases, we've also been more substantially reducing discounts," he said, adding that "domestic token manufacturing costs are also extremely low, much lower than widely perceived or externally estimated." This reframes the bear case on China cloud margins, which has assumed relentless price deflation. Cloud revenue growth accelerated from high-teens percentage in the first quarter to low-20s percentage in the second, and management flagged that capacity constraints, not demand, remain the binding limit. Business Services and FinTech combined grew revenue 9% to RMB 60 billion, with wealth management aggregate customer assets rising on automated investment strategies, a smaller but steady contributor investors should track alongside cloud.
Xiaowei's Agent-to-Agent Ambition, and Its Guardrails
Tencent disclosed the prototype rollout of Xiaowei, an embedded agentic assistant inside Weixin built on a proprietary small model called WeLM, designed explicitly for privacy and cost efficiency rather than raw capability. Lau's framing of the long-term vision was the most expansive language on the call: "In the future, if they can actually send a complex instruction to an agent and the agent can actually sort of start helping the user to execute transactions... longer term, there will be even user, each user has got an agent, and they can actually interact with each other to execute a transaction." Notably, management is deliberately throttling the product today, keeping Xiaowei configured to require user confirmation at each step "as safety measures," even though the underlying model can technically handle more autonomous workflows. Lau also addressed a bear concern directly, that agent-mediated transactions could cannibalize higher-margin ad inventory, dismissing it by comparing the AI transition to the QQ-to-Weixin migration, which he said "magnified QQ's value by more than 10x." Investors should treat that comparison as a management thesis rather than a proven outcome; Xiaowei has not yet scaled beyond prototype. On investment intensity, Lau offered a useful anchor: Xiaowei's cost envelope will run below what Tencent spent annually on Yuanbao, its existing consumer chatbot, suggesting this is a controlled bet rather than a major new drag line.
Hunyuan 3's Traction Numbers, and the Case for Hunyuan 4
The production version of Hunyuan 3 generated a roughly 6x increase in average daily token usage compared to the preview version and now ranks among the top three models globally on OpenRouter by token usage, a genuinely strong external validation point for a model management describes as intentionally small and cost-optimized rather than frontier-scale. Lau was unusually direct about the strategic logic: "Hunyuan 3 is a very small model even in today's terms, but it's actually very widely used... it's actually focused on use cases rather than just benchmark beating." The larger Hunyuan 4 model is due later this year, and management's justification for the added training spend centers on unlocking a standalone token-selling business and enabling WorkBuddy to handle higher-value, more complex agentic tasks, rather than any specific near-term Weixin use case, since Lau confirmed Xiaowei's WeLM model does not depend on Hunyuan reaching state-of-the-art status at all.
Margins, Cash Flow and the Capital Allocation Signal
Excluding new AI products, non-IFRS operating profit grew a healthier 19% year-on-year to RMB 86.1 billion, with margin excluding AI products up 2.8 points to 42%, underscoring that the core business, games, advertising, fintech, continues to generate substantial operating leverage even as the AI segment drags headline profitability down to 9% growth. R&D expenses jumped 35% year-on-year to RMB 27.2 billion. Net cash fell to RMB 58.2 billion from RMB 146.9 billion in the prior quarter, reflecting both the CapEx surge and RMB 41.6 billion in dividend payments. On buybacks, which started accelerating in May even as CapEx ramped, Mitchell was candid that capital allocation will flex dynamically: "if we identify that there's superior returns from capital expenditure... then we'll steer more cash toward the capital expenditures than we had in the past and therefore, potentially less cash towards buybacks." Lau added a useful clarification for modeling purposes, that the current AI infrastructure spend is closer to a one-time lump sum for this year and next rather than a permanently elevated run rate, since model-building costs are largely fixed rather than compounding annually. Investors should treat that as a soft commitment rather than firm guidance, given management repeatedly declined to provide specific timing on when AI investment turns earnings-accretive.