We recently published a report on Chinese AI apps on our Substack channel.
You can now download a free PDF version with improved layout on our homepage. Enjoy!
https://t.co/0Zs7QfCY5W
We're writing on Tencent's AI strategy this week, and this is one of a veritable flurry of data points and strategic talking points that the company has released so far this month.
@TencentGlobal says its cloud AI platform, TokenHub, is already processing 5 trillion tokens a day and could reach 50 trillion daily by 2027. That's a staggering increase in AI usage in a very short period of time.
But Tencent also says it still doesn't have enough GPUs. At the company's June 5 conference, senior EVP Dowson Tang said compute remains the main bottleneck, especially for external cloud customers. Tencent is prioritizing its own products first, including Hunyuan model training, WeChat, Tencent Meeting, and Yuanbao.
That helps explain Tencent's AI strategy. Rather than focusing on selling API access, the company is embedding AI into products people already use. It now offers more than 20 AI agents covering coding, office work, knowledge management, design, and other tasks. CodeBuddy is already used by 95% of Tencent engineers, while WorkBuddy has become China's most-used desktop AI agent by daily active users. Tang said monetization is not the immediate goal; adoption and usage come first, with pricing mainly used to manage limited compute resources.
New chief AI scientist Yao Shunyu made a similar point in his first major public appearance since joining Tencent. He argued that the next challenge in AI is not finding a better model architecture but finding real-world problems to solve. That's one reason he joined Tencent: the company has a large ecosystem of products and users that allows AI teams to build directly against actual user needs rather than benchmark tests.
Viewed through that lens, the 5 trillion daily tokens are less a cloud metric than a product metric. Most of that usage is flowing through Tencent-owned surfaces such as developer tools, workplace software, Yuanbao, and eventually WeChat. Tencent appears willing to absorb the infrastructure costs today in order to drive adoption across its ecosystem, betting that AI becomes more valuable when it is deeply integrated into products than when it is sold as a standalone model.
#Tencent
A few prior pieces we wrote that covered this exact topic:
- The Taobao Inside Qwen: Why Alibaba's AI Gambit Is About Re-Architecting the Internet (Apr 2026): https://t.co/M6KLoe8SPQ
- Tencent’s e-Commerce Revival. Part 3: Challenges & Outlook (Sep 2025): https://t.co/F2CEJ7fxhl
The Tencent AI news keep on coming!
https://t.co/SqsrDxUkKo and Tencent are connecting their AI Agent systems, with JD contributing supply-chain and fulfillment infrastructure and Tencent providing traffic and distribution through WeChat and partnerships with five major phone makers: Huawei, Xiaomi, OPPO, vivo, and Honor.
The mechanics: phone-level AI assistants can now invoke WeChat functions and JD's fulfillment API without users opening standalone apps. A shopping request made through a phone's native agent can route through JD's product database and land in its logistics network, creating a closed loop from intent to delivery. WeChat's 1.3 billion monthly active users sit at the center of this system, and Tencent is separately developing a WeChat-embedded AI assistant that would layer into social scenarios directly.
The architecture mirrors what Alibaba has been building with Qwen and Taobao, which we've covered before. Both pursue the same strategic prize: capturing user intent at the earliest touchpoint, before users navigate to individual shopping apps. But the routes differ: Alibaba owns the full stack, from model to marketplace to payment to logistics. The JD-Tencent version splits the stack, with Tencent controlling the entry points and JD providing fulfillment as a service.
For JD specifically, the bet is that in a mature e-commerce market where price competition has diminishing returns, turning its supply chain into an API that plugs into any AI agent surface could reduce per-transaction friction costs more effectively than competing for app opens. JD has already integrated its agent with Huawei, OPPO, and Honor devices.
The unresolved question is the same one that hung over Alibaba's Spring Festival push: whether agent-driven commerce creates lasting user habits or whether the behavior fades once the promotional incentives disappear. At least for now, the competition is less about which AI model performs better and more about which combination of entry-point control and fulfillment infrastructure proves stickiest.
For background on China's AI model and deployment landscape, these are the earlier TBC pieces we would put next to the new item:
- China's Overlooked AI Model Makers: Xiaomi, Meituan, and StepFun (Jun 2026): https://t.co/jaamzaaPrK
- Forget the Leaderboard: Mapping the Ten Business Systems Behind China's AI (May 2026): https://t.co/QUpvuQipZ1
Moonshot AI (Kimi) is in talks to raise as much as $2 billion at a pre-money valuation of $30 billion, according to Bloomberg, barely a month after closing a previous $2 billion round at $20 billion, making it the third financing round in under six months.
Total disclosed funding for the Beijing-based company now exceeds RMB 37.6 billion (~$5.2 billion), per Huafeng Capital. Annual recurring revenue passed $200 million in April, split between paid subscriptions and API usage, but that still leaves a wide gap between current revenue and a $30 billion valuation.
The speed suggests a deliberate pre-IPO capital grab, with OpenAI raising $122 billion this year at an $852 billion valuation and Anthropic closing a $65 billion round at $965 billion just last week. Chinese labs appear intent on matching that firepower before the IPO window and any consolidation wave arrive.
Moonshot’s product footing gives the raise a sharper edge. The new Kimi Work agent, now in beta, moves toward local desktop task execution. Moonshot released the K2.6 model in April, and it benchmarks competitively against GPT-5.4 and Claude Opus 4.6. But in a market where ByteDance and others bundle assistants into larger platforms, converting user attention into durable paid revenue is the operational question.
Raising something close to $4 billion within two months would give Moonshot a balance sheet that can sustain both model R&D and agent distribution even if the consumer pricing war intensifies. Moonshot flagged IPO preparations in March, suggesting the company is positioning for a path where public-market discipline meets the heavy spending the category still demands.
#ChinaAI
Alibaba is not retreating from instant retail. It is spending half as much overall but pouring capital into physical stores and front-end warehouses, a strategic pivot from delivery subsidies to retail infrastructure.
Taobao Flash Purchase, Alibaba’s instant retail unit, set FY2027 goals: keep external delivery market share stable and reach monthly unit economics breakeven. Daily orders hover around 60 million (including Tmall 4-hour delivery and Hema), down 30-40% from last summer’s peak but stable. The loss per order is about 1.5 yuan (~$0.21).
What changed is where the money goes. Alibaba slashed total spending on this unit by half versus the prior year. But it is tripling its Taobao Convenience Store target from 1,000 to 3,000 stores and adding Hema front-end warehouses. The logic: shorten fulfillment distance (“远转近”), reduce last-mile costs, and build supply chain assets that are harder to replicate than a delivery platform.
This is the most concrete operational data we have seen from Alibaba’s instant retail strategy since the sector became a profitability trap. The question is whether physical store expansion adds another layer of cost without breaking the unit economics, or whether Alibaba’s existing retail assets (Hema, Tmall Supermarket) finally give it a structural advantage over pure delivery players.
We have tracked this battle for years; the Thirty-Minute Trap is still unbroken, but Alibaba is trying a different way out.
A few prior pieces that frame this pretty directly: Together, they give the background for why we read this as part of an existing pattern.
- Forget the Leaderboard: Mapping the Ten Business Systems Behind China's AI (May 2026): https://t.co/QUpvuQipZ1
A researcher profile Career path, degrees, notable work, quoted remarks with sources, and media, all linked back into the talent graphs.
- China's Overlooked AI Model Makers: Xiaomi, Meituan, and StepFun (Jun 2026): https://t.co/jaamzaaPrK
Tencent's Hunyuan team just published a new sparse attention algorithm called Stem that was accepted to ICML 2026. The headline claim is impressive: at a 128K-token context window, Stem reduces first-token latency by 3.6x while using only about 25% of the compute required by traditional dense attention.
The key idea is simple. Instead of treating every token in a long conversation as equally important, Stem identifies which tokens are most likely to influence the model's answer and spends its compute budget there. Two techniques, Token Position Decay (TPD) and an Output-Aware Metric (OAM), help the model ignore information that contributes little to the final response. Tencent also built a companion HPC operator library to ensure those theoretical gains translate into real-world speedups.
Why does this matter? Because Tencent isn't building AI for benchmark leaderboards. It's building AI for products used by hundreds of millions of people.
Tencent has reportedly been testing a WeChat-native AI agent. In that environment, long-context performance determines how much chat history, personal information, and prior interactions the assistant can remember. Latency determines whether responses feel instant or sluggish. And inference costs determine whether the service can be offered broadly or only sparingly.
A 3.6x reduction in latency at 128K context makes it significantly cheaper to run AI assistants that can remember more and respond faster. That matters not just for WeChat, but also for Tencent's gaming, enterprise software, and collaboration products.
Tencent is open-sourcing both the algorithm and the operator library. But the bigger story is that inference efficiency is increasingly becoming a competitive advantage. If AI is going to be embedded across Tencent's ecosystem, lowering the cost of every query may ultimately matter as much as making the model itself smarter.
A few prior pieces that frame this pretty directly: We are suggesting these because they give the prior context for the mechanism in the first post.
- The Thirty-Minute Trap, Mutated: A Retrospective on China’s Instant Retail Reckoning (May 2026): https://t.co/7tXALU65L5
If DS continues to open source their most core innovations, anyone else can choose to follow in their footsteps to heavily optimize for Huawei chips. The bigger question is how long it will take for that supply to come online reliably at the expected specs (and cost!) for the whole ecosystem to use ...
Five days after DeepSeek permanently cut V4-Pro API prices by 75%, Xiaomi slashed MiMo-V2.5 pricing by up to 99%, bringing it almost exactly in line with DeepSeek.
At first glance, it looks like another move in China's AI price war. But Xiaomi's disclosures suggest something more interesting.
On May 30, the MiMo team led by Luo Fuli, formerly of DeepSeek, published a paper detailing a series of optimizations around Hybrid SWA, MoE, and KV cache management. The reported result is a system that achieves 93-95% cache hit rates, increases effective KV cache capacity nearly 5x, and reduces the prefill cost of non-cache-hit tokens dramatically. Luo noted that even after the price cuts, the service is operating near full capacity and roughly breaks even.
For a standalone model company, break-even at these prices would not be particularly encouraging.
As we noted in our February analysis of Zhipu and MiniMax, China's foundation model sector increasingly faces a business challenge rather than a technical one. Usage continues to grow, but cloud API revenue remains difficult to scale relative to compute costs. Market pressure may therefore accelerate the sector's maturation from research-led competition toward business model discipline.
Xiaomi is operating under a different set of incentives. It does not need MiMo to be a major profit center. If lower API pricing drives developer adoption across HyperOS, EVs, smartphones, and smart home devices, Xiaomi can capture value elsewhere in the ecosystem.
That creates an uncomfortable dynamic for pure-play model providers. Xiaomi can treat inference as customer acquisition. They cannot.
The question is not whether API prices will fall. They probably will. The more important question is who still has business leverage once they do.
A lot of the initial discussion around Xiaohongshu's World Cup deal has focused on demographics: Xiaohongshu wants more male users, while Douyin no longer needs to pay for traffic.
There's probably some truth to that, but it doesn't fully explain why Xiaohongshu was willing to buy the rights.
Douyin reportedly paid more than RMB 1 billion ($138 million) for the 2022 World Cup. Four years later, it can probably afford to sit out. Whether or not it owns the matches, users will still go there for highlights, commentary, memes, and fan reactions. It already captures much of the attention around major events.
Xiaohongshu is in a different position. The company has been trying to expand beyond its core identity as a lifestyle and consumption platform, and the World Cup gives it a rare piece of content that can pull in entirely new audiences.
What makes the deal interesting is that Xiaohongshu doesn't need users to watch the match on Xiaohongshu forever. It needs the tournament to generate content that fits naturally into its ecosystem.
A 3 a.m. match can become restaurant recommendations, watch-party guides, jersey styling posts, travel itineraries, snack orders, and fan stories by the next morning. That's much closer to Xiaohongshu's core product than live sports broadcasting.
The company has said it wants to grow daily active users from roughly 100 million toward 300 million. The World Cup may help with that, but the bigger opportunity could be expanding the categories of content, search, and advertising inventory that live on the platform long after the tournament ends.
The question is whether that happens. Sports audiences are notoriously event-driven. If the content created around the World Cup becomes part of Xiaohongshu's long-term search and recommendation ecosystem, the rights purchase may look smart. If not, it risks becoming a very expensive marketing campaign.
A few earlier pieces help explain why we read it this way: They are useful because they show the earlier TBC framing behind this post, not just the latest news peg.
- The State of Chinese AI Apps 2025 (Oct 2025): https://t.co/YBrGDkXWmh
- The Taobao Inside Qwen: Why Alibaba's AI Gambit Is About Re-Architecting the Internet (Apr 2026): https://t.co/L49FwiTgFL\
Three announcements hit in quick succession on June 2, 2026. Meituan founder Wang Xing said Meituan’s AI assistant Xiaomei will integrate into Tencent’s Yuanbao, letting users order takeout and complete delivery entirely inside Yuanbao’s chat interface. The Financial Times reported that Tencent is testing an autonomous AI agent prototype that could call WeChat mini-programs directly, with a compliance review possible this month. Separately, WeChat opened interfaces to Huawei, Honor, Xiaomi, OPPO, and vivo voice assistants, enabling them to initiate WeChat audio/video calls and send messages on behalf of users. Tencent’s Hong Kong shares jumped more than 10%, adding about 415.8 billion HKD (~$53.3B) of market value in a single session, the biggest one-day gain since January 2021.
Taken together, the moves mark a sharp departure from the cautious AI posture Tencent maintained through much of 2025. Rather than chasing ByteDance’s Doubao (340 million MAU) or Alibaba’s Tongyi Qianwen in the race for standalone AI assistant market share, Tencent is embedding WeChat, with its 1.4 billion users, as the connective execution layer for other services. The Xiaomei-Yuanbao link is agent-to-agent: Meituan provides real-time local service data and fulfillment; Yuanbao provides the conversational entry point. The mini-program agent would let WeChat orchestrate transactions across sectors without requiring users to open separate apps. The OEM deals turn WeChat into a service pool that external AI assistants can call, rather than a walled garden they need to bypass. In each case, Tencent’s value is not the smartest model or the biggest chat app, but the ability to complete actions—payments, ride-hailing, messaging—inside the ecosystem Chinese users already inhabit.
When we wrote in October 2025 that China’s AI app economy was entering a scale phase where distribution and bundling would favor large platforms, Tencent looked like the giant still deciding what to do. The current sequence suggests it has found its answer. Alibaba used Qwen to capture user intent inside its own commerce loop. Tencent is moving in the opposite direction: opening WeChat to external agents and positioning the super-app as shared infrastructure, not as an AI destination. The question, as always, is whether the investment will be tolerated before the revenue appears. Tencent’s Q1 capex reached 31.9 billion yuan (~$4.4B), roughly 42% of non-IFRS operating profit, and the company has said the WeChat agent launch depends on regulatory approval, a process that a 1.4-billion-user platform cannot rush. The bet, essentially, is that controlling the execution layer in an AI-first world is worth the price of building it.
Gu Quanquan's departure from ByteDance Seed is attracting attention because he led two very different parts of the organization: LLM pre-training and AI4S research, including protein folding and drug design.
Some observers are interpreting this as a signal that ByteDance is becoming more commercially focused and less interested in long-horizon scientific research. We are not convinced that follows.
ByteDance is certainly under pressure to show returns on its AI investments. Doubao is beginning to roll out paid tiers, and the company is reportedly planning to spend more than RMB 200 billion ($27.6 billion) on AI infrastructure this year. But that alone does not necessarily imply a reduced commitment to AI for science.
If anything, many leading AI labs globally appear to be moving in the opposite direction. Biology and AI for science have become increasingly important research areas for companies such as Google DeepMind and OpenAI, and there is substantial investor and industry interest in the space.
What seems clearer is that AI4S operates on a very different timeline from consumer AI products. A breakthrough in protein design or molecular modeling may ultimately prove extremely valuable, but the path to commercialization is generally longer and less direct than improvements to a chatbot, coding model, or agent product.
ByteDance has reportedly chosen to reorganize the AI4S team rather than spin it out or shut it down. That suggests the company still sees strategic value in the work.
For now, we would be cautious about reading a single departure as evidence of a broader retreat from AI for science. The more interesting question is how large AI labs balance increasingly product-driven priorities with research programs whose payoff may take years to materialize.
One thing we've been thinking about recently is whether China's AI story ultimately ends up being less about model companies and more about distribution systems. Much of the discussion still focuses on foundation models and leaderboard performance, but some of the most interesting companies don't fit neatly into that narrative.
Xiaomi, Meituan, and StepFun are good examples. Xiaomi is building models, but its real objective is strengthening the ecosystem around phones, cars, and connected devices. Meituan is building models, but the value it is trying to create sits inside transactions, merchants, logistics, and local commerce. StepFun increasingly looks less like a standalone lab and more like a company using partners and terminals as its route to users.
The interesting question is not whether these companies can compete directly with the leading model labs. It is whether AI ultimately becomes more valuable when embedded into an existing economic system than when sold as a standalone product. China's strengths in hardware, consumer platforms, local services, and distribution may make that path particularly important. Our sense is that many people outside China are still underestimating how significant this possibility could become.
That's what we explore in this week's piece. https://t.co/jaamzabnhi
If you want to follow the thread back, these are the pieces we would suggest you pull up:
- Hands, Not Bodies (May 2026): https://t.co/tVdn9lu3nO
- On the Ground in China’s Humanoid Robotics Moment (May 2026): https://t.co/O9mOKwlDIs
- Unitree Can Build the Body, Can It Build the Mind? (Mar 2025): https://t.co/BcVLFKpmme
BrainCo, the Hangzhou-based brain-computer interface company originally known for prosthetic hands, expects its robotic hand sales to surge this year. It now counts dozens of Chinese humanoid robot makers as customers, including Unitree Robotics and Leju Robot.
The shift offers a useful real-time check on where China's humanoid industry is actually spending money. After two years of demos that emphasized walking and backflips, developers are now hunting for real application scenarios, and the hand layer is where the hard engineering problems still sit: precise grasping, object manipulation, and fine-motor tasks remain unsolved.
BrainCo's background adds a layer the generalist robot companies cannot easily replicate. Its original business decoded neural signals to control prosthetic fingers for amputees, building data and algorithms around how human intent translates into dexterous motion. That expertise now feeds directly into the robotic hand market, where control precision and energy efficiency matter more than raw actuation speed.
The customer list is notable. Unitree, best known for quadrupeds and low-cost humanoid bodies, sources hands rather than building them in-house. Leju Robot, which discounted its way into the top tier of China's humanoid rankings, does the same. Both appear to be betting that component specialization, especially at the hand, will yield faster progress than vertical integration.
This fits the pattern observed earlier with LinkerBot, the world's largest high-DOF dexterous hand supplier, which shipped over 10,000 units in 2025 and aims for 10,000 per month by end-2026. The hand is shaping up as a separate category with its own economics, customer relationships, and scaling dynamics, not merely a subsystem absorbed into the robot OS. If BrainCo can build similar volumes from its prosthetic-to-robot pivot, the component supply chain may prove more durable than the body brands themselves.