Tech Buzz China

Exclusive Insights into China’s Tech & Innovation Landscape. Trips, bespoke research, and an investor-focused newsletter.

[Tech Buzz China @TechBuzzChina]

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

[Tech Buzz China @TechBuzzChina]

One theme we’ve been writing about for a while is that AI monetization is increasingly shifting away from advertising and engagement-driven business models toward direct payments and global distribution.

Kuaishou’s latest results offer an interesting example of that transition.

Q1 2026 live-streaming revenue fell 13.5% year-over-year to RMB 8.49 billion ($1.17 billion), its lowest level in 16 quarters. A regulatory crackdown on low-quality content following a RMB 119 million ($16.4 million) fine in late 2025 appears to have removed some of the tipping behavior that platforms had long benefited from. At the same time, daily active users grew just 1% to 413 million, suggesting the core platform is entering a more mature phase.

Against that backdrop, Kling is becoming increasingly important. Kuaishou’s AI video-generation product generated more than RMB 650 million ($90 million) in revenue during the quarter, up over 300% year-over-year, with an annualized revenue run rate approaching $500 million. The company is reportedly evaluating a spin-off at a $20 billion valuation.

What makes Kling particularly interesting is that its growth appears largely independent of Kuaishou’s core traffic engine. Rather than monetizing engagement through advertising or tipping, Kling monetizes directly through paid usage. Its user base is also increasingly global, with the app reportedly ranking among the top downloads in more than 40 countries.

The proposed spin-off may also reflect the economics of the AI video race. ByteDance’s AI infrastructure spending is on a completely different scale, with reports suggesting around RMB 200 billion ($27.6 billion) of capex this year versus roughly RMB 26 billion ($3.6 billion) for Kuaishou. An independent Kling would have greater flexibility to raise capital and compete for talent.

Kling still represents only a small portion of Kuaishou’s overall revenue today. But the willingness to separate it from the parent company suggests management believes the market may ultimately value AI assets very differently from a mature content platform.

[Tech Buzz China @TechBuzzChina]

@Lunexalith It is interesting to observe this while DeepSeek and MiMo have been cutting prices. Each company sees a different path forward!

[Tech Buzz China @TechBuzzChina]

MiniMax’s Hong Kong-listed shares fell 14.6% on Monday, the same day the company released what may be its most impressive model yet.

The new M3 model introduces MiniMax’s sparse attention architecture (MSA) and native multimodal capabilities. According to the company, MSA reduces per-token compute to roughly 1/20th of the previous generation while supporting a 1 million-token context window. MiniMax reports more than 9x faster prefilling and 15x faster decoding versus its prior architecture.

The most interesting result may not be a benchmark score. In one demonstration, M3 autonomously optimized CUDA kernels over a 24-hour coding session, making nearly 2,000 tool calls and 147 benchmark submissions while improving Hopper FP8 utilization from 7.6% to 71.3%, a 9.4x gain. On agent-focused benchmarks such as SWE-Bench Pro and Claw-Eval, M3 also performs at or near the top of the field.

So why did the stock fall?

One possibility is that investors are increasingly separating technical progress from commercial outcomes. MiniMax disclosed losses of roughly $186 million through the first three quarters of 2025, and the company’s latest pricing continues to reflect an intensely competitive market: RMB 2.1 ($0.29) per million input tokens and RMB 8.4 ($1.16) per million output tokens before launch discounts.

This is a dynamic we’ve discussed before. Better models and more efficient architectures are necessary, but they do not automatically translate into stronger economics. As inference costs fall, prices tend to fall as well.

The market’s reaction may be less a judgment on M3 itself and more a reflection of a broader question facing model developers everywhere: how much of the value created by better models ultimately accrues to the model company?

[Tech Buzz China @TechBuzzChina]

Unitree's updated STAR Market IPO filing offers one of the clearest looks yet at the economics of China's humanoid robotics industry.

Revenue reached RMB 1.7 billion ($234 million) in 2025, up from just RMB 159 million ($22 million) in 2023. The company also turned profitable last year, reporting net income of RMB 278 million ($38 million). The filing is now at the registration stage after clearing review in 73 days, reportedly the fastest STAR Market timeline this year.

When we wrote about the emerging robotics IPO wave last year, the companies that stood out were generally the ones delivering measurable productivity gains in logistics, manufacturing, and healthcare, not humanoid demos. Unitree has often been viewed as a showcase of China's humanoid ambitions, but the commercial story was still an open question.

This filing is somewhat reassuring on that front. Unitree disclosed cumulative production of roughly 11,000 units for a single humanoid model, a 2025 gross margin of 60%, and Q1 2026 revenue growth of 68% year-over-year to RMB 423 million ($58 million). Meituan, Tencent, Alibaba, and Ant Group are all strategic investors.

The planned use of proceeds is also noteworthy. Nearly half of the RMB 4.2 billion ($580 million) raise will go toward intelligent robot model R&D, addressing the AI sophistication gap we highlighted earlier this year when the prospectus was first released. The remainder will fund robot body R&D, new product lines, and manufacturing capacity.

That allocation reflects something we've been hearing repeatedly across the industry: the challenge is increasingly not building the robot, but making it useful. Better software, stronger models, tighter hardware-software integration, and clearer applications may ultimately matter more than the next generation of hardware.

[Tech Buzz China @TechBuzzChina]

Kuaishou's Q1 2026 results highlighted an increasingly important reality: much of the AI value in the company may now sit inside Kling.

While Kuaishou reported record quarterly revenue of RMB 33.72 billion ($4.65 billion), adjusted net profit fell 26% year-over-year to RMB 3.37 billion ($465 million). Meanwhile, Kling generated more than RMB 650 million ($90 million) in quarterly revenue, with ARR approaching $500 million, up roughly 5x from a year ago. The business is reportedly seeking outside funding at a $20 billion valuation.

What's notable is that Kling's growth appears largely independent of Kuaishou's core short-video business. Around 70% of revenue reportedly comes from overseas markets, driven by professional users in advertising, film, gaming, and short dramas rather than Kuaishou's domestic consumer traffic. That makes a spin-off easier to justify: Kling gains access to capital and talent incentives without sacrificing a major distribution advantage.

The timing also reflects how competitive AI video has become. Management says inference costs for Kling 2.5 Turbo have fallen nearly 30% and margins are approaching breakeven, but this remains a scale business. ByteDance is reportedly planning up to $70 billion of AI infrastructure spending this year, while Kuaishou's 2026 capex target is around RMB 26 billion ($3.6 billion).

One thing we have argued before is that defensibility in AI increasingly comes from distribution and integration rather than model performance alone. Kling's foothold among overseas professional creators may be one such advantage. The question is whether it can maintain that position as competition intensifies and some of the engineers who helped build it are now working elsewhere.

[Tech Buzz China @TechBuzzChina]

BYD @BYDCompany sold 383,453 vehicles in May, a rise of just 0.26% year-on-year. BUT! Overseas deliveries reached 160,644 units: a new monthly record and the clearest sign that the company’s growth axis has shifted beyond its home market. Cumulative new energy vehicle sales now top 16.5 million.

Breaking down the total: the mass-market Dynasty and Ocean series contributed 330,215 units. Fang Cheng Bao posted 30,186 deliveries, a number that suggests some traction for a mid-tier premium badge. Denza recorded 16,303 deliveries, and Yangwang, the ultra-luxury line, moved 286 vehicles.

The brand stratification shows that at home, volume growth has flatlined as price competition intensifies and penetration matures. The 0.26% domestic reference figure points to a saturated mass market where BYD’s scale advantage is hard to expand further. Overseas expansion is now the primary lever, with exports driving incremental volume. The premium push remains nascent and uneven: Fang Cheng Bao is finding an early audience, but Yangwang’s single-digit daily sales highlight the difficulty of breaking into the high-end segment, even with vertical integration and technological ambition. BYD still defines itself by value-for-money scale, with the transition to higher-margin brands a work in progress.

[Tech Buzz China @TechBuzzChina]

A data point that caught our attention this week: Wang Yunhe, 34, the former head of Huawei's Pangu large model and the Noah's Ark Lab, left Huawei in late March and has already launched a new AI agent startup called 基元律动 (Jiyuan Lvdong). The company has closed a financing round at a $100 million valuation (roughly 678 million yuan (~$93.5M)).

Jiyuan Lvdong officially registered in Shanghai on April 8, 2026, with Wang as CEO and Han Kai, previously a chief researcher at Noah's Ark Lab, as CTO. The startup is hiring across AI agent development, algorithm, and security/compliance roles, with annual salaries ranging from 600,000 to 1 million yuan (~$83,000 to $138,000). Multiple informed sources confirmed the company already has a stable state-owned enterprise (SOE) client and plans to release new products in the coming months.

One way to read the move is that it reflects a broader migration of top-tier AI talent from foundational model architecture toward the enterprise application layer. Wang spent nearly nine years at Huawei building extremely large models; now he is building AI agents for large, likely regulated, organizations. The presence of SOE clients from the outset suggests the company is targeting a segment that values integration, security, and compliance as much as raw model capability.

The valuation and the willingness of "top-tier venture capital firms and major internet companies" to back this early venture hint at a specific bet: that specialized AI agent companies can capture value in enterprise deployment that platform-controlled assistants may not easily reach. If Jiyuan Lvdong can embed its agents into the workflows and systems of institutions that demand tight control and customization, it may be able to build a defensible position against larger incumbents. For now, it's one startup and one funding round, but it reflects the ongoing recomposition of China's AI talent toward concrete commercial applications, not just model scale.

[Tech Buzz China @TechBuzzChina]

Doubao will reportedly launch paid subscription tiers in late June at RMB 68 ($9), RMB 200 ($28), and RMB 500 ($70) per month.

What’s interesting is that ByteDance reportedly is not setting paid-user penetration targets for 2026. The subscription appears to be as much about managing demand and inference costs as it is about generating revenue. User growth has already slowed in part because ByteDance reduced advertising spend for the app.

More broadly, this feels like another sign that China’s AI market is increasingly a distribution and ecosystem game, not just a model game.

Doubao’s 336 million monthly active users are impressive, but they are also largely a function of ByteDance’s distribution advantage through Douyin and its broader ecosystem. Average daily usage is reportedly around 10 minutes. Charging for advanced features in a market where most competitors remain free will likely result in relatively low conversion rates. That’s not unusual: even ChatGPT converts only a small percentage of active users into paying subscribers.

The bigger story may be elsewhere.

According to IDC, ByteDance’s Volcano Engine already accounts for 49.5% of China’s public cloud LLM call volume. That’s a much more direct way to monetize AI adoption, since MaaS revenue scales with actual usage and compute consumption.

At the same time, ByteDance is increasingly connecting Doubao to commerce. Early tests reportedly show more than 3% of users clicking through from AI shopping recommendations to product pages, and the company is working toward enabling purchases directly within Doubao by leveraging the existing Douyin commerce infrastructure.

The subscription tiers will attract headlines, but they are unlikely to be the main event. The larger opportunity for ByteDance is integrating AI into businesses where it already has significant advantages in distribution, transaction volume, and infrastructure: cloud services, advertising, and e-commerce.

[Tech Buzz China @TechBuzzChina]

Alibaba's May 13 earnings for the March 2026 quarter offered a useful look at the economics of the AI transition.

Cloud revenue grew 38% year-on-year, while AI-related product revenue posted its eleventh consecutive quarter of triple-digit growth. Alibaba says its AI business is now approaching an annualized revenue run rate of RMB 40 billion.

What's equally notable is the cost side. Capital expenditures have risen to roughly RMB 120 billion, while profits have come under pressure. At least for now, Alibaba's AI business looks less like a traditional software business with expanding margins and more like an infrastructure business requiring sustained investment.

When we wrote about China's AI chip market earlier this year, one observation was that the competitive focus had shifted from peak benchmark performance to delivery continuity, ecosystem depth, and long-term control of the stack. Alibaba's recent results suggest it is pursuing exactly that strategy: developing in-house chips through Pingtouge, building a large model ecosystem around Tongyi Qianwen, and leveraging an enterprise footprint that reportedly reaches around 63% of China's A-share listed companies.

The strategic value may not be the model itself, but the infrastructure layer underneath it. Once an enterprise fine-tunes models, builds workflows, and deploys applications on Alibaba's platform, future iterations, scaling, and optimization become increasingly tied to Alibaba's cloud and AI services.

One way to interpret the revenue growth is that enterprises are moving from buying software licenses to purchasing AI capabilities and compute consumption. If that trend continues, the companies that own the underlying infrastructure may capture a disproportionate share of the value created by AI adoption.

$BABA

[Tech Buzz China @TechBuzzChina]

China's AI race is now all about serving people more efficiently, and at lower cost. Not about large models in and of themselves.

Xiaomi has given the public all the detaisl on its MiMo-V2.5 models, including a hybrid attention architecture that reduces KVCache storage to roughly one-seventh of conventional approaches, a redesigned caching system with a 93% prefix cache hit rate, and scheduling improvements that cut time-to-first-token by 30%.

This means 99% lower API prices, 2.3× faster early token generation, and a reduction in one-hour video processing time from 156 seconds to just 23 seconds.

[Tech Buzz China @TechBuzzChina]

So far, CATL’s choco-swap battery rollout looks less like a pilot and more like a serious national infrastructure buildout. There are already more than 1500 swap stations working across China. They cover more than 100 Chinese cities. There is a longterm plan to build up to 10,000 stations, working with gas stations run by Sinopec. Right now, it looks like 200 stations are added a month.

These kind of battery swaps are great for taxis and ride-hailing cars. For them, switching battery fast is more cost-effective than waiting 20 minutes for a fast charge.

The station count is impressive, but the more important number over the next 2–3 years will be active swap-compatible vehicles per station. But given China has seen infrastructure booms before where deployment has outpaced utilization, if vehicle adoption lags, thousands of stations could become an underused network...

[Tech Buzz China @TechBuzzChina]

CATL's newest ¥3 billion project turns out to be a battery lie detector. The mega Xiamen campus is being tasked with evaluating the cells, modules, battery packs and entire grid-scale storage systems developed by the battery giant, and doing so in real-ish stress-test settings.

Critical infrastructure has to be reliable!

CATL executives have come clean that some projects promising 10,000+ cycles are giving people just 3–5 years of battery power. We are entering an era in which verified results matter more than what the label says.
And CATL knows it needs to test this in-house to position as trustworthy.

[Tech Buzz China @TechBuzzChina]

On June 1, 2026 Meituan CEO Wang Xing told analysts that the company’s AI assistant Xiaomei will soon be accessible through Tencent’s Yuanbao AI agent. Users will be able to ask Yuanbao for food delivery or other local services, with Xiaomei handling the transaction directly inside the chat interface. Wang added that serving “AI Agents” (To A) is becoming as important as serving consumers (To C) and merchants (To B).

When we wrote about Tencent’s e-commerce revival in September 2025, the story centered on WeChat Stores and mini-programs weaving a transaction layer into the messaging app. The Yuanbao integration with Meituan pushes the logic further: it inserts a conversational AI agent between the user and the service catalog, making the AI a new layer of distribution for local commerce.

Meituan’s recently opened LongCat-2.0-Preview model, which surpasses one trillion parameters and was trained end-to-end on domestic compute, powers Xiaomei. The larger strategic turn is that Meituan is positioning its local-services infrastructure, merchant network, review graph, and delivery fleet, as the fulfillment backend for any AI agent, not just its own super-app. As Wang framed it, Meituan’s advantage lies in having “more complete business coverage, more comprehensive merchant data and user reviews, and a more reliable fulfillment system.”

The partnership tests whether a super-app can stay relevant as an AI-accessible service provider. If AI agents become the primary ordering interface, Meituan’s moat shifts from owning the user session to being the indispensable execution layer every agent needs, a position that competitors without its offline density would find difficult to replicate.

[Tech Buzz China @TechBuzzChina]

No surprise that @Tsinghua_Uni dominates the education paths out of the 68 core Chinese AI researchers and operations we profiled. Total affiliations outnumber the next 3 universities combined. Check out our China AI Atlas project!

[Rui Ma @ruima]

3. Speaking of Z ai and its Tsinghua connection, we actually originally started the project with a simple analysis of talent flows. Where did the top talent in Chinese AI come from?

Some expected and unexpected answers based on the 68 well-known researchers and operators that we analyzed:

- Tsinghua has 27 affiliations in the network, more than the next three schools combined. The elite Tsinghua CS track shows up repeatedly across StepFun, Tencent, Meta MSL, and related leadership paths.
- Microsoft Research Asia (which we categorized as "Western") is the top Western-lab link. But we should note that it is based in China!
- China's AI 1.0 computer vision giants Megvii and Sensetime were where the founders of StepFun and Minimax came out of.
- About 1/3 of those folks whose academic records are known to us do not have any Western academic / lab affiliation.

https://t.co/25LVyo02yc

[Rui Ma @ruima]

I meant to post this last Friday with our newsletter but hey, the @TechBuzzChina China AI Atlas is now live! It's a free and interactive "field guide" to the top labs, talent and capital building China's foundation models and it is one of quite a few data tools we are building to map our "coverage universe" in China properly (the others being robotics / physical AI, advanced manufacturing, EVs, biotech, and more). The atlas itself is at https://t.co/WKDoMOWZxo and we are thankful to @Gracemzshao of AI Proem and @CRC_8341 China Research Collective for their contributions and help! Any errors that survive are ours.

A thread on the Atlas & what the data shows in this alpha version.

1. As I mentioned before, one of the first things we did was to map the most important researchers and try to give you a flavor for their technical strengths, tech/product/life philosophies, personal journeys, so that you can get a better feel for how they differentiate from each other. To make it slightly more fun and interactive, we made it so that each of the top ~50 profiled researchers got a "stat card" that shows off their relative strengths and weaknesses in a few core metrics. And you can even have the labs "face off" against each other in a mock head-to-head, lol.

*Scores are data-based but ultimately subjective and meant to spur discussion / be entirely for fun!!

[Tech Buzz China @TechBuzzChina]

When we wrote about CATL's supplier-financing platforms last November, the interesting part wasn't that CATL made better batteries. It was that the company kept expanding beyond manufacturing into the parts of the ecosystem that shaped purchasing decisions, financing, and long-term customer relationships.

Its recent $1.45 billion push into data center infrastructure suggests it may be trying something similar in AI.

CATL spent roughly $565 million to take control of Zhongheng Electric, a major supplier of data center power equipment whose HVDC systems are already used in facilities operated by Alibaba, Tencent, ByteDance, and Baidu. It also spent $942 million to become the largest shareholder of 21Vianet, one of China's biggest carrier-neutral data center operators.

Taken together, the deals look less like passive financial investments and more like an attempt to move closer to the center of AI data center buildouts. Zhongheng gives CATL a stronger position inside data center power infrastructure, while 21Vianet gives it exposure to a large and growing base of power demand tied to AI workloads.

The broader pattern is familiar. In EVs, CATL steadily expanded beyond battery manufacturing into materials, components, financing, and partnerships with automakers. The company did not need to control the entire industry to gain influence over it.

Something similar may now be happening in AI infrastructure. CATL already sits at the center of battery and energy storage supply. Zhongheng extends its reach into power systems, while 21Vianet connects it more directly to data center demand.

The AI infrastructure market is obviously far more crowded and competitive, with players ranging from NVIDIA and Huawei to Vertiv, Delta, utilities, and cloud providers. But CATL's recent moves suggest the company sees AI infrastructure not just as another battery customer, but as a system where tighter integration between power supply, storage, and data centers could matter over time.

The bigger question is whether these are simply opportunistic investments or the early stages of a longer-term strategy around AI energy infrastructure. If it is the latter, CATL may end up playing a much larger role in the economics of AI data centers than people currently expect.

[Tech Buzz China @TechBuzzChina]

When we wrote about BYD's export ambitions in July 2025, the EU's 17.4% countervailing duty on Chinese BEVs had just taken effect. April 2026 registration data from ACEA shows that tariff barrier hasn't choked off momentum.

EU+UK+EFTA new car registrations rose 7% YoY to 1.15 million units, with electrified vehicles jumping 21% and capturing over two-thirds of sales, while gasoline and diesel fell 15% and 17% respectively.

@BYDCompany registrations surged 114.5% to 27,008 units, more than double Tesla's 10,654 (up 46.5%), while Chery posted roughly 322% growth from a smaller base, with both still reliant on imports rather than local production.

The shift to electrified powertrains is accelerating, and Chinese brands are capturing a disproportionate share of that growth. Local production remains the key variable: once BYD's Hungary plant ramps and eliminates the tariff, the cost advantage could convert even faster share gains.

[Tech Buzz China @TechBuzzChina]

People probably forgot that Baichuan began as one of the original 6 independent foundational model makers in China. Now, only 4 survive (Zhipu/Z ai, Minimax, Moonshot/Kimi and StepFun). Kaifu Lee's 01 basically folded and Baichuan has bet its future on a specific vertical, healthcare.

Wang Xiaochuan, formerly founder and CEO of Sogou, one of China's leading search engines way back when, has cut the Baichuan team to under 300 and shelved general models to build healthcare-specific AI. The new M4 medical model tops HealthBench’s hard and professional subsets, and the patient-facing “Baixiaoyi” agent now operates inside Beijing Children’s Hospital alongside physicians, with diagnosis agreement rates reaching 95% and deployment already extended to 150 county hospitals in Hebei.

The approach uses reinforcement learning rather than doctor replication. Physicians build reward functions from diagnostic pathways, not case data alone, forcing the model to learn clinical reasoning instead of mimicking individual doctors. A WeChat bot version handles proactive follow-ups, medication reminders, and health data management, aiming to create a family doctor rather than a search engine.

[Tech Buzz China @TechBuzzChina]

Did you know BYD has its own self-designed 4nm automotive chip, the Xuanji A3?

BYD has been building chips since 2002. Today, it operates five wafer fabs and controls nearly the entire semiconductor value chain, from IC design to manufacturing and packaging.

The company has invested more than 100 billion yuan (~US$13.8 billion) into chips and employs over 7,000 semiconductor R&D engineers. Three Xuanji A3 chips together deliver more than 2,100 TOPS of compute, and BYD says its hardware-software co-optimization effectively doubles utilization.

What's most interesting isn't whether this brings L4 autonomy any sooner. It's whether BYD's deep vertical integration into semiconductors starts reshaping automotive gross margins in ways that simple price-per-vehicle comparisons miss.

After all, depreciation on five fabs doesn't fluctuate with vehicle volumes.

[Tech Buzz China @TechBuzzChina]

Embodied intelligence job postings in China rose 15x year-over-year in the first four months of 2026, according to Maimai. Average monthly pay increased from 59,000 yuan (~US$8,200) to 62,000 yuan (~US$8,600).

At the high end, Shenzhen-based UBTECH is offering chief scientist roles paying 450,000–500,000 yuan per month (~US$62,000–69,000), or 5.4–6.0 million yuan annually (~US$750,000–830,000). Similar searches advertised by headhunters often have no stated upper limit, reflecting a severe talent shortage rather than normal growth hiring.

To address the gap, China's Ministry of Education approved nine leading engineering universities to launch embodied intelligence undergraduate programs starting with the 2026 gaokao cycle. Combined enrollment is only a few hundred students annually, with the first graduates not expected until 2030.

The shortage spans AI researchers working on VLA models, world models, and reinforcement learning; hardware and systems engineers focused on robotics platforms and integration; and commercialization talent capable of turning prototypes into products. For the foreseeable future, companies will continue recruiting heavily from autonomous driving, industrial robotics, and foundation model teams, keeping salary pressure structurally high.