Tech Buzz China

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! techbuzzchina.com

NIO's chip spinout has moved faster than expected. After a Q4 2024 net income of 1.2 billion yuan (~$171 million), the company restructured its chip unit into Anhui Shenji Technology in June 2025, offloading its most capital intensive division. Shenji has since co-founded a joint venture with Aixin Yuanzhi, a domestic ADAS chip designer. This is now the designated channel through which the M97 chip reaches third-party automakers. The structure matters as much as the product. By routing external sales through a JV rather than selling directly, @NIOGlobal insulates its core brand from the friction of competing with customers. The real test is whether the M97 can win design-ins at volume before the JV's cost base outpaces revenue.

BioMap, an AI life sciences company co-founded by Baidu founder Robin Li and former Baidu Ventures CEO Liu Wei, has filed for a Hong Kong IPO. Founded in 2020, BioMap is not building drugs. It is building the computational infrastructure others use to find and develop them. This replaces parts of the trial-and-error lab process with computational prediction across drug discovery, target identification, protein design, and chassis organism engineering. Biodata isn't standard like text, so translating model performance into clinical outcomes requires extra regulatory and experimental validation cycles. BioMap's core product is the xTrimo model series, trained on DNA, RNA, protein, and cell data at roughly the 268-billion parameter tier. Market conditions are favorable. Insilico Medicine raised $337 million in its 2025 Hong Kong IPO and gained roughly 130% post-listing. Hong Kong healthcare fundraising in 2025 has exceeded $130 billion, a four-year high. The Hang Seng Biotech Index is up more than 30% over the past year. Whether BioMap's positioning holds value depends on how investors eventually price model leverage against clinical progress.

Feng E Zushi, a Chinese unmanned retail operator, has deployed an AI agent system across 180,000 smart vending cabinets, making roughly 100 million operational decisions a day. Human interventions? around 4,000. A single restocking worker now manages roughly 120 cabinet locations, up from 60-70 in 2022. The system, called FLOW Pilot, runs four sub-agents covering sales, merchandise planning, supply chain, and logistics. COO Zhu Tao compares the automation ratio to L4 autonomous driving. The company committed 250 million yuan in R&D over four years following a SoftBank-led Series A in 2021, building a 120-130 person algorithm team drawn largely from Meituan, ByteDance, and Tencent.

GGII, a Chinese robotics research firm, just put out a notably optimistic, data-heavy take on how quickly humanoids are moving toward commercialization. Here are some of its key claims: GGII says average unit cost fell to 100,000 yuan (~$14,300) in Q1 2026, down from 150,000 yuan (~$21,400) in 2025, while the average payback period shortened to 12 months, versus 18 months in 2025 and 48 months in 2023. There is also a price war. In March, UBTech reportedly cut its industrial humanoid to 128,000 yuan (~$18,300), Fourier launched the GR-3 at 115,000 yuan (~$16,400), and Unitree priced the G1 at 99,000 yuan (~$14,100), making it the first major model to break below 100,000 yuan (~$14,300). GGII frames 12 months as an important threshold. They quote a manufacturing CFO saying traditional automation usually pays back in 18-24 months, so if humanoids can reliably stay under 12 months, they become much easier to justify as capex. GGII projects 80,000 units shipped in 2026, up 185% from 28,000 in 2025. The piece also says Q2 2026 shipments alone could exceed 25,000, and that full-year shipments could be revised up to 100,000 if prices keep falling. Factory deployments are the main driver of that demand. GGII says factory use cases will account for 72% of 2026 demand, up from 65% in 2025, with the remaining 28% coming from logistics, security, and commercial services. One of the most concrete anecdotes is a factory case study. A Chinese auto-parts company reportedly bought 50 humanoid robots in 2025 for welding and assembly, invested 7.5 million yuan (~$1.07 million), then after 12 months had saved 6.8 million yuan (~$971,000) in labor costs, improved yield by 3.2 percentage points, and reached a combined payback period of 11.5 months. The market is already looking fairly concentrated. UBTech, Fourier, and Unitree together hold 70% share, with UBTech at 30% and 8,500 units, Fourier at 22% and 6,200 units, and Unitree at 18% and 5,100 units. Geographically, the Yangtze River Delta and Pearl River Delta still dominate, accounting for 68% of the market. But the piece says Sichuan, Hubei, and Anhui each saw more than 200% year-on-year growth, driven by auto and electronics clusters. The report is bullish, but it does acknowledge meaningful technical limits. Fault rates in unstructured environments are still 8-12%, versus 2-3% for conventional industrial robots. It also suggests current prices are not yet low enough for true mass adoption. According to the GGII survey cited, 65% of surveyed companies see 50,000-80,000 yuan (~$7,100-$11,400) as the real psychological price band for large-scale adoption. The labor implications are large if these shipment numbers prove right. The piece estimates each humanoid could replace 1.5-2 assembly-line jobs, implying that 80,000 shipments could theoretically affect 120,000-160,000 jobs. On competition, GGII notes that Tesla, Xiaomi, and Huawei are all looming in the background. The piece says Tesla Optimus is expected to begin small-batch deliveries in H2 2026, while Xiaomi and Huawei have announced humanoid strategies but remain pre-deployment.

A group of Chinese tech companies have launched what they are calling the country's first commercial optical circuit switching supernode, connecting 128 GPUs. Traditionally in AI data centers, every single signal has to be converted from light to electricity, before being processed and converted back to light. This burns energy and adds delay. At the scale of tens of thousands of GPUs training a frontier model, inefficiencies like this can turn into a serious bottleneck. Optical circuit switching removes one step in the conversion. Light is routed directly through the network without ever becoming an electrical signal, saving a huge amount of energy. This stack was a concept in mid-2024 and is now commercially deployed. With no access to Nvidia's most advanced chips, Chinese AI companies have to squeeze more out of their hardware through better architecture. The companies involved are Shanghai Yidian, Biren Technology, ZTE, and Xizhi Technology, whose silicon photonics chip is the core of the system.

Nvidia's GTC 2026 keynote had a single premise: agent workloads are so heavy they tend to break existing architecture, so Nvidia has had to rebuild. The company’s new platform Vera Rubin comprises seven chips and five rack systems, and Jensen Huang called it a “generational leap”. The new architecture has some elegance to it: Integration with Groq has split AI tasks into prefill and attention, and the decoding of tokens, as they require different latency levels. In the Nvidia-led US system, design comes first. It is standardized via CUDA and exported globally wherever possible. In China, hardware, models and applications are being tested in deployment, not just design. It may be messier this way, but it is definitely more adaptive.

Alibaba Group has set up a token hub, steering its AI efforts into one of five sub-units: Tongyi Labs, MaaS platform, Qianwen, Wukong, and AI Innovation. Its a bet that AI won’t be monetarized as SaaS. It’ll be metered. So a smart internet giant will be looking to consolidate its strategy. Agents burn through 10–100x more tokens than chat. So whoever owns the meter wins.

Everyone’s watching humanoids. Meanwhile farms are getting automated. Juewu Technology is already deployed at scale, after brutal testing in Xinjiang. Cold, rough, real-world. This robot is now primed for Australia, Northern Europe, Canada...all places where workforces are slim on the ground.

Just how popular is Moonshot AI's K2.5 model? It was only released in January, and in 20 days had made the company more money than in the whole of last year. Now it was making more money abroad than in China, giving it a higher ranking than Gemini, and Claude, albeit for a little while. OpenClaw lists it as a default model. And why? Costs, according to one developer, are one-eighth those of Claude workflows. The technical foundation is an agent cluster architecture coordinating up to 100 parallel AI instances on a single task, trained via a proprietary method the company calls Parallel Agent Reinforcement Learning (PARL). Moonshot is skipping the handouts and freebies that Alibaba and Tencent are bribing consumers with. Red packets and milk teas, you will have to get yourself. Moonshot is building leverage in the global developer ecosystem instead. Cash reserves stand at nearly 100 billion yuan (~$14.3 billion), the company reached a $10 billion valuation in just over two years with 300 employees, faster than ByteDance (4 years) or Pinduoduo (3 years) at comparable milestones.

TEMU WATCH #12: UPDATE ON MARKETS, ADVERTISING & LOGISTICS STRATEGIES In @EdSander's final report for Tech Buzz China, he shares new insights into the global progress of Pinduoduo’s Temu. The report was compiled from over 15 interviews with Chinese experts close to Temu. This first report, Temu Watch #12, is available for free to read on Tech Buzz China. We share updates on Temu’s business development across global regions, as well as the progress of different operational models, merchant margins, advertising spending in the last quarter, and adjustments to logistical processes. Ed would like to thank you all for your support over the years. The report includes information on how you can continue to follow him. techbuzzchina.substack.com/p/temu-watch-1…

AI Will Design Your Next Peptide (And it's getting faster at doing so) MetaNovas (元星智药), an agentic AI materials platform backed by Hillhouse and Fang Yuan Capital, has closed back-to-back Series A+ and A++ rounds. - CEO 王梅杰 (Wang Meijie), formerly at Nvidia's Silicon Valley HQ building AI infrastructure for biocomputing; CTO 余论 (Yu Lun), MIT PhD in nuclear science, engineering and AI, previously chief data scientist at UnitedHealth Group - Fastest pipeline: concept to human efficacy testing in 12 months - First-pass molecule success rate: >60%; molecular generation validity >95% across 10^60 chemical space - Anti-aging Senoreversing peptide validated in 42 molecules, 2 iterations; antimicrobial AMP33 achieved medical device master file (MDF) registration For context on capital model: Lila Sciences, a US AI-native drug and materials discovery company, raised $550M and owns full-stack robotic labs internally. MetaNovas runs CDMO joint ventures for scale-up and brand co-development for commercialization, keeping its own capital deployment low. MetaNovas also records failed experiments alongside successes (of course). That library of failures is fed back into the model, and the company claims this as the main reason its predictions keep getting more accurate over time.

China Humanoid Supply Chain Watch: Zero Difference Cloud Control (零差云控), a Shenzhen robot joint maker founded in December 2016, closed a C+ round. The company's core customers are in surgical robotics, factory automation, and industrial applications where a joint failure has direct consequences. Modular redesign cuts 7-DOF arm component counts from 40 to 10, which reduces assembly cost. - 2025 revenue more than doubled YoY; traditional business contributed ~35% of that growth, humanoid/embodied robots ~65% - Humanoid now accounts for 30%+ of total revenue; industrial and medical segments together ~50% - eRob joints and eCoder encoders serve 2,000+ clients globally, with sales ranked first in the industry - China robot joint module market: 2.24M units in 2024, projected 4.82M units and 68.9B yuan (~$9.8B) by 2030

The head of the Qwen model team at Alibaba, Lin Junyang, has just quit giving no explanation. His public message was brief: essentially a “Bye my beloved Qwen.” He is the third leadership exec to leave Qwen in 2026, and it is only March. The exit surprised many in the AI community because he was considered one of the core technical architects behind recent Qwen models like Qwen3-Max and Qwen3.5. Mostly likely, Lin left because of structural change inside Alibaba’s Tongyi Lab, where Qwen is being developed. The Qwen group had been vertically integrated, with pre-training, post-training, multimodal work and model releases all under the same team’s control. Alibaba may be wanting to split the team into specialized groups, and Lin reportedly believes everyone should stick together. Another underlying factor has to be an internal drive to make money from Qwen, and fast, by pivoting to AI agents and enterprise deployment. The commercialization pressure can clash with researchers who want to prioritize long-term model architecture research.

Global AI glasses shipments reached 8.7 million units in 2025, up 232% year-on-year, according to Omdia data published this week. Meta accounts for 7.4 million of those units (85.2% share, +281.3%), without selling a single unit in mainland China. China's domestic market shipped nearly 1 million units, capturing 10.9% of global volume. Rokid, a Hangzhou-based AR hardware company, and Xiaomi ranked second and third globally by brand, both on the strength of China sales. Chinese firms hold 71% of that sub-segment, which grew from 3.3% to 8.4% of the overall market between 2024 and 2025. Meta's dominant position is built on audio-AI and a fashion-brand partnership. Chinese vendors are converging on a different product architecture: display integration at locally competitive prices. Omdia projects 2026 global shipments above 15 million units. How aggressively can Chinese OEMs price their way into broader distribution?

For EV maker Xpeng, 2026 had a tough beginning. A sluggish year end market, combined with increasingly conservative buyers who would prefer lifetime maintenance over a humanoid robot at the wheel, pushed He Xiaopeng’s high-profile criticism of existing L2-level assisted driving as a "Frankenstein's monster," and prediction that fully autonomous driving would arrive in the next 1-3 years, into the eye of the storm. “L2”, which relies on stitching together various sensors and systems, “is inefficient”, He said. The China Passenger Car Association (CPCA) shows that in January this year, national passenger car sales amounted to 1.544 million units, a year-on-year decrease of 13.9%. In a turbulent market, other automakers are using "refrigerators, TVs, and comfortable sofas" and substantial discounts to stabilize their customer base. He Xiaopeng is going out a limb pushing for the auto industry to leapfrog L3 and move straight to L4 autonomy. Looking at February delivery data: HarmonyOS led with 28,212 vehicles, followed by Leapmotor with 28,067 vehicles. Li Auto maintained its core market share with 26,421 vehicles, while NIO led with a year-on-year sales increase of 57.6%.

Qwen AI Glasses G1 series Song Gang, head of Qwen Consumer AI Hardware, predicts 2026 will be a breakthrough year for the AI glasses market, with more than 2 million units of the new Qwen AI glasses expected to ship within China. That's up from 500,000 in 2025. Song said, “We believe glasses should be wearable all day; suitable for daily use.” The Qwen AI Glasses G1 series was launched at MWC 2026 in Barcelona, and will hit the shelves this month. The G1 starts at 1,997 yuan after subsidies. Qwen differentiates from competitors by integrating Alibaba's ecosystem (Taobao, Amap, Fliggy). The device will support 'one-sentence task completion' via OTA update by end of March. Over 4 million users aged 60+ used Qwen’s 'one-sentence ordering' feature during Chinese New Year 2026, validating demand and informing development.

Will WeChat get an AI agent?? The market is hopeful @TencentGlobal is having a real moment. The stock jumped 7% in Hong Kong, and the ADR kept running, up about 10%. What seems to be driving it is a one-two punch on AI. First, reports this morning said Tencent is rolling out a new workplace AI product called WorkBuddy, "an openclaw AI agent for workplace tasks that supports local installation without cloud deployment… allowing users to automate tasks such as information retrieval, report generation, and content drafting." Then later in the day, @theinformation reported something even bigger: Tencent is also working on an AI agent for WeChat. That is the part that really gets people excited. If Tencent can bring agent-style AI into WeChat itself, you are talking about distribution at an almost absurd scale. The reported plan is for the agent to live inside WeChat as its own chat, which is such an obvious and potentially powerful interface. No new behavior to learn, no separate app, just open WeChat and use it.

Apply here! techbuzzchina.com/trip-new-energ…

Hey friends, we’re about a month out from our New Energy & Robotics Trip (4/20-24), and we’re still open to 1 last-minute joiner. This one is going to be very, very cool. If you’re from one of the 50+ countries that now have unilateral visa-free entry into China,* all you need to do is book your ticket. We’ll be heading to Shanghai and Shenzhen to see humanoid robotics, an EV factory, and more. If you’ve been curious about where China is actually pushing the frontier in hard tech, this is a pretty special chance to see it up close.

World's leading battery maker, China's @catl_official Printed $29M in Daily Profit Last Year, Sending $1.16B to Founder Robin Zeng - CATL's 2025 results: revenue 423.7bn yuan (~$60.5bn, +17%), net profit 72.2bn yuan (~$10.3bn, +42%). - Battery systems output hit a record 748 GWh (+44.96% YoY). Capacity utilization rose from 76% to 96.9% on a 772 GWh installed base, with 321 GWh under construction - Power battery: 541 GWh sold (+41.85%), global share 39.2% (+1.2pp), #1 for 9 consecutive years; but domestic share slipped to 43.42% (-1.67pp), second-largest decline in the top-15 ranking - Energy storage: 121 GWh sold (+29.13%), #1 globally for 5 consecutive years; storage gross margin 26.71% vs. 23.84% for power batteries - Domestic share decline reflects tier-2/3 battery makers gaining ground; analysts at Zhuochuang note quality improvements alongside price competition; EV makers including Li Auto, @XPengMotors, Leapmotor, and GAC Aion have introduced second and third-tier suppliers to reduce single-source risk and cut procurement costs - Overseas revenue 129.6bn yuan (~$18.5bn, 30.6% of total, +17.5%); overseas gross margin 31.44% vs. 24% domestic; H-share listing on Hong Kong Stock Exchange raised 41bn HKD (~$5.3bn USD), earmarked for the Hungary plant and global expansion - The billion dollar payout for the chairman? That came from the 36.1bn yuan (~$5.2bn) dividend which was half of net profit