Tech Buzz China

China's embodied AI companies are racing toward scale, and the bottleneck isn't the robots anymore, it's real-world data. At GEIS, a robotics conference in Silicon Valley, the consensus became clear: synthetic data alone doesn't work because it can't capture friction coefficients, latency, tactile feedback. So the industry has converged on hybrid training. Magic Atoms collects about 16,000 data points daily from real deployments, then synthesizes that 10,000x. Unitree posted 5,500 units shipped in 2025 ($17.07 billion revenue, over 50% from overseas). The companies winning at scale aren't the ones with the best algorithms, they're the ones with the most efficient data loops from actual deployment sites. The hard part isn't building better robots. It's getting robots into messy real environments fast enough to find failure modes labs never see, wet floors, rust, bright light, multiple systems running. OpenMind founder Jan Liphardt made this explicit: deploy early or fail late. That's why new energy vehicle manufacturing is now the primary data mine for training. The factories are already moving parts; robots just need to learn by doing it alongside humans who can correct them. On embodied "brains," VLA (vision-language-action) dominates because touch sensors are still immature. Amazon AI researcher Haozhi Qi noted the architecture choice is really just engineering pragmatism: vision sensors work, so use them to compensate for weak tactile systems. Meanwhile, dexterous hands are splitting into three paths: linkage (cheap, simple), tendon (fine manipulation, expensive), direct drive (balanced, heat management unsolved). The emerging winner is hybrid: tendon structure for precision plus AI for control. The bottleneck in dexterity isn't physics, it's learning efficiency. A robot that can't learn from failure stays expensive.

Tencent is preparing to add a new assistant to WeChat that users can invoke with a right swipe, letting them book cars, order meals, buy groceries, or check tickets without leaving the chat. It’s the first time in the platform’s 15‑year history that this kind of agent is being placed on par with the core messaging interface. The move comes just weeks after Pony Ma told shareholders that Tencent was still struggling to find its footing in AI. Yuanbao, Tencent’s own chatbot, had slipped back to a bit over 40 million monthly active users once the Lunar New Year cash giveaways stopped—about one‑sixth the size of ByteDance’s Doubao. The company could have pushed harder on Yuanbao, but instead it went back to basics: in April it launched Hunyuan 3.0, a rebuilt model overseen by former OpenAI researcher Yao Shunyu and reporting directly to President Martin Lau. A month earlier it folded AI Lab into Hunyuan’s teams. First‑quarter spending underlined the shift, with about RMB 37 billion in capital expenditure and RMB 22.5 billion in R&D, more than 70 percent of which went to staff. When tools like OpenClaw went viral, Tencent reacted quickly, rolling out multiple agent‑style products across at least three business groups. The logic is that the next battleground is not just who has the best model, but who controls the entry points and service networks. In that sense Tencent has a strong hand: its chat apps reach 1.4 billion users and connect millions of mini‑programs, payments, and social relationships. The design being tested is cautious. The assistant sits behind a swipe gesture rather than taking over the chat list or feed. That reflects two unresolved issues: WeChat would need deep access to users’ chats, contacts, and payments to work at its best, and an all‑purpose agent could disrupt the ecosystem of merchants, creators, and advertisers who rely on search rankings and storefronts. Those challenges aren’t solved by the June 2 preview. What’s changed is that Tencent is openly giving an assistant an equal footing with messaging. Over the coming months, how tightly this tool integrates with the mini‑program ecosystem will show whether it’s a conservative test or the start of a bigger rethink.

MiniMax listed in Hong Kong three months ago with a 109% first-day pop, and its chief agent architect is now saying fewer than five large model companies will survive the next two years. That's a pointed claim from a company that spent only $449 million on total model training and cloud costs between 2022 and September 2025. MiniMax listed in HK three months ago with a 109% first-day pop. Its chief agent architect is now saying fewer than five large model companies will survive the next two years. That's a pointed claim from a company that only spent $449 million on total model training and cloud costs between 2022 and September 2025. The bet is that as AI agents replace chatbots as the primary interface, raw model performance converges and the delta shifts to how efficiently you extract value from each token. MiniMax's read is that harness design, sandbox iteration speed, and inference productivity become the actual competition. China's daily token call volume hit 140 trillion in March 2026, up more than 1,000x from early 2024. MiniMax framed this four years ago as tokens becoming a commodity energy unit, like electricity, and built accordingly. Today, 73% of revenue comes from overseas, serving 236 million individual users and 200,000+ enterprise clients and developers across more than 100 countries. The strategic upgrade announced post-IPO, from large model company to "AI platform company," is the logical endpoint of that positioning. Defining the infrastructure layer for token delivery globally is a different business than competing on benchmark scores.

China's securities regulator has approved Unitree Technology (宇树科技)'s IPO registration on the STAR Market. Unitree, founded in August 2016 by Wang Xingxing and headquartered in Hangzhou, is a leading domestic developer of civil quadruped and humanoid robots and a global leader in four‑legged robot sales. It is a national high‑tech enterprise and a national‑level 专精特新"小巨人" enterprise. As of June 2025 the company had more than 1,000 employees. Unitree's core advantage is that it self‑develops the full stack for joint motors, reducers, and controllers, with key parts costing about one‑third of comparable imported products. Our previous coverage questioned whether Unitree's commercial case matches its hardware reputation. The company clearly builds the robot body and low‑cost, capable actuators. What is less clear is how it builds the software, AI, and applications that bring repeat sales or service fees. The regulator's approval does not resolve that question. Unitree is now set to list. Pricing and first‑day trading will show whether investors are paying for the hardware story alone or demanding proof that Unitree can sell software, services, or repeat contracts.