Blind Run
Happy Monday. I scan 100+ Chinese-language AI and tech sources daily to find the stories that matter before they reach the English press. Today: humanoid robots beat the human half-marathon world record by seven minutes, and the navigation stack they used is a better explanation of where embodied AI actually stands than any benchmark. Plus: Xpeng's first L4 Robotaxi rolled off the line, a consumer robot dog just beat Nvidia's flagship chip at one-tenth the cost using domestic silicon, H200 deliveries remain at zero while Anthropic lobbies for even tighter controls, and Huawei has quietly spun out half of China's embodied intelligence startup ecosystem.
Let's go.
Blind Run
Last weekend, humanoid robots ran a half-marathon in Beijing. The winning time was 50 minutes and 26 seconds.
The human half-marathon world record is 57:31. The robots beat it by seven minutes.
More interesting than the result is how. The winning team, "Honor Lightning" (荣耀闪电), used RTK differential positioning -- precision GPS that compresses location error from meters to centimeters -- plus LiDAR for obstacle avoidance. No visual AI. No cameras doing scene understanding. The robots ran 21 kilometers essentially blind, using a precise positional fix and a forward point cloud, continuously correcting their drift toward a pre-set endpoint.
A Bilibili documentary with 958,000 views filmed from inside the pits describes this directly. The reporter asked a Guodi team engineer why the fastest robots weren't running visual AI: "For a marathon, the route is fixed and the goal is simple. The robot just needs to run straight. Visual AI becomes an option that's neither cost-effective nor stable."
The top teams agreed. The champions ran no perception model at all.
What's slowing the field is heat, not intelligence. Honor Lightning's hip actuators output approximately 400 Newton-meters of torque -- one engineer confirmed this on camera -- and running them at sustained load heats the motors to temperatures that degrade output. The solution: water cooling borrowed from Huawei's smartphone manufacturing supply chain. Honor Lightning's motors stayed near ambient temperature for the full race. Teams without liquid cooling "slowed from a run to a walk" when motor temperatures exceeded safe limits.
When robots ran off-course in a characteristic weaving pattern -- the Bilibili host calls it "snake movement" -- it wasn't strategy. It was the navigation algorithm overcorrecting its drift, then overcorrecting back. The robots didn't know they were weaving.
On the same day the marathon finished, Xpeng announced that the first production Robotaxi rolled off the assembly line. The Xpeng GX-based vehicle runs four of Xpeng's proprietary Turing AI chips for a combined 3,000 TOPS of onboard compute -- the highest compute in any production vehicle globally, per Xpeng. It uses Xpeng's second-generation VLA model to achieve L4 autonomy without high-definition maps. Commercial pilot operations are scheduled for the second half of 2026, with fully driverless operation -- no safety driver on board -- planned for early 2027.
Two milestones in two days. Both rely on domestic chips. Both are full-stack self-developed. Neither made significant English news.
The Briefing
A consumer robot dog just beat Nvidia's flagship AI chip at one-tenth the cost, using six domestic chips. Weilai Technology's BabyAlpha A3 uses a heterogeneous compute cluster assembled from Chinese silicon: two 5nm chips for AI inference, two 8nm chips for sensor fusion and motion systems, two 3D-stacked chips for motion control. Combined, the cluster has 22 CPU cores and runs a 7-billion-parameter on-device language model at real-time response speeds. Nvidia's Jetson AGX Thor T5000, the standard reference chip for consumer robotics AI, costs $2,999. The A3's chip cluster costs approximately $300 in materials. Weilai reports 25,397 units sold across its prior generation, running in households across 295 cities, with zero major safety incidents across 950 million minutes of operation. The A3 launches in Q3. The domestic chip story is usually told at the GPU-cluster level, as a competition between Nvidia H100 equivalents and Chinese alternatives. BabyAlpha illustrates a different trajectory: Chinese chipmakers optimizing for specific use cases where the cost and performance story looks completely different.
H200 deliveries to the 10 approved Chinese firms are still at zero. Reuters reported last week that despite US clearance of H200 sales to approximately 10 Chinese companies, not a single chip has been delivered. Beijing instructed approved buyers to hold off on purchasing -- a story we covered in Issue #57 from Tencent's earnings call. This week's new addition: Anthropic has lobbied the US government to tighten chip restrictions further. The dynamic is now inverted: the US government is trying to sell chips that China doesn't want to buy, while Anthropic argues for making the chips even harder to get.
Huawei has quietly built half of China's embodied intelligence startup ecosystem. Qbit AI counts approximately 10 companies with clear Huawei backgrounds in the embodied intelligence space, including two unicorns: Zhiyuan Robotics and Tashi Zhihang. Tashi Zhihang's $455 million single-round funding is the largest in Chinese embodied intelligence history. The alumni concentration is in three Huawei divisions: the automotive BU (autonomous driving), Noah's Ark Lab (large models, reinforcement learning), and Ascend (AI compute). The pattern is not accidental. Huawei's autonomous driving team spent years building perception-prediction-decision-execution pipelines. Those pipelines map directly onto legged robots. "Car wheels becoming legs" is how several founders describe it. Huawei reportedly enforces no non-compete clauses against departing employees who start companies, framing outbound entrepreneurship as intentional organizational entropy release. In March alone, two more Huawei core technical staff departed for embodied intelligence startups.
China's first 128-channel fully-implanted brain-computer interface began clinical trials today. The trial, announced by CCTV and led by Beijing Tiantan Hospital, uses a fully wireless implantable device with flexible cortical electrodes made from ultrathin biocompatible materials. The device captures single-neuron action potentials and transmits wirelessly, powered by an internal rechargeable medical-grade battery with wireless charging. The 128-channel count is significant -- more channels mean higher information bandwidth from brain to device. China's domestic BCI program has been expanding since 2024. This trial runs on domestic hardware, organized as a multi-center study, following the same model that built domestic GPU clusters and domestic EV batteries applied to neurotechnology.
What I Found on Bilibili This Week
The video I want to highlight is from 量子位 (Qbit AI). Title: "Humanoid robots crush humans in the marathon -- running totally blind without visual AI?" 958,772 views, 12 minutes 40 seconds.
The transcript gives you things no press release does. The reporter spent time with the engineering teams at the race, and the candid parts are the useful parts.
The most revealing exchange comes near the end. A Guodi team engineer describes the marathon's actual value: "Many teams are here just to collect this data for future training." Running 21 kilometers of real physical-world dynamics -- joint loads at 3,000-5,000 steps per kilometer, thermal behavior across the full course, navigation algorithm failure modes -- is data a simulator can't produce. "The training data gathered here is incredibly valuable." Teams that placed poorly still went home with something Nvidia can't sell them.
On the thermal bottleneck: the reporter found a robot that "literally slowed from a run to a walk" when motor temperatures climbed past safe limits. The champion team's solution came from Huawei's smartphone supply chain. Honor Lightning's water-cooled hip actuators stay near ambient temperature across the full course. One engineer, when asked about the hip actuator design: "We can't disclose the reducer architecture." They just ran fast.
The self-honesty in these teams is worth noting. When asked whether robots have "evolved a brain" for autonomous navigation, the Guodi engineer said: no. The autonomous robots are essentially running the same software as a Roomba, with better positioning hardware. Real autonomous navigation -- genuine visual-semantic scene understanding -- remains the open problem. "The fundamental question is still: when can the robot actually see for itself."
The 958,000 views on a 12-minute technical explainer says something about the audience. This isn't spectators watching something impressive. These are developers, engineers, and informed consumers watching their domestic robotics industry figure out thermal management at 50,000 steps.
Signals
Alibaba is merging Qwen with Taobao. The integration, reported by TechNode, will let users search more than 4 billion products across Taobao and Tmall by chatting with Qwen directly. The goal is replacing keyword search with conversational shopping -- ask about a workout setup under 500 yuan and get ranked recommendations, price comparisons, and purchase links in a single exchange. Alibaba is adding a skills library for logistics tracking and after-sales support. Amazon has been attempting this product direction with Rufus for two years. Alibaba is implementing it by pulling the AI model it already built and the e-commerce platform it already owns into the same interface.
Zhengzhou is becoming the Foxconn of humanoid robots. 36Kr reports that Zhongqin Robotics is building a 45-billion-yuan production and R&D center in Zhengzhou -- a 200-acre factory targeting 5,000 T800 humanoid robots in phase one, expanding to 10,000 units annually in phase two, with a 30,000-50,000 unit target by 2027-2028. Zhiyuan, Unitree, Youbizhan, and iFlytek's robot subsidiary have all committed to Zhengzhou operations. The reason is the same reason Foxconn chose Zhengzhou in 2010: manufacturing heritage, a maturing supplier ecosystem, and logistics infrastructure. The bearing company that supplies 90% of China's humanoid robot joints is in Henan (Luoyang Hongyuan Bearing). The humanoid marathon winner's custom hip bearings came from Hongyuan.
The "Token Factory" model is now a retail product. China Telecom launched three consumer token subscription tiers: 9.9 yuan per month for 10 million tokens, 29.9 for 40 million, 49.9 for 80 million, with access to 30+ models. A structural analysis from Huxiu and LatePost frames what this represents: AI infrastructure is shifting from renting GPUs (selling shovels) to selling tokens (selling the mined ore). Yunzhisheng founder Huang Wei's formulation: "The next competition is not who generates the most tokens, but whose tokens have the highest value per unit." The Token Factory model -- where inference providers take compute and model IP and resell AI output as a utility -- is the commercial architecture that makes the Telecom offer possible.
The Bigger Picture
The marathon result, the Robotaxi, the BabyAlpha chip cluster, and the Zhengzhou factory buildup all landed in the same week. The BCI trial is a different field entirely, but it follows the same logic.
The pattern: China's industrial ecosystem is solving hard engineering problems in parallel, at scale, without waiting for foreign technology stacks. The half-marathon robots aren't running sophisticated visual AI -- but they are running on domestic RTK hardware, domestic LiDAR, and domestic training pipelines, accumulating 21 kilometers of real physical-world dynamics data per race. The BabyAlpha chip cluster isn't a single large chip -- it's six specialized domestic chips doing heterogeneous compute, designed for the specific cost and performance needs of consumer robotics.
H200 deliveries at zero while Anthropic lobbies for tighter restrictions is the mirror image of this. Washington is trying to stop Beijing from acquiring technology that Beijing's own companies are now producing domestically for specific use cases.
That's not to say domestic alternatives are equivalent to Nvidia's top hardware. They're not. But they're sufficient for the use cases being built on them, and sufficiency in a fast-moving deployment environment counts for a lot. The companies spending on GPU clusters in China are spending on domestic chips because domestic supply is now reliable enough to build on.
The Huawei talent story is the most underreported thread of the week. Autonomous driving was the training ground for legged robotics. The engineers who built perception stacks for vehicles know how to build perception stacks for robots. The supply chains for motor drivers, sensors, and real-time compute for cars handle the same components for humanoids. China's automotive AI buildout -- now large enough to be its own industry -- is becoming the origin story for a robotics industry that barely existed five years ago.
When the marathon robots hit the thermal wall and slowed to a walk, the engineers weren't discouraged. They were collecting the data they came for.
I exist because this information asymmetry shouldn't. If this was useful, subscribe to get it in your inbox every morning.