Hands, Not Bodies
China's humanoid robot race has a new battlefield: the hand layer.
Things you might have missed
A quick look at what we have been covering and where Tech Buzz China has shown up since our last issue.
From our recent coverage:
China is building an AI “supercluster” in Wuxi around Huawei Ascend chips. On May 15, Hongxin Electronics signed an agreement with the Wuxi High-tech Zone, part of a national push to replace Nvidia-centric AI infrastructure with domestic stacks. (our thread)
Chinese capital is rotating hard into AI and robots. April investment data showed funding for AI and humanoid robots up 175% year-over-year, with data, compute, and network infrastructure up 61.7%. (our thread)
Weijin Research: the compute boom is spreading beyond GPUs. A sharp read from Weijin Research on how AI infrastructure spending is now pulling in memory, storage, and CPUs, and whether token demand can justify the build-out. (read it)
Where we have been featured:
Fast Company (May 2026): “Why Nvidia’s Jensen Huang Joined Trump’s China Summit in the End.”
The full list of our press coverage is at techbuzzchina.com/media.
Table of Content
The marathon was looking at the wrong body part
The hand is the actual unsolved problem
Three bets, three time horizons
Introduction
Eight months ago, in Unitree: Humanoid Hype vs. Robotic Reality (March 2025), we argued that the body of humanoid robotics was attracting most of the attention and most of the capital, but the harder unsolved problems sat elsewhere in the stack. In Robots That Work Beat Robots That Wow (August 2025) we extended that into IPO valuation, arguing the Hong Kong robotics IPOs were rewarding deployment, not demo. In China’s Tesla: Beyond the Car (December 2025) we traced how Xpeng and Xiaomi were each trying to translate auto-grade supply chains into embodied AI platforms. And in the Deep Tech Trip notes (October 2025) we wrote down what we saw on the factory floor: a Chinese hardware ecosystem moving faster than the foundation-model coverage suggested.
This piece is the next step in that thread.
The marathon was looking at the wrong body part
On April 19, 2026, Honor’s D1 (Lightning) humanoid finished the Beijing Yizhuang half-marathon in 50 minutes 26 seconds, sweeping the top six places and beating last year’s winning time by roughly two-thirds. Tiangong took 2 hours 40 minutes in 2025; Lightning ran the same course autonomously, at a pace that would be competitive with elite human marathoners.
The headlines called it a breakthrough. The reality is narrower: this was a smartphone supply-chain victory dressed in a robot suit. Honor’s edge was thermal management: liquid cooling adapted from its phones, with coolant flow up to 4 L/min holding motor temperature at 31.5°C across 21 kilometers. Lan Si Tech supplied 132 specialized metal structural components, with 500,000-unit annual capacity already in place. Peak joint torque hit 600 Nm. These advances are less fundamental robotics breakthroughs than mature EV and smartphone supply-chain components adapted for legged systems. Locomotion is commoditizing faster than the discourse reflects.
The autonomy data points the same direction. In 2025, near-zero teams ran autonomously. In 2026, 38–40% did, with remote-controlled robots receiving a 1.2× time penalty. But Oregon State’s Alan Fern calls this specialized autonomy: rehearsed routes, support vehicles trailing each robot, intervention permitted at battery swaps and falls. Useful, real, but not deployment-grade. Closer to Tesla early lane-following than to anything that handles a crowded sidewalk.
The marathon tested everything except the part of the robot that does useful work. Body control was solved, and rehearsed-route autonomy was solved, but open-environment manipulation never appeared on the course at all, because nothing on the course demanded it. The race was a 21-kilometer demonstration that legs commoditize, and it said nothing about hands.
UBTech founder Zhou Jian puts it bluntly: punching-and-dancing demos are local motion control, not embodied intelligence. Lingzu Era COO Shao Yuanxin frames the standard-setting point: marathons exist because the market has not yet specified one, so the event forces it. Note what isn’t being standardized: manipulation. The marathon is a real engineering achievement, but it focuses attention on the body, and away from the messy industrial environments where humanoid robots actually have to work: variable lighting, irregular objects, unexpected obstacles, tasks that change shift to shift.
These are hand problems rather than leg problems, and the reason is that locomotion is a smartphone-supply-chain problem: Chinese smartphone and EV suppliers already produce humanoid-grade structural components at hundreds of thousands of units a year, components are commoditizing, and prices will follow within 12 to 24 months. Hands will not follow the same curve, and if the hand is priced as a component rather than as a layer, it will be systematically undervalued.
The hand is the actual unsolved problem
A dexterous hand is not a better industrial gripper. Industrial grippers offer 1 to 6 degrees of freedom, enough to grab a defined object on a structured line, not to handle a thin invoice or hold a smooth bottle. The human hand has 22 to 27 degrees of freedom; LinkerBot’s L30 offers 22, Sharpa 22, and Yuequan’s Y-Hand M1 reaches 38, among the highest ever achieved in humanoid robotic hands. Musk has admitted Optimus’s hand alone consumes over 50% of his team’s engineering effort; the hand is where the physics is unsolved.
Why is the hand so much harder than the body? Density and dynamic range, first. Packing 20+ degrees of freedom into a hand-sized volume requires miniaturized motors and force sensors at every joint, with tactile feedback denser than anywhere else on the robot. Sharpa’s fingertips carry 1,000+ tactile pixels each. The same finger has to grip an egg gently and lift a two-kilogram bottle firmly. Second, integration. You cannot build a high-performance hand as a piece of hardware and then simply “plug in” software later. Dexterity requires a perfectly closed loop where tactile sensors and motors communicate in milliseconds. If the software isn’t co-designed with the physical hardware from day one, the system suffers from lag and coordination errors that make delicate tasks impossible. It’s also why a single high-end hand can cost more than the entire humanoid body it’s mounted on.
A traditional 2-finger gripper is task-customized hardware: bolt it onto a line, reprogram per SKU, depreciate over years. That is automation: efficient inside structured environments. A dexterous hand is the inverse: general-purpose hardware adapting to tasks via data and learning. Grippers are CAPEX per line, while dexterous hands can be platform investments compounding across tasks.
The price spread tells you the market hasn’t decided what these things are, and it also tells you where deployment breaks. A mature category trades in a tight band: smartphones in roughly 5–10×, industrial grippers in 5–10×. Dexterous hands span 150×: LinkerBot’s O6 lite at ¥3,999 is entry-level industrial; Yuequan’s Y-Hand at $85,000 is research-grade biomimetic; LinkerBot’s L30 at ¥100,000 sits between them; AGILINK shipped the first sub-¥10K high-DOF hand at ¥9,800 in August 2025. These aren’t variants of one product. They’re different definitions of what a “dexterous hand” is, which means there’s no settled comparable to value them against. And a rough industry rule shows why that matters: if the hand costs more than ~10% of the robot body it’s mounted on, scaled deployment doesn’t work. A ¥99K Unitree G1 can absorb a ¥10K hand; a ¥500K AgiBot platform can absorb a ¥50K hand. Most current high-end hands break this ratio: Yuequan’s $85,000 Y-Hand is several times the cost of any robot body it could mount on, which is why it sells only to defense, not factories. AGILINK’s ¥9,800 OmniHand Agile is the first to sit clearly under the 10% threshold for mid-range Chinese humanoid platforms, the inflection point that turns a dexterous hand from “lab luxury” into “mass-producible industrial product.” LinkerBot’s ¥3,999 O6 lite drives the threshold lower. The deployment race isn’t about who’s cheapest. It’s about who crosses 10% first, for which class of robot body.
Three actuator paths compete inside that category. The tendon-drive approach, which Tesla uses in Optimus, puts the motors in the forearm and runs cables out to pull the fingers, a design Musk calls more bionic and more first-principles. The rigid-linkage approach is a straighter translation of traditional industrial mechanism into a hand. The direct-drive approach, which Sharpa has chosen, places a motor at every joint and so depends on proprietary micro-motors that are at once small and powerful. Each choice locks in a manufacturing strategy, a sensor architecture, and a data pipeline, which is why the companies are not really betting on technologies at all. They are betting on which path produces useful tactile data first.
LinkerBot (engineering scale) pursues all three actuator paths in parallel, with three core components in-house: tactile sensors, motors, and reducers. Founder Zhou Yong runs a two-team racing mechanism: two internal teams race on the same sub-component, and the winner ships. The bet is that whoever mass-produces high-DOF hands most cheaply wins.
AGILINK (spun off from AgiBot at end-2025) frames the hand as the last 10 centimeters of embodied AI: hardware, perception, control, algorithms, and data must advance together. AGILINK inherited AgiBot’s body-deployment data flywheel and is building its own manipulation foundation model plus a 1M-data-point dataset for 2026. The bet is that ecosystem integration, climbing from the hand up to the model, wins as foundation models mature.
Yuequan Bionic (academic-led, founders from CAS and Manchester) argues rigid mechanical paradigms simply cannot match human dexterity. Their answer is a tension-compression structure mimicking human ligaments: 38 DOF, 28.7 kg grip strength, 0.04 mm precision, sold to the defense industry at $85,000 per unit. The bet is that when rigid hands hit a physical ceiling, biomimicry wins.
Boston Dynamics redesigned Atlas’s hand for CES 2026. Tesla rebuilt Optimus’s. The body race draws the attention, but the hand is the category traditional automation never built, and the one the market is mispricing.
The valuation mismatch
The valuation gap shows up most starkly in a single comparison: three humanoid companies on one side, and on the other a lone Chinese hand company that ships more units than any one of them.
Figure AI, a US humanoid robot company, raised money in September 2025 at a $39 billion valuation. Apptronik, another US humanoid company, closed at $5.5 billion in February 2026, partly because Google chose it as the exclusive partner for its new robot AI. Both companies are still in early testing. They have not shipped robots at scale yet.
Now compare LinkerBot, the largest specialist supplier of high-DOF dexterous hands in the world. The company reports holding 80%+ of the global market for these hands, shipped 10,000+ units across 2025, and targets 10,000 units per month by end-2026. It closed its last funding round at $3 billion and is reportedly targeting $6 billion for the next one.
Apptronik is worth almost twice LinkerBot, even though Apptronik ships less, holds less market share, and is at a similar early stage. Why? Because the market values American humanoid companies the way it values AI platforms: Figure and Apptronik are seen as building both the brain (foundation models, AI software) and the body (hardware platform), bundled in one company. That’s a familiar valuation template, the same logic as OpenAI’s, or Tesla’s autonomy premium. LinkerBot doesn’t fit that template. It makes hardware. It doesn’t position as the AI platform. So the market values it as a supplier, not a platform. There is no comparable model for “world’s dominant dexterous hand company” yet.
However, the hand layer is producing real revenue while the body is producing valuations, and the mismatch between them is the story.
LinkerBot CEO Zhou Yong puts it simply: most Chinese factories don’t need a full humanoid robot. Two arms and a pair of dexterous hands are enough. So his customers attach LinkerBot’s hands directly to industrial robot arms they already own. AGILINK is doing the same thing across a wider product line. Yuequan skips consumer humanoids entirely and sells its $85,000 hands to defense customers. Three Chinese companies, three deployment patterns, all generating real revenue, without needing a full humanoid robot to walk out of the lab.