Power density is the binding constraint in a humanoid actuator, and the first per robot GaN chip count we have seen in print puts a mainstream humanoid at about 300 devices, climbing past 1,000 once the hands and waist are fully articulated [1]. Torque, endurance and how many joints a body can carry all trace back to how much power each drive moves per gram and per cubic centimetre. GaN is the clearest lever on that number, and the count below is what it looks like when a robot maker acts on it.
How much GaN a robot carries
The count comes from a July 2025 report on Zhiyuan, the Shanghai humanoid maker, fitting Innoscience GaN devices into its joint drives [1]. The robots actually shipping carry GaN in three motors of key joints such as the neck and elbows, three chips per motor [1]; the larger numbers are the reporter's calculation for a fully fitted mainstream humanoid. A mainstream humanoid carries 30 to 40 joint motors. Small joints such as fingers take 3 to 6 chips, medium joints such as elbows about 12, and the largest joints 24 [1]. On that arithmetic the baseline is about 300 GaN chips per robot, and adding five finger dexterous hands and waist torsion joints pushes it past 1,000 [1].
Two caveats on the number, both ours. First, it is a calculation published by a business outlet alongside one maker's adoption, not a count from a bill of materials; nobody else has published a per robot device count we can set beside it. Second, the count depends on how many discrete devices sit in each drive. TI's humanoid reference counts about 40 servo motors before the hands and more than a dozen micro motors in each hand [4], which is consistent with the joint count but says nothing about chips per joint. Treat 300 as an order of magnitude with a named source, not a specification.
Why the count is going up, not down
The same report gives the direction of travel. Joint motor power is rising from about 80 W toward 200 W, and a finger joint motor has to go from stationary to 500 rpm within 0.1 s, which the report says requires the drive to support a higher PWM frequency [1]. The report puts silicon switching loss at about 23 percent of the motor's total power consumption, GaN switching loss 85 percent lower and power device chip area more than 50 percent smaller, as general device statements, and attributes the roughly 50 percent reduction in joint drive board volume to Innoscience's own tests [1]. Less loss in less volume is the definition of higher power density, and that is why the maker put GaN into three motors of key joints such as the neck and elbows first and has since installed the devices across hundreds of robots [1].
Our reading of the chip count against the power trend: more power per joint at higher PWM frequency does not lower the device count, it raises it, because the thermal budget per joint is fixed by the housing and the only way to move more watts through the same volume is to switch with less loss. A robot that adds hands and a waist adds sockets; a robot that raises joint power adds parallel devices or larger dies inside the sockets it has.
The market frame
Third party figures, which we flag as such, put the 2026 global GaN power device market near 920 million dollars, up 58 percent on 2025, on a path toward roughly 3 billion dollars by 2030 at a 42 to 44 percent compound rate; the numbers are attributed to Yole and TrendForce by the reporting outlet, and we have not opened the primary reports [2]. The two endpoints do not reconcile with the rate: 0.92 billion compounded for four years at 42 to 44 percent is 3.7 to 4.0 billion, our arithmetic, which is one more reason to carry the figures as reported and no further.
The humanoid unit forecast we do trust is the one in our own market brief: about 8,000 robots in 2025, 136,000 in 2030 and 2.1 million by 2035 on Yole's humanoid analysis [3]. Multiply the 2035 figure by the baseline count and you get on the order of 600 million GaN devices a year in humanoids alone, our arithmetic, before a single hand is counted. That is the scale at which a per joint device count stops being a curiosity and becomes a supply chain.
What we would push on
The 300 figure is a count of discrete devices, and in our reading the interesting engineering question is whether it survives integration. A monolithic half bridge with its driver in one package replaces several discretes and drops the count while raising the density. If the industry follows that path, the chips per robot number falls even as the GaN content in watts rises. The number to watch is not chips per robot but watts of GaN switching per kilogram of robot, and nobody has published that yet.
References
- 36Kr Europe, Is Zhiyuan's Equipping with Innoscience Chips a New Inflection Point for GaN?, 15 July 2025, https://eu.36kr.com/en/p/3379671551140609, figures as reported.
- OFweek Advanced Materials, GaN, Full Takeoff, byline Semiconductor Industry Insights, 16 March 2026, citing Yole and TrendForce, figures as reported.
- Aleph-B, GaN and the Rise of Humanoid Robotics, market brief, 2026, reporting Yole Group, Humanoid Robots 2026, YINTR26535.
- Texas Instruments, Application of GaN FET in Humanoid Robots, SNOAAB3, January 2025.