A market that grows from under a billion dollars to fifty one billion in a decade is not a market for GaN. It is a demand curve for GaN, because the three things a humanoid needs most, control precision, efficiency and miniaturisation, are hard to reach without it. The numbers below come from Yole Group's Humanoid Robots 2026 analysis as we reported them in our market brief [1][2]. The argument about where GaN sits inside those numbers is ours.
Millions of machines, not thousands
Yole projects the humanoid market from roughly 0.6 billion dollars in 2025 to 6 billion dollars in 2030, a 56 percent compound annual rate, and on to 51 billion dollars by 2035 [1]. Unit shipments climb on the same curve: about 8,000 robots in 2025, 136,000 in 2030 and 2.1 million by 2035, on a six year replacement cycle [1].
Adoption is expected in three waves: industrial first, with active pilots at Apptronik, Figure AI and Agility Robotics; then consumer, driven by aggressive price compression from Chinese makers; then medical, once safety and regulation mature [1].
Read the unit curve rather than the dollar curve. Each of those machines carries around 40 servo motors before the hands are counted [3]. At 2.1 million robots a year that is on the order of 80 million joint inverters a year as a floor, a figure we derive from the two stated numbers and which nobody has to believe yet. What matters is the shape: in our reading the joint inverter is the highest volume power electronics socket in the machine, and its volume scales with units, not with price.
Semiconductors ride the same wave, with a falling share
Semiconductors make up about 7 to 8 percent of the humanoid bill of materials in 2025, easing toward roughly 5 percent by 2035 as silicon costs fall with scale [1]. The BOM section below counts 16 percent once every subsystem is included; our reading is that the two figures use different scopes, and we quote both as the brief states them. The absolute number still compounds: total semiconductor revenue tied to humanoids is projected to grow from 41 million dollars in 2025 to 2.3 billion dollars in 2035 [1].
Compute processors claim the largest single slice [1]. The power electronics layer is smaller per robot but it is embedded across nearly every electromechanical subsystem: joint inverters, DC to DC converters, chargers and even LiDAR laser drivers [1]. That is the structural point for a GaN supplier. Compute is one socket that one company can dominate. Power is dozens of sockets spread across the body, and no one owns them yet.
The BOM tells you where GaN goes
On a humanoid BOM of about 32,000 dollars, semiconductor content runs near 5,015 dollars, roughly 16 percent once every subsystem is counted [1]. The subsystems that carry GaN bearing power stages are also the most expensive and the most actuator dense. Dexterous hands alone are about 31 percent of the BOM at 10,000 dollars, followed by rotary actuators at 7,000 dollars and linear actuators at 6,300 dollars [1].
Each of those three blocks embeds joint inverters, gate drivers and current sensors, so GaN content is spread across all of them rather than concentrated in one line item [1]. Yole names GaN explicitly as a key technology in three of the five functional blocks of the power chain: the 48 V bus switches, the DC to DC stages and the joint inverters, with EPC and Infineon called out as central GaN suppliers and Texas Instruments embedded through drivers, current sensors and MCUs [1].
The number we would underline is the hand. Ten thousand dollars of a 32,000 dollar machine, holding more than a dozen micro motors [3], is where the 100 kHz waveform argument for GaN is least abstract. TI's own comparison shows the phase current resolving into a far cleaner sinusoid at 100 kHz than at 10 kHz [3], and for a hand grasping an egg that resolution is the difference between a demo and a deployable machine [2].
Market structure favours a merchant GaN supplier
The 2025 unit leaders are Unitree at 37 percent, Tesla at 21 percent, and AgiBot and UBTECH at 12 percent each [1]. All four are chasing cost reduction and higher performance at the same time, which is exactly the pair GaN addresses by shrinking magnetics, cutting cooling overhead and improving motor efficiency.
More than half of the 60 plus active humanoid makers are based in China and are building vertically integrated supply chains [1]. Our reading of that: the power electronics layer is more fragmented and more open to multiple GaN suppliers, Western and Chinese alike, than the compute layer, and a vertically integrated maker still buys its power stages from somebody.
The hurdles are thermal and they are system level
GaN is part of the solution, not the whole of it. Thermal management was exposed at the August 2025 World Humanoid Robot Games in Beijing, where many units overheated during endurance events [1]. GaN lowers waste heat per inverter, yet 30 to 40 actuators packed into one body still create a cumulative thermal load that the whole cooling architecture must address [1]. Miniaturisation is acute in hands that can hold more than 50 sensors and several micro actuators in a human sized form [1]. And field oriented control runs on every joint, so 100 kHz capable inverters only pay off if gate drivers and current sensors keep pace across the whole signal chain from MCU to winding [1]. That last point is the thread that runs through every article in this notebook.
What we would argue about
The forecast is a forecast. What we would push on is the denominator: 8,000 units in 2025 is a pilot line, and a supplier that sizes its capacity for 2.1 million in 2035 is betting on a curve that has to survive ten years of price compression. For a robot working a four to five hour industrial shift on hot swappable packs, every watt saved at the inverter extends productive runtime [2], and that is the argument a joint drive supplier can win today, on a bench, before the volume arrives.
References
- Yole Group, Humanoid Robots 2026: Market and Technology Trends Report, YINTR26535, 2026, as reported in [2].
- Aleph-B, GaN and the Rise of Humanoid Robotics, market brief, 2026.
- Texas Instruments, Application of GaN FET in Humanoid Robots, SNOAAB3, January 2025.