Humanoid Semiconductor Stack

Tesla Optimus · Figure · Unitree · AgiBot · UBTech · Agility · Boston Dynamics

A 40-joint humanoid carries about 260 to 530 active chips, and only one or two are the AI processor. The rest are motor drive transistors, encoders, current sensors, battery monitors and power regulators, mostly analog and power parts from Texas Instruments, Analog Devices, Infineon, ams-OSRAM and Bosch, and the robots carrying them are now built mostly in China: AgiBot and Unitree shipped about three-quarters of the 22,000 humanoids sold in the first half of 2026.

Active semiconductor devices in a 40-joint humanoid robot, low to high estimate: GaN motor drive switches 80 to 160, current sense amplifiers 40 to 80, gate drivers 40 to 80, position encoders 40 to 60, communication ICs 20 to 40, force-torque signal chain 10 to 30, PMICs 8 to 20, motor MCUs 8 to 20, BMS ICs 4 to 12, memory 4 to 8, IMUs 3 to 8, image sensors 2 to 6, and one or two AI inference SoCs CHIPS PER 40-JOINT HUMANOID, LOW TO HIGH ESTIMATE 0 40 80 120 160 GaN motor drive switches 80-160 Current sense amplifiers 40-80 Gate driver ICs 40-80 Position encoder ICs 40-60 Communication ICs 20-40 Force-torque signal chain 10-30 PMIC regulators 8-20 Motor control MCUs 8-20 BMS ICs 4-12 Memory (LPDDR, NAND) 4-8 MEMS IMUs 3-8 Image sensors 2-6 AI inference SoC 1-2 POWER ANALOG AND MIXED-SIGNAL SENSORS DIGITAL COMPUTE AND MEMORY About 260 to 530 active chips, and one or two are the AI processor.
Active chips per robot
260-530
AI SoCs per robot
1-2
Encoders per robot
40-60
Shipments, 1H 2026
22,000+
AgiBot and Unitree
~74%
Hub
Sectors

Six layers

Six Layers of Silicon in a Humanoid

A production humanoid with about 40 degrees of freedom carries roughly 1,100 to 1,500 electronic parts: 800 to 1,200 passives and about 260 to 530 active semiconductors. The same parts serve robotaxis, warehouse robots, cobots and drones, but humanoids set the ceiling, with the most joints and the tightest size, weight and power limits.

The humanoid semiconductor stack
LayerFunctionDevicesPer robotLeading suppliersSupply risk
Joint motor driveThree-phase inverter per joint for about 40 brushless motorsGaN half-bridges, gate drivers, shunts80 to 160 GaN devicesInfineon (with GaN Systems), Texas Instruments, EPC; Innoscience in ChinaMedium
Battery managementMonitors, protects and balances the lithium-ion packCell monitor AFEs, coulomb counters, protection ICs, BMS MCU4 to 12 ICsTexas Instruments (BQ), Analog Devices (LTC)High
Compute power deliverySteps the 48 to 96V bus down to sub-1V SoC railsDC-DC controllers, multiphase VRs, PMICs, LDOs8 to 20 ICsTI, Analog Devices, Renesas, Infineon, Monolithic Power SystemsMedium
Position sensingRotor angle at each joint to close the motor loopMagnetic and inductive encoder ICs40 to 60 ICsams-OSRAM (AS5047P), Texas Instruments (TMAG5170); Novosense in ChinaCritical
Inertial sensingBalance and motion estimationSix-axis MEMS IMUs3 to 8Bosch Sensortec (BMI088), Analog Devices (ADIS)High
Force-torque sensingContact force at wrists and anklesStrain gauge assemblies with instrumentation amplifiers and 24-bit ADCs; no MEMS force-torque IC in production2 to 6 sensorsTI and ADI signal chain; ATI, Bota Systems, in-house bodiesCritical
Inference computePerception, planning and whole-body controlAI SoC with LPDDR5X and NAND1 or 2 SoCsTesla in-house chips, NVIDIA Jetson Thor; Horizon Robotics, Rockchip and Huawei Ascend in ChinaMedium

Power and signal

Most Chips Sit in the Joints

Power and signal paths in a humanoid: the battery pack and BMS feed a 48 to 96 volt bus; PMICs step it down for the AI SoC, memory and sensors; the same bus feeds about 40 joint drives, each with a GaN power stage, gate driver, current sensing, an encoder and a motor control MCU; IMUs, force-torque sensors and cameras report to the SoC POWER (AMBER) AND SIGNAL (BLUE) PATHS IN A HUMANOID Battery and BMS TI BQ, ADI LTC 48-96V bus FEEDS JOINTS AND COMPUTE PMICs and VRMs TI, ADI, RENESAS, INFINEON, MPS AI SoC and memory TESLA, NVIDIA JETSON THOR; LPDDR5X Joint drive, x40 GAN STAGE, GATE DRIVER CURRENT SENSE, ENCODER MOTOR CONTROL MCU INFINEON, TI, EPC, AMS-OSRAM Sensors BOSCH IMUS, CAMERAS, FORCE-TORQUE CONTROL COMMANDS (CAN, ETHERNET) THE JOINT DRIVES TAKE MOST OF THE POWER AND MOST OF THE CHIPS; THE SOC TAKES MOST OF THE SPEND.

Analog and mixed-signal parts outnumber digital compute parts by about 20 to 50 times by unit count. The encoder ICs, 40 to 60 per robot, are the highest-count single risk; the GaN switches, 80 to 160 per robot, are the highest-count power devices. A robot maker that grew out of AI software runs, at scale, an analog procurement business.

Compute

The AI Processor: One or Two per Robot

The inference SoC is the most expensive part in the bill of materials and the most covered, but each robot carries one or two. It runs perception, language and whole-body control at under 10 milliseconds per cycle, with LPDDR5X working memory and NAND for model storage, inside a sealed body with little airflow.

Humanoid inference compute
PlatformWhat it isUsed by
NVIDIA Jetson ThorBlackwell GPU, up to 2,070 FP4 TFLOPS, 128GB LPDDR5X; shipping since August 2025Early adopters named by NVIDIA include Agility Robotics, Boston Dynamics and Figure
Tesla in-house chipsFSD-class inference chips designed by Tesla, made at TSMC and SamsungOptimus
Qualcomm robotics SoCsSnapdragon-based robotics platformsService and industrial robots
Chinese SoCsHorizon Robotics, Rockchip RK3588 class, Huawei Ascend where NVIDIA is restrictedUnitree, UBTech, AgiBot and others

Most fleets train in the cloud and push compact models to robots over the air during charging, the same loop autonomous vehicles use; see Edge AI SoCs.

Production

Chinese Makers Ship Nearly All Humanoids

Counterpoint counted more than 22,000 humanoids shipped in the first half of 2026, up nearly 300% from a year earlier. AgiBot and Unitree shipped about three-quarters of them.

  • US-based maker
  • Abroad
Humanoid shipments, first half of 2026 (Counterpoint)
MakerHomeUnitsShare
AgiBotChinaAbout 9,70043%+
UnitreeChina7,000+31%
GalbotChina1,100+About 5%
UBTechChina1,000+4.4%
Leju RoboticsChinaAbout 6502.9%
Tesla OptimusUnited StatesInternal only; no customer units-
Figure, Agility, Boston Dynamics, 1XUnited States and NorwayPilot deployments-

Tesla says Optimus V3 is ready for mass production on lines being built at Fremont; robots coming off them are used inside Tesla, with output reported at several hundred a week by August 2026. Optimus 4 will be built at a new Optimus factory at Giga Texas, planned for about 10 million robots a year, 10 times the Fremont line's 1 million; Tesla expects production there to start around summer 2027. Hyundai plans to deploy Boston Dynamics' electric Atlas at its Georgia Metaplant.

Fleet demand

Chip Demand at Fleet Scale

Annual chip demand by production volume
LayerPer robot10K robots a year100K a year1M a yearFirst constraint
GaN motor drive80 to 160800K to 1.6M8M to 16M80M to 160MGaN-on-silicon wafer allocation at 100K robots a year
BMS ICs4 to 1240K to 120K400K to 1.2M4M to 12MCompetes with EV BMS programs for TI and ADI supply
PMICs8 to 2080K to 200K800K to 2M8M to 20MMultiphase VRs compete with AI server programs
Position encoders40 to 60400K to 600K4M to 6M40M to 60MCatalog allocation runs out between 10K and 50K robots a year
IMUs3 to 830K to 80K300K to 800K3M to 8MBMI088 competes with automotive ADAS demand
Force-torque sensors2 to 620K to 60K200K to 600K2M to 6MInstrument makers cannot reach 10K robots a year

Bottlenecks

Humanoid Semiconductor Bottlenecks

Bottlenecks by severity
BottleneckLayer20262029Path out
No MEMS force-torque ICForce-torqueCriticalHighA three- to seven-year MEMS program; Bosch or STMicroelectronics most likely; none announced
ams-OSRAM encoder concentrationPosition sensingHighHighTI TMAG5170 ecosystem, Melexis qualification, Novosense in China
Bosch BMI088 single sourceInertialHighMedium-highMurata SCHA634 as an automotive-grade alternative
GaN-on-silicon capacityMotor driveLow-mediumMedium-highTI GaN in Dallas and Aizu, Infineon 300mm GaN at Villach, Innoscience in China; TSMC exits GaN by July 2027
TI and ADI BMS duopolyBatteryMediumMediumRenesas and NXP add limited diversity
No volume tactile skinHandsCriticalCriticalFlexible substrates fall outside standard IC manufacturing

Suppliers

Texas Instruments and Analog Devices Span the Most Layers

Four decades of precision analog process work put TI and Analog Devices in more humanoid layers than any other supplier, and both make most of those parts in their own US fabs.

  • Made in US
  • Some US volume
  • Made abroad
Supplier concentration
SupplierLayersHomeFab model
Texas InstrumentsMotor drive, BMS, PMIC, encoders, signal chainUnited StatesOwn fabs; GaN in Dallas, Texas and Aizu, Japan; 300mm analog in Texas and Utah
Analog DevicesBMS, PMIC, tactical IMUs, signal chainUnited StatesOwn fabs plus TSMC and GlobalFoundries; MEMS in Wilmington, Massachusetts
InfineonGaN motor drive, VRsGermanyOwn fabs (Villach, Kulim) plus foundry
Bosch SensortecIMUsGermanyOwn MEMS fab, Reutlingen
ams-OSRAMEncodersAustriaFabless for encoders (TSMC)
RenesasBMS, VRsJapanOwn fabs plus foundry
STMicroelectronicsBMS, PMIC, IMUs, MCUsFrance and ItalyOwn fabs (Agrate, Crolles)
Monolithic Power Systems48V PMICs and VRsUnited StatesFabless (TSMC)

Two supply chains

Western and Chinese Robots Are Splitting Their Supply Chains

Western programs (Tesla Optimus, Figure, Agility, Boston Dynamics, 1X) and Chinese programs (Unitree, UBTech, AgiBot, Fourier, Xiaomi) are building parallel supply chains. The split is incomplete: Chinese robots still balance on Bosch's BMI088.

Supply by layer, West and China
LayerWestern programsChinese programsSplit
GaN motor driveInfineon, TI, EPCInnoscience, Sanan ICActive
BMS ICsTI BQ, ADI LTCTI and ADI; BYD Microelectronics captiveIncomplete
PMICTI, ADI, Renesas, Infineon, MPSSouthchip, Silergy on secondary railsPartial
Position encodersams-OSRAM, TINovosense NSM2116Active
IMUsBosch BMI088, ADIBosch BMI088 for balance; ACEINNA, Senodia elsewhereBlocked until about 2028
Inference SoCNVIDIA, Tesla, QualcommHorizon Robotics, Rockchip, Huawei Ascend; NVIDIA where allowedActive and widening

Qualification

Design Choices Now Set Supply for Years

Validating a new part in a safety-critical joint takes 12 to 24 months, and humanoids have no equivalent of the automotive AEC-Q100 and ISO 26262 regime, so each maker writes its own. Moving an encoder from ams-OSRAM to TI takes three to six months of firmware and angle-error work; swapping a BMI088 means retuning the balance controller. Parts chosen in the 2025 to 2027 design phase will set supply for three to five years of production, and second sources are cheapest to add before the first production run.

Record

Page Record

Covers
The humanoid semiconductor stack, chip counts, compute, production, fleet demand, bottlenecks, suppliers and the West-China split; absorbs the former Humanoid Compute Stack page
Sources
Counterpoint humanoid shipments (August 2026) via Forbes; Tesla shareholder letter (October 2026); NVIDIA Jetson Thor specifications; supplier datasheets (TI, ADI, Infineon, ams-OSRAM, Bosch); SemiconductorX device-count model
Last verified
October 2026