Humanoid robot shipments surged nearly 300% in H1 2026, but battery life and compute architecture remain the hard limits

Global humanoid robot shipments exceeded 22,000 units in H1 2026, rising nearly 300% year-on-year, according to Counterpoint Research - a pace that has moved the question from "will humanoids work?" to "what is stopping them from scaling?" The answer, increasingly, sits inside the chip stack.^[1]

The compute problem is not one problem

Humanoid robots do not run on a single processor. The common denominator across all robot architectures is a heterogeneous mix of MCUs, MPUs, CPUs, GPUs, image signal processors, and digital signal processors, with no settled standard for how those elements connect. Some designs mirror software-defined autonomous vehicles, with distributed zonal compute; others use a centralised drone-style architecture.

Ronald Stärz, system architect for humanoid robots at Infineon Technologies, describes the likely convergence point: "Humanoid compute is likely to be a hybrid model of a central powerful processing system with distributed or zonal compute for pre-processing of data from places such as the fingers. This mirrors the shift automotive already made, from centralised ECUs to zonal architectures, driven by latency, bandwidth, and fault tolerance requirements."

The automotive parallel is instructive. Zonal architectures in cars took roughly a decade to standardise. Humanoid robots are arriving at the same conclusion under far greater time pressure.

Battery life is the deployment ceiling

Whatever the compute topology, the robot still has to run on a battery - and current cells are not up to the job. Most humanoid robots in 2026 are limited to 1-4 hours of active operation on a single charge, making continuous industrial shifts impractical without dedicated charging infrastructure. Industrial customers typically expect 95-99% uptime; most deployed units today manage 30-90 minutes before needing intervention.

Gallium nitride power devices are one route toward extending that window. Infineon's Adam White has noted that GaN-based motor drives can reduce joint size and cut power draw simultaneously - a meaningful gain when dozens of actuators are running in parallel across a humanoid's limbs and hands.

The semiconductor supply chain is not ready

The chip industry's exposure to humanoid robots goes well beyond processors. A full humanoid draws on:

• Motor-control ICs and gate drivers for every joint
• MEMS sensors for force, torque, and environmental perception
• Single-pair Ethernet PHYs and CAN-FD controllers for the in-robot network
• Power-management ICs and battery-management systems
• Edge-AI inference accelerators for vision and sensor fusion

Qualcomm launched the Dragonwing IQ10, a humanoid-specific processor, working with Figure AI and Neura Robotics in 2026, while Infineon, NXP, STMicroelectronics, and Texas Instruments all formalised humanoid-specific product lines covering motor control, real-time communications, sensor fusion, and safety logic. The supply chain is organising - but capacity investment decisions made now will determine whether the production ramp of 2028-2032 is supply-enabled or supply-constrained.

US robotic sales reached $11.4 billion in 2026, up 29% year-on-year, and Berg Insight forecasts annual humanoid shipments growing from roughly 16,000 units in 2026 to 26 million by 2040 at a CAGR of 63.7%. Those numbers imply semiconductor demand on a scale the industry has not yet begun to provision for.

The near-term signal to watch is whether H2 2026 pilot deployments in manufacturing and logistics convert into multi-year volume commitments. That is the trigger that will force dedicated tooling and capacity investment across the motor-control, MEMS, and power-device supply chains - and determine whether the humanoid ramp runs ahead of its chips or behind them.^[1]