Compute & AI chips
Onboard processors that run perception, VLA models and real-time control.
Data updated
Why it matters
Humanoids need GPU/NPU compute for vision-language-action models plus safety-rated real-time cores for motor control. NVIDIA Jetson Thor (Blackwell) is the Western default, Qualcomm's Dragonwing IQ10 and Horizon/D-Robotics' RDK S600 compete, and Tesla uses its own AI chips.
Why it is a bottleneck. Running large models onboard within power and thermal limits drives chip choice. US export controls and China's push for domestic chips split the market. The control-system share of BOM is ~10% in BofA's 2030 projection.
Cost. BofA 2030 BOM: control system ~10%.
Suppliers 3
- D-Robotics (Horizon Robotics)RDK S600 (560 INT8 TOPS) / S100 robot compute
- NVIDIAJetson Thor onboard computer; Isaac Sim, GR00T foundation models
- QualcommDragonwing IQ10 robotics platform (up to 700 sparse INT8 TOPS)
Sources for this category
- NVIDIA Blackwell-powered Jetson Thor now available (NVIDIA Newsroom) (Aug 25, 2025)
- Qualcomm introduces a full suite of robotics technologies (Qualcomm) (Jan 5, 2026)
- RDK S600 developer kit (D-Robotics) (accessed Oct 8, 2026)
- Physical AI, part 2: Humanoid robots (Bank of America Institute) (Mar 12, 2026)
- NEURA and Qualcomm: advancing physical AI together (NEURA Robotics) (Mar 9, 2026)
- D-Robotics completes $400 million Series C (PR Newswire) (Sep 17, 2026)