IT Home On October 26, according to the latest global humanoid robot tracking report by Counterpoint Research, it is expected that by 2030, the total shipments of global humanoid robot AI accelerators will exceed 1 million units, and the compound annual growth rate (CAGR) during 2026 to 2030 will be as high as 144%.
In terms of computing architecture, due to the gradual integration of the Visual-Language-Action (VLA) model with the world model, GPUs will remain the mainstream architecture for humanoid robot computing power in the short term. Meanwhile, AI ASICs are also rising rapidly, and vertical integration manufacturers such as Tesla and Xpeng are transferring their in-house chip technologies from the automotive sector to their humanoid robot product lines.
NVIDIA dominates the humanoid robot GPU market with its complete full-stack ecosystem ranging from computing chips, models to simulation tools. It has established deep collaborations with leading robot bodies such as iVoiR, Ustwo AI, Uestar, Yibei Select Robotics, Galaxy General, Figure AI, Neura Robotics, and Noetix Robotics.
Qualcomm has entered the market with its heterogeneous computing architecture, focusing on energy efficiency advantages and lower Bill of Materials (BoM) costs. It has already reached design collaborations with Neura Robotics, VinMotion, Booster Robotics, and Figure AI, but there is still a significant gap in production volume compared to Nvidia.
Soumen Mandal, chief analyst at Counterpoint, said that the computing power ecosystem of humanoid robots is shifting from GPU dominance to the coexistence of multiple AI computing architectures. As robot models evolve from VLA to World Action Model (WAM), the value of dedicated computing power in terms of lower power consumption, smaller cooling constraints, and higher efficiency will become even more prominent.
Counterpoint Vice President Ethan Qi pointed out that GPU remains the preferred choice for current advanced embodied model on-device deployment, and automotive companies like Tesla will become the key promoters of ASIC architectures. As the next-generation AI accelerator capable of supporting general artificial intelligence (AGI) workloads enters mass production after 2027, market competition is expected to intensify further.
In the vertical integration camp, Tesla and Xpeng Automobile, relying on automotive manufacturing experience, self-developed chips, proprietary AI models, and a full-stack software and hardware layout, are expected to reshape the landscape of humanoid robot AI accelerators.
In China, D-Robotics (Sweet Potato Robotics), which was separated from Horizon, has reached cooperation agreements with manufacturers such as Leju Robotics and LIMX Dynamics. Its RDK S100 integrated development kit adopts a “CPU+BPU+MCU” architecture, featuring a BPU neural network processing unit based on the Nash architecture with a computing power of 80 TOPS. This continues to enhance the autonomous capabilities of the domestic humanoid robot computing power ecosystem.