Humanoids are as much a compute story as a robotics one

Barclays analysts have warned that humanoid robots may create demand for data center infrastructure, AI accelerators, memory, and edge processors even before they become widely used. While investment has largely focused on visible hardware components like motors, actuators, sensors, and batteries, the biggest near-term challenge lies in the intelligence required for these robots to navigate and adapt to their surroundings.

The computing stack consists of three main layers: data center systems generate synthetic training data through simulations, foundation models translate perception into action, and processors inside each robot execute decisions locally with strict latency, power, and safety constraints.

Simulation is crucial as developers lack the massive datasets needed to train language models. Digital environments can teach robots skills like locomotion, object manipulation, and more before they are deployed in the real world. However, physical testing is still necessary to cover variables that simulations cannot accurately reproduce.

This demand for compute may scale before robot production begins, benefiting chip and infrastructure suppliers such as Nvidia, AMD, Qualcomm, Micron, SK Hynix, Samsung Electronics, and TSMC. Nvidia already offers an integrated robotics stack, including Omniverse for digital environments, Isaac Sim for training, GR00T foundation models, and Jetson processors for on-robot computing.

The humanoid market is currently valued at $2 billion to $3 billion, with forecasts ranging from $10 billion to $25 billion by 2030, and optimistic projections reaching up to $200 billion by 2035. However, widespread economic deployment may take longer, closer to 2035, as safety, reliability, and autonomy issues remain unresolved.

While hardware will become more critical as production increases, current low robot volumes limit standardization and supplier investment. Actuators account for an estimated 30% to 50% of a humanoid's component cost, compared to 10% to 15% for onboard compute, but the latter is still crucial for the robust performance of these advanced robots.