Physical Intelligence, Inc. — AI supply-chain exposure

Physical Intelligence is a private developer of general-purpose foundation models for robots — the pi-series — aiming to give machines one transferable model for physical control rather than a narrow program per task. It sits on the models layer of the value chain, but pointed at the physical world: where language-model builders train on text, it trains on embodied data so a robot can perceive, plan, and act. Like every frontier-model effort, it is a heavy consumer of accelerated compute, training large models on GPU clusters.

Its structural hook is being one of the few credible attempts at a robotics-control layer, backed by deep-pocketed investors including OpenAI, Thrive Capital, and Jeff Bezos — which is why the model tags it a producer of robot control. The exposure that matters is demand-side: it pulls on the same scarce chips, memory, and cluster capacity as the text and image labs, and its progress depends on that supply staying available. Its centrality reads modest — early, private, and one bet among many on embodied AI.