Physical AI and robotics companies have raised close to $48 billion year to date in 2026, according to PitchBook. That is a figure in the same order as the entire generative AI funding wave of a few years ago, for a category that has shipped comparatively little.
What the money is actually buying is data. Companies in this space are collecting movement and touch data at scale, because that is the input nobody has enough of. NEURA Robotics plans five data-collection gyms by the end of 2026 - facilities whose output is not robots but recordings of bodies doing things.
That reframes the whole sector. Language models had the internet; robotics has no equivalent corpus, and no amount of compute substitutes for never having seen a hand turn a doorknob. The bottleneck is not actuators or chips. It is that the training data has to be physically generated, one motion at a time, which is why a robotics company now looks more like a recording studio than a factory. Set it against Sam Altman's framing on OpenAI's own humanoid - that the brain is harder than the body - and the investment pattern makes sense.