For a couple of years the giant checks in AI went to models: foundation labs raising rounds that used to be reserved for entire industries. That is still happening, but a second front opened, and it is loud. Robotics startups have raised more money in 2026 so far than in any full year before, and the physical-AI foundation-model layer is now the most heavily funded emerging category after the frontier labs themselves. The money is moving from bits to atoms.
The clearest single data point is Neura Robotics, which raised a Series C of up to $1.4 billion this year at around a $7 billion valuation, backed by Nvidia, Amazon, Qualcomm, Bosch, and Tether, with an order pipeline already north of a billion dollars and a stated goal of millions of robots by 2030. That is not a science project. That is a company being funded like an industrial giant.
Why the capital rotated toward hardware
The logic is straightforward once you say it out loud. The software side of AI is commoditizing fast. Models converge, prices collapse, and the wrapper businesses get cloned in a weekend, all of which I have written about. Investors chasing durable advantage started looking for places where the moat is not a prompt. Atoms are one of those places. You cannot fork a factory. Actuators, supply chains, real-world safety validation, and the years it takes to make a robot reliable are exactly the kind of hard, slow, physical barriers that software stopped providing.
Physical AI also has an obvious enormous market if it works: any job that involves moving through the world. The bet is that the same wave of capability that made models good at language is about to make robots good at manipulation and navigation, and whoever owns the platform layer for that captures something the size of a category, not a product.
What it means if you build software
Most of us are not going to start a humanoid company, so the honest read-through is about where software attaches to this, not about welding.
The interesting layer for builders is the same one that turned out to matter in language AI: not the hardware, the platform on top of it. Neura talks about a skill-sharing platform for its robots. That phrasing should sound familiar, because it is an app store for physical capabilities, and app stores need developers. As robots become programmable platforms, the demand for people who can build, orchestrate, and verify robot behavior in software goes up, even for people who never touch a servo. The agent patterns we have been discussing, planning, tool use, verification loops, are the same patterns a robot needs, just with a body attached.
The caution is the same one I would give about any funding wave. Record capital into a category is a statement about belief, not about revenue, and a lot of these bets will not pay off on the timeline the valuations imply. Humanoid robots have been five years away for longer than five years. The money is real and the direction is probably right, but "funded like an industrial giant" and "is an industrial giant" are different sentences, and the gap between them is where a lot of this capital will get burned.
Still, the signal is worth holding onto. When the biggest checks rotate from software toward the platform layer of physical AI, it tells you where the next decade of hard, defensible building is expected to happen. If your skills are in orchestration and verification, that layer is going to need them, body attached or not.
If you have moved from pure software into robotics or physical AI, I would like to hear what transferred and what did not, because that map is still mostly private.
Top comments (0)