The money pouring into "physical AI," the effort to give artificial intelligence a body that can act in the real world, has never been louder, and this week the industry's own analysts spent most of their time explaining why it is harder than it looks. At Citi's Robotics and Physical AI conference, which wrapped on Tuesday, founders and investors kept circling back to a single, unglamorous constraint: data. Not chips, not clever models, but the raw record of robots doing real tasks in real places. Roughly 20 billion dollars has flowed into physical AI over the past two years, and the binding limit is something you cannot simply buy.
The contrast with the chatbot boom is the point. Large language models were trained on an internet's worth of text that already existed. A robot that needs to fold laundry or inspect a phone on a production line has no such archive; every hour of useful training data has to be generated by machines actually doing the work. One workforce firm at the conference estimated that even the tens of millions of hours being collected in 2026 amount to "basis points, not percentage points" of what high-level robotic performance will finally demand. That is why Citi's analysts concluded the winners will be whoever owns proprietary, task-specific data, solves a narrow labor problem, and rents robots out as a service to lower the cost of adoption.
You can already see that logic in who is making money. RoboSense, a Chinese lidar maker, shipped more than 719,000 sensor units in the first half of the year, with its robotics business up 510 percent, because it sells the perception hardware nearly every robot builder needs regardless of which one eventually wins. It is the classic picks-and-shovels position, the same one Nvidia occupies with its chips, simulation software, and robot foundation models. When the ultimate prize is uncertain, the surest profits go to whoever equips all of the contestants.
Meanwhile the demonstrations keep coming, and they are genuinely improving. AgiBot, a Chinese firm, livestreamed eight humanoids working a real tablet production line for more than 64 hours, reporting a 99.99 percent task success rate. Startups keep raising, too: Zeroth, which builds home robots, just pulled in 74 million dollars led by Ant Group on the back of 600 percent revenue growth. But a livestreamed factory run in a known, tightly bounded environment is exactly the kind of "closed" setting analysts say will commercialize first. The messy, open world of homes and streets is a far longer road.
The honest read is that physical AI is real and slow at the same time. It is not a bubble in the sense of having no substance; closed environments like warehouses and factories are already producing measurable returns. But it is not the next overnight ChatGPT either. The value looks set to accrue across a decade, first to the sellers of compute and sensors, then to specialists who crack one expensive labor problem, and only much later to the general-purpose humanoids that hog the headlines. The tell worth watching is not how smoothly a robot moves at a launch event. It is whether the data it gathers afterward compounds into something a rival cannot copy.