The money arriving in humanoid robotics this month is spectacular, and slightly ahead of the machines. In a single week, a UK startup bluntly named Humanoid raised $152 million to become what it calls Europe's first pure-play humanoid unicorn, at a $1.35 billion valuation; Travis Kalanick's physical-AI venture Atoms pulled in $1.7 billion led by Andreessen Horowitz; and at AMD's Advancing AI expo, the chipmaker announced that Foundation's Phantom robots would run on its processors, opening a new front in the fight for what the industry now calls robot brains, a market one report pegs at $50 billion a year by 2035.
The awkward backdrop to all this capital is that today's humanoids mostly do not work yet, at least not usefully. TIME's profile of Unitree, the Chinese firm that ships more humanoids than anyone, contains the number that deflates the hype: only 9 percent of its sales go into industrial use, while 74 percent go to universities, research labs and individual developers, that is, to people trying to build better robots. The whole modern humanoid industry, as one analyst puts it, is essentially selling robots to people trying to build robots. Current machines run at 30 to 50 percent of human efficiency at tasks as simple as stacking boxes.
The barrier is not the body but the brain. Hardware has raced ahead: Unitree has cut its flagship G1 from $16,000 to $13,500 in eighteen months and sells a robot dog for under $2,000. The unsolved problem is what engineers call the generalisation gap. A robot trained to fold a shirt fails when the lighting changes or the shirt shifts position. Unlike a language model, which can feast on the entire internet, a robot needs manipulation data gathered painstakingly in the physical world. That is why Tesla is strapping camera backpacks onto assembly workers in Germany to record how they grip tools, and why a New York startup is cleaning apartments free of charge simply to harvest the training data.
The capital is splitting into two philosophies about how to cross that gap. One camp chases the general-purpose humanoid, a single machine that can eventually be told "go to the grocery store" and simply do it. The other, exemplified by Kalanick's Atoms, rejects the human shape entirely, betting that heavy industry, mining, construction and food production, wants purpose-built machines rather than costly robots that walk. Ben Horowitz, who is backing Atoms, argues specialised industrial machines suit tough environments far better. It is a genuine fork in the road: flexibility and general intelligence on one side, ruggedness and focus on the other.
The chip layer adds its own contest. Nvidia has spent years making itself the default for robotics, from its Jetson edge modules to its Isaac simulation stack. This week Foundation broke ranks, choosing AMD and citing its FPGAs, reprogrammable chips well suited to the deterministic, low-latency control a robot hand needs to coordinate 23 degrees of freedom. Foundation is not a demo shop: it says its robots run three shifts a day and contributed to more than 24,000 cars built last year, earning $100 million in contracted recurring revenue, which puts it on the very short list of humanoid makers with real income.
So the sector sits in a peculiar spot: real revenue at a handful of firms, real deployments beginning, and a wall of money betting the generalisation gap closes soon, all layered on top of machines that Unitree's own founder calls "strangely theatrical," bodies dressed as intelligence, rehearsing a role they have not yet learned to live. Morgan Stanley thinks 13 million humanoids could be walking among us by 2035. Whether that is foresight or froth depends entirely on a data problem that no amount of funding has so far solved.