Physical Intelligence just crossed $1 billion in funding with a $5.6 billion valuation, and co-founder Lachy Groom still won't tell his backers when they'll see a return. The two-year-old startup, backed by Khosla Ventures, Sequoia Capital, and Thrive Capital, is building foundation models for robots - think ChatGPT but for mechanical arms learning to fold laundry and peel vegetables. While competitor Skild AI just hit a $14 billion valuation by focusing on commercial deployment and revenue, Groom is betting that pure research wins long-term.
Inside a nondescript San Francisco building marked only by a pi symbol, robotic arms are fumbling through the mundane. One struggles to fold black pants. Another tries turning a shirt inside out with mechanical determination. A third peels a zucchini with surprising competence, depositing shavings into a container. This is where Physical Intelligence is building what it hopes becomes the brain behind every robot.
"Think of it like ChatGPT, but for robots," Sergey Levine, UC Berkeley associate professor and co-founder, explains while gesturing toward the mechanical ballet. The setup looks deliberately unglamorous - these robotic arms cost about $3,500 each, with material costs under $1,000 if manufactured in-house. A few years ago, roboticists would've laughed at the idea these cheap arms could do anything useful. But that's precisely the point, according to TechCrunch's exclusive look inside the startup.
Data collected from robot stations here and in warehouses, homes, and test kitchens trains general-purpose foundation models. When researchers build a new model, it returns to these stations for evaluation. The sophistication lies not in the hardware but in the intelligence compensating for it. Even the espresso machine nearby isn't a staff perk - it's there for robots to learn from.
