From the street, you could walk right past Physical Intelligence’s San Francisco headquarters and never know it’s there. No neon sign. No startup bravado splashed across glass walls. Just a pi symbol on a door, slightly off-color, like an inside joke meant only for those who already belong.
Step inside, though, and the place hums.
It’s a raw concrete box softened by long blond-wood tables that feel part cafeteria, part lab bench. On one table: half-crushed Girl Scout cookie boxes, jars of Vegemite betraying Australian roots, and condiment baskets that have clearly exceeded their original mandate.
On the others: robotic arms, tangled wires, spare motors, and monitors glowing with experiments mid-thought. This isn’t a showroom. It’s a workshop in the truest sense.
One robotic arm is struggling to fold a pair of black pants. It’s failing, but with conviction. Another is locked in combat with a shirt, trying to turn it inside out. A third seems to have found peace in life, efficiently peeling a zucchini and dropping the shavings into a container. Progress is uneven, but visible.
“Think of it like ChatGPT, but for robots,” Sergey Levine tells me, waving toward the mechanical chaos. Levine, an associate professor at UC Berkeley and co-founder of Physical Intelligence, has the calm patience of someone who’s spent years translating hard ideas into human language.
Teaching Robots Without Teaching Each Robot
What’s happening here is part of a loop. Data is collected from robot stations in warehouses, kitchens, and homes. That data feeds general-purpose robotic foundation models. New models get trained, then sent back into the world to be tested again. Pants-folding, shirt-turning, zucchini-peeling — all of it becomes feedback.
The goal isn’t to master pants. It’s to understand motion, force, dexterity, and generalization. If a robot can peel a zucchini today, can it peel an apple tomorrow? A potato it’s never seen? That’s the bet.
Physical Intelligence even runs a test kitchen, stocked with off-the-shelf hardware and a suspiciously fancy espresso machine. No, it’s not for the engineers. It’s for the robots. Foamed lattes are training data, not employee perks.
The hardware itself is aggressively unsexy. The robotic arms cost about $3,500 each, and Levine says that price includes “an enormous markup.” In-house, the material cost would be under $1,000. A few years ago, most roboticists wouldn’t have expected much from gear like this. That’s precisely the point. Intelligence, if good enough, can compensate for mediocre hardware.
The Founder Who Didn’t Want to Be an Investor
As Levine slips away, Lachy Groom appears, moving fast, scanning the room like someone mentally juggling a dozen tasks. Groom is 31, Australian, and has that Silicon Valley “boy wonder” glow that follows people who’ve been improbably successful very early.
He sold his first company nine months after starting it — at age 13. After Stripe, where he was an early employee, Groom spent about five years angel investing. Figma. Notion. Ramp. Lattice. A ridiculous hit list. But he insists investing was never the endgame.
“I was looking for five years for the company to go start post-Stripe,” he says. “Good ideas at a good time with a good team — that’s extremely rare.”
Robotics pulled him back. His first bet, Standard Bots in 2021, reignited a childhood obsession that began with Lego Mindstorms. Following the academic work of Levine and Chelsea Finn — now running a robotics lab at Stanford — Groom sensed something converging. When he heard they might be starting a company, he tracked down Karol Hausman, a Google DeepMind researcher tied to the effort.
“It was just one of those meetings,” Groom says. “You walk out and it’s like, This is it.”
A Billion Dollars, No Timeline
Physical Intelligence is just two years old. It has already raised over $1 billion and carries a valuation north of $5.6 billion, backed by firms like Khosla Ventures, Sequoia Capital, and Thrive Capital. What’s unusual isn’t the money — it’s what Groom doesn’t promise in return.
“I don’t give investors answers on commercialization,” he says, almost casually. No roadmap to revenue. No deadline for monetization. That’s… not normal.
Most of the burn goes toward compute. And Groom is candid: with the right partners and terms, he’d raise more. “There’s no limit to how much money we can really put to work,” he says. Compute scales ambition.
This tolerance won’t last forever. Groom knows that. Which is why the company is capitalized now, while belief still outweighs impatience.
Any Robot, Any Task
Quan Vuong, another co-founder and DeepMind alum, frames the strategy more clinically. It’s about cross-embodiment learning. Train intelligence once, deploy it everywhere.
If a new robot platform appears tomorrow, Physical Intelligence doesn’t want to start from zero. The model should transfer. Knowledge shouldn’t care about form factor. “The marginal cost of onboarding autonomy to a new robot platform,” Vuong says, “it’s just a lot lower.”
The company is already working quietly with partners across logistics, grocery, and even a chocolate maker across the street. In some cases, Vuong claims, the systems are already good enough for real-world automation.
A Philosophical Split in Robotics
They’re not alone in chasing this. The race for general-purpose robotic intelligence is accelerating fast. Pittsburgh-based Skild AI raised $1.4 billion this month at a $14 billion valuation. Unlike Physical Intelligence, Skild is already commercial. Its “Skild Brain” generated $30 million in revenue last year across warehouses, security, and manufacturing.
Skild has also taken public shots at competitors, arguing that many “robotics foundation models” are just vision-language models dressed up as physical intelligence. Real understanding, they argue, comes from physics-based simulation and embodied data — not internet-scale pretraining.
It’s a real divide. Skild believes deployment creates the data flywheel. Physical Intelligence believes patience creates better intelligence. Someone will be right. It will take years to know who.
Slow Growth, Hard Problems
Physical Intelligence has about 80 employees and plans to grow carefully. Groom says the hardest part is hardware. It breaks. It ships late. Safety constraints complicate everything. This isn’t software. There’s no undo button for physics.
Still, the internal clarity is striking. “A researcher has a need,” Groom says. “We go collect data. New hardware? We get it. It’s not externally driven.”
Their original 5–10 year roadmap? Gone in 18 months.
As Groom rushes off, the robots keep practicing. The pants remain imperfectly folded. The shirt resists inversion. The zucchini peeler, at least, keeps winning.
The Bet Beneath the Noise
There are fair questions. Do people actually want robots in their kitchens? What about safety? Pets? Trust? And the bigger one: does solving general intelligence solve problems people care about, or just create impressive demos?
For now, Physical Intelligence is betting that timing, talent, and patience will matter more than early revenue. Silicon Valley has made that bet before. Often it fails. Sometimes it changes everything.
FAQs
1. What does Physical Intelligence do?
Physical Intelligence builds general-purpose robotic foundation models that allow robots to perform many tasks across different environments without retraining from scratch.
2. How is Physical Intelligence different from other robotics startups?
Unlike many competitors, it prioritizes research over near-term commercialization, focusing on transferable intelligence rather than specific products.
3. Who founded Physical Intelligence?
The company was co-founded by Sergey Levine, Chelsea Finn, Karol Hausman, Quan Vuong, and Lachy Groom.
4. How much funding has Physical Intelligence raised?
It has raised over $1 billion and is valued at approximately $5.6 billion.
5. When will Physical Intelligence release commercial products?
The company has not committed to a public commercialization timeline.
