Why Bezos is backing robot ‘brain’ developer Physical Intelligence

Physical Intelligence, cofounded by a Russian-born robotics professor, has raised more than $2.1 billion in two years. Its investors include billionaire Jeff Bezos and even potential rival OpenAI. What makes the company so promising? 

What connects home-rental provider Airbnb and Physical Intelligence? The answer may be unexpected: a former industrial building on Brannan Street in San Francisco. Airbnb had its headquarters there in the early 2010s, and PI recently subleased some of that office space. The deal covers 80,964 square feet, versus around 60,000 square feet under discussion early this year. Physical Intelligence has also expanded south of San Francisco, leasing 233,500 square feet at a business complex near Google’s headquarters.

Physical Intelligence is snapping up real estate as it works to create a universal “brain” for real-world robots – a universal AI model that can control any robot to perform any task – that could require computing infrastructure and large amounts of space in which to train physical robots and test and refine models.

The company's AI models for robotics are known as vision-language-action, or VLA, models, which can perceive their environment, understand instructions, and control a machine’s actions. The latest is π0.7, unveiled in April. The company says it can transfer capabilities across robots and tasks. For example, it can enable a new robot to fold laundry, even when there is no laundry-folding data for that robot, or to learn to use a household appliance through language instructions.

The company also describes experiments conducted in real-world settings. Robots powered by its models are operating at partner companies’ facilities, performing tasks such as packing orders. Physical Intelligence had around 80 employees as of January, according to Revenue Memo.

“I’m not much of an investment person. So I can’t claim to know how to do this,” Physical Intelligence cofounder and University of California, Berkeley, robotics professor Sergey Levine said in a recent episode of the Peterman Post podcast. Levine previously worked at Google and received the Presidential Early Career Award for Scientists and Engineers (PECASE). According to Cade Metz’s book Genius Makers, Levine grew up in Moscow, where both his parents worked as engineers on the Soviet Buran space shuttle project.

Physical Intelligence was founded by Levine and six others: Karol Hausman, a former Google DeepMind researcher and Stanford lecturer; Chelsea Finn, a Stanford professor; Brian Ichter, a former Google Brain and Google DeepMind researcher; Quan Vuong, a former Google DeepMind researcher specializing in reinforcement learning and robotics; Adnan Esmail, who led engineering at Anduril after working at Tesla; and Lachy Groom, a former Stripe employee and prominent angel investor whose investments include Figma, Notion, and Ramp.

The company does not disclose its financial results. Still, we know Physical Intelligence has already raised around $2.1 billion from investors. According to Revenue Memo, its backers include OpenAI, Jeff Bezos, Alphabet’s CapitalG fund, and Thrive Capital. OpenAI has participated both directly, in funding rounds in November 2024 and November 2025, and through the OpenAI Startup Fund, in March 2024. Physical Intelligence does not disclose the exact amount invested by each backer. Bezos has participated in two rounds. He was one of the lead investors in the $400 million Series A in November 2024 and later joined the $600 million Series B in November 2025.

In July, financial data firm Forge included Physical Intelligence among the 10 most in-demand private companies of the second quarter and listed its valuation at $11.19 billion. Private-equity platform AXEVIL estimates that Physical Intelligence may now be worth $633.50 a share. Nasdaq Private Market, by contrast, valued the company at $500.59 per share as of September 16.

The key risk with Physical Intelligence is that it must turn its pilot projects and research into full-fledged paying contracts relatively quickly, argues a Startup Diligence report. A lack of commercial revenue by the end of 2026, without a compelling customer portfolio, could become a breaking point for its investment appeal, the report states.

Startup Diligence, which specializes in private-market research, also cites competition from Skild AI and Google DeepMind, as well as Physical Intelligence’s dependence on Google technology: the architecture of the company’s first VLA model uses PaliGemma, a base model developed by Google DeepMind.

The report also points to the company’s dependence on individual key employees. The departure of Sergey Levine or Chelsea Finn before commercial launch could undermine its investment appeal.

At the same time, the team is identified as one of Physical Intelligence’s strengths. Its researchers come from different fields, giving the company the expertise to train models across different types of robots. That could create a data flywheel: the more real-world robots use the model, the more data the company receives to train robots further.

Levine previously said that robotics is still searching for a technology that can be scaled reliably. With the large language models used in AI, increasing computing power, data volume, and model size already produces reasonably predictable improvements in capabilities. No such relationship has yet been established in robotics.

Neither company representatives nor Levine responded to questions from Oninvest about plans to raise additional capital or go public.