Anthropic is pursuing an acquisition of Physical Intelligence, according to a single, uncorroborated account from Robert Scoble (@Scobleizer). In a July 18th post on X, Scoble said an unnamed investor told him Anthropic was buying the robot foundation-model developer, though the deal "hasn't closed yet." (x.com)
Scoble later described the source as a "well known investor," without identifying the person or providing a proposed price, transaction structure or timeline. As of July 18th, neither Anthropic's newsroom nor Physical Intelligence's website carried an acquisition announcement. The claim remains unconfirmed, and the terms are unknown. (anthropic.com)
The rumor surfaced at an open house for Dimensional, an open-source robotics software developer. Scoble described Dimensional as making an American operating system for Chinese robots. Dimensional's repository describes an agentic operating system that connects language-model agents to humanoids, quadrupeds and drones, including hardware from Chinese robot maker Unitree. The repository had about 3,400 GitHub stars when recently indexed. (github.com)
A direct route into robot control
Physical Intelligence was founded in San Francisco in 2024 by a group that includes CEO Karol Hausman and robotics researchers Sergey Levine, Chelsea Finn and Brian Ichter, alongside former Stripe employee and investor Lachy Groom. Hausman described the founding objective as building "a universal model that can power any robot" in Physical Intelligence's first public interview. (bloomberg.com)
That thesis makes Physical Intelligence a strategically coherent target for Anthropic. Anthropic's Claude models can write software, interpret visual inputs and operate computers, but Anthropic has acknowledged that its own robotics experiments did not address low-level actuation policies, the systems that translate decisions into precise physical movement. Physical Intelligence has spent two years working directly on that missing control layer. (anthropic.com)
Physical Intelligence's models are designed to operate across different robot bodies and tasks. Its open-source pi 0 release included model weights and examples for single-arm, dual-arm and mobile platforms. Physical Intelligence said developers could fine-tune the model with their own robot data, while warning that the experimental release would not work reliably on every setup. (pi.website)
Physical Intelligence has since moved beyond research demonstrations into partner deployments. In a February 24th technical post, Physical Intelligence said its models were controlling Weave laundry-folding robots and Ultra warehouse robots. Weave reported that adding its deployment data to pre-training reduced missed grasps by 42% and human interventions by 50%. Ultra reported a continuous order-packaging shift at 96.4% autonomy. Those figures came from Physical Intelligence's partners and have not been independently audited. (pi.website)
The latest published model, pi 0.7, was released on April 16th. Physical Intelligence said it could control an unfamiliar robot, respond to language-based coaching and combine previously learned capabilities when handling new tasks. The results point toward a reusable robot intelligence layer rather than software built for one machine and one workflow. (pi.website)
A highly priced target
Any acquisition would require Anthropic to absorb one of the most heavily financed independent robotics labs. Physical Intelligence raised $600 million in November 2025 at a $5.6 billion post-money valuation, with Alphabet's CapitalG leading and Lux Capital, Thrive Capital and Jeff Bezos participating, Bloomberg reported. OpenAI was also among Physical Intelligence's early backers, putting investors tied to two of Anthropic's largest model competitors on Physical Intelligence's capitalization table. (news.bloomberglaw.com)
In March 2026, Physical Intelligence was discussing another roughly $1 billion financing at a valuation above $11 billion, according to Bloomberg and TechCrunch. Founders Fund was expected to participate, while Lightspeed Venture Partners, Thrive and Lux were also involved in the talks. That financing was described as preliminary, and no completion was publicly confirmed in the reporting reviewed for this story. (bloomberg.com)
Physical Intelligence's valuation therefore sets a demanding baseline for any negotiated sale. A buyer would also be paying before Physical Intelligence has disclosed a conventional commercial model. Groom told TechCrunch earlier this year that Physical Intelligence had no commercialization timeline, even as Physical Intelligence employed about 80 people and continued spending on models, data and computing infrastructure. (techcrunch.com)
Anthropic has become more willing to buy technical teams. Anthropic acquired computer-use developer Vercept and said the transaction followed its earlier purchase of Bun. Anthropic also acquired API infrastructure developer Stainless in May. Those deals added software and engineering capabilities adjacent to Claude; Physical Intelligence would move Anthropic into the harder problem of controlling machines in unstructured environments. (anthropic.com)
Anthropic has already been testing that boundary. In June, Anthropic reported that Claude Opus 4.7 completed several robot-programming tasks substantially faster than human teams in its Project Fetch experiment, while still struggling with precise physical manipulation. Physical Intelligence's control models address the part Anthropic said Claude had not solved. That technical fit explains why an acquisition discussion would make sense. Scoble's unnamed-source post does not establish that Anthropic and Physical Intelligence will complete one. (anthropic.com)