In the span of a single week, two separate robotics startups have entered funding discussions that together imply a combined valuation of around $22 billion. Physical Intelligence, the two-year-old San Francisco company whose founders describe their goal as "ChatGPT, but for robots," is in talks to raise approximately $1 billion at a valuation exceeding $11 billion. Separately, a robotics startup founded by former Google DeepMind researchers is reportedly in discussions for funding at a similar valuation. Neither company has shipped a product at commercial scale. Both are now worth more than most established manufacturers that actually build things.
Physical Intelligence's trajectory captures the speed of this particular moment. The company raised a $600 million round last year at a $5.6 billion valuation. If the new round closes as reported, its valuation will have roughly doubled in under four months. Founders Fund is said to be leading the deal, with Lightspeed Venture Partners also in talks. The founding thesis is that robots are currently limited not by their hardware but by their software: give them a general-purpose AI model that can transfer learned dexterity across tasks, and you get machines that can fold laundry in one factory and assemble electronics in another without separate bespoke training.
This framing, "foundation models for robotics," has become the organising idea for a wave of investment that extends well beyond these two companies. Skild AI hit a $14 billion valuation in January. Rhoda, an AI robotics startup, raised at $1.7 billion in March. Agile Robots this week announced a strategic partnership with Google DeepMind to integrate the Gemini Robotics foundation models into its industrial bots. The pattern is consistent: investors are betting that the same breakthrough that made large language models useful across text tasks will now do the same for physical manipulation.
The comparison to the LLM moment is instructive, but it is also the main source of uncertainty. Language models benefited from an almost inexhaustible supply of training data: the entire written output of the internet, ready to scrape. Physical manipulation data is scarce and expensive to collect. You cannot download a billion labelled examples of a robot folding a shirt. Companies like Physical Intelligence are spending heavily on robot farms to generate their own data, which is part of what makes their capital requirements so large, and part of what makes the moat argument more plausible than it might look at first glance. The company with the best proprietary dexterity dataset has a durable advantage over anyone who arrives later.
What remains genuinely uncertain is timing. The labour markets most vulnerable to robotic displacement, warehouse fulfilment, basic assembly, food preparation, are also the ones where humans remain stubbornly cheap and adaptable by comparison to what robots currently cost per unit. The optimistic case is that foundation models collapse that cost curve within a few years, the way transformer scaling collapsed the cost of natural language understanding. The pessimistic case is that physical interaction turns out to be harder in ways that do not yield to scale: that manipulation in messy, unstructured real-world environments is not simply "more training data" but requires architectural advances that have not arrived yet.
For now, the funding momentum is itself the story. Investors who missed the LLM wave are not willing to wait for certainty. And the companies raising at these valuations are not being shy about the scope of their ambitions: the target is not a specialised robot that does one thing well in one controlled environment, but a general-purpose robotic agent that can learn new tasks from a handful of demonstrations. Whether the hardware of 2026 is actually ready for that software is the open question these valuations are betting on.