Physical Intelligence in Talks for at Least $1 Billion in New Funding, Valuation Could Reach $20 Billion

QbitAI ROBO | WeChat Official Account AI4ROBO

How do you achieve a net worth of $20 billion without earning a single cent?

A robotics startup founded only in 2024 is creating a new capital myth.

The latest news: the American robotics AI company Physical Intelligence (PI for short) is negotiating a new round of financing, planning to raise at least $1 billion, with a post-money valuation expected to reach about $20 billion, equivalent to roughly RMB 133.3 billion.

Last November, this company's valuation was only $5.6 billion; in less than a year, the valuation may rise to 3.6 times its original level.

You should know that PI has so far generated no commercial revenue at all, nor does it have any robot products of its own in mass production.

But well-known investors including Bezos and Google's CapitalG have already placed their bets one after another. The reason lies in what PI develops—

This company, regarded by investors as "the OpenAI of the embodied brain field," is now pushing the valuation of this track to a new height.

First, let's look at PI's breakneck speed in the capital market.

According to Bloomberg, people familiar with the matter said that Physical Intelligence is negotiating a new round of financing of at least $1 billion, and existing investors Avenir and Lux Capital are considering leading the round.

If the deal can be completed smoothly, PI's valuation is expected to reach about $20 billion, nearly doubling from $11.5 billion earlier this year.

Looking further back, when the company secured $600 million in financing last November, its valuation was still $5.6 billion, nearly 2.6 times higher in a year.

Its list of investors is packed with top names. In addition to Avenir and Lux Capital, which are negotiating to lead this round, there are also a host of well-known investors such as Alphabet's CapitalG, Thrive Capital, and Sequoia Capital—even Bezos dug into his own pocket early on to bet on it.

For reference, the neighboring unicorn Skild AI, which also works on general-purpose robot brains, was valued at $14 billion in January this year. And by September, Skild AI had already announced that its annualized recurring revenue (ARR) had officially surpassed $100 million.

And PI?

So far, it has not generated any commercial revenue.

Moreover, the company's co-founder revealed this year that the company's main expenses are concentrated on computing resources, and it has not promised investors a specific commercialization timetable—everything said between the lines is that it is in no hurry at all to make money.

No way? With absolutely no commercialization revenue, capital has still given it a higher valuation premium... Incredible. Have these usually shrewd old-money and VC types collectively lost their minds?

You might think this is an outright bubble. But if you look at PI's founding team roster, you can understand why top venture capital dares to place heavy bets without looking at revenue.

This company was founded in 2024 and is based in San Francisco, USA. Its core founding team comes from Google DeepMind, Stanford University, and UC Berkeley, making it nothing short of an all-star lineup at the ceiling of the global robotics academic and industrial worlds:

Co-founder and CEO Karol Hausman previously spent many years at Google and Google DeepMind, was a core scientist responsible for frontier robotics research, and is also an adjunct professor at Stanford University, having long led Google's frontier exploration in robotics.

Co-founder Sergey Levine is a professor at UC Berkeley and a grandmaster-level figure in robot learning and reinforcement learning. Many students studying embodied intelligence have probably read his papers.

And co-founder Chelsea Finn, an assistant professor at Stanford University, a leading scholar in meta-learning and robotics (and formerly Levine's PhD student).

There is also another co-founder, the aforementioned Lachy Groom, who founded his first company at the age of 13, later became an early Stripe employee, and then transitioned into a well-known Silicon Valley angel investor, having invested in famous tech companies such as Figma, Notion, and Ramp (he also has considerable ties with OpenAI's CEO).

At the same time, PI has also brought together former Stripe executives and a group of talents including Brian Ichter, a core technical backbone from DeepMind.

It can be said that a group of top scholars in the field of robotics research have come together with investors well versed in Silicon Valley entrepreneurship and capital operations, winning this company enough attention.

But team glamour alone is probably not enough to support a $20 billion valuation. From another dimension, the large-model industry has currently reached a relatively special stage:

In the eyes of top investors, writing PI a sky-high check now is buying an absolute ticket to the foundation models of the physical world.

So then, why is it PI, which has not yet commercialized, that supports today's top-tier premium?

In fact, this group of top scholars never intended from day one to build robot hardware themselves.

In their view, traditional robotics R&D has long been trapped in silos. Those building robotic arms write robotic arm control code, those building bipedal robots build dedicated gait algorithms, and changing the hardware base means tearing down and rewriting the entire control system.

So the team established a native, pure-software positioning from the very start—building a general-purpose robot foundation model capable of adapting to all kinds of third-party hardware.

Their core technical path is to introduce action-generation capabilities on top of vision-language models (VLMs), training vision-language-action models (VLA) to further translate perception and understanding of the real environment into actions that robots can execute.

For example, a robot observes its surroundings through a camera, identifies the position and shape of objects and the current task, and then the model generates the corresponding motion commands to control the robotic arm's position, joint movement, and grasping actions.

From understanding the world to actually manipulating it, PI aims to bridge exactly this technical gap in between.

Co-founder Levine, in a previous interview, also described the company's goal as a "ChatGPT for robots."

However, bringing AI from the digital world into the physical world sharply raises the difficulty. To this end, since its founding, PI has continuously iterated its robot foundation model, progressing from π0 and π0.5 all the way to this year's π0.7, continually improving the model's generalization ability across tasks and devices.

π0.7, released in April this year, demonstrated a rather interesting advance.

Researchers had a dual-arm robot that had previously received no specialized training in clothes folding use knowledge the model learned from other robots to complete a clothes-folding operation.

Note that the robotic arms used by this robot differ markedly from those of the robots that previously collected clothes-folding training data in size, structure, and manner of movement.

But π0.7 was already able to recombine experience from different sources to complete new operations. It can also, under language guidance, use kitchen appliances it had not been specifically trained on and complete a variety of fine manipulation tasks.

Among these, the most important advance is that robots have begun to possess a certain combinatorial generalization ability.

Simply put, in the past, when robots faced new tasks, they often needed to recollect data and conduct specialized training; but π0.7 can already, in some scenarios, recombine skills learned in the past to solve new problems.

This capability is precisely an important foundation for a general-purpose robot foundation model.

Going further, PI's technology has already begun to move out of the lab and into real production scenarios.

Since the beginning of this year, PI has successively published research results on directions such as long- and short-term memory for robots and reinforcement learning for fine manipulation.

Just this month, PI took another important step, launching a robot foundation model API beta, beginning to provide model fine-tuning and inference services to more external developers.

It was precisely this move that fully ignited investors' sentiment.

No matter how capable a single startup team is, it is impossible for it to independently bear the data collection and engineering tuning for all physical scenarios across the entire industry.

Opening up the developer ecosystem means they have begun to move from closed demos in the lab toward a truly scalable developer co-creation system.

And looking at the entire track, the valuation race in embodied intelligence has already entered a white-hot stage.

Figure AI's valuation has surged to $39 billion, Skild AI has reached $14 billion, and Meta, Apple, and DeepMind are all massing heavy forces, not to mention Tesla and OpenAI, which are eyeing the front line with hostility.

PI is actually taking a completely different pole. It avoids the mire of asset-heavy hardware manufacturing and single-mindedly aims to build underlying software that empowers all third-party hardware.

This direction sounds incredibly sexy, but frankly speaking, the difficulty and cruelty of the physical world far exceed imagination.

In the digital world, when a model generates a piece of erroneous code, there is usually still a chance to check it, modify it, and re-execute it; but in the physical world, if a robot misjudges distance or the amount of force applied, and the robotic arm comes down, it could mean a shattered porcelain plate, damaged equipment, or even a production safety accident.

Even PI itself, when opening the public beta, poured a rare bucket of cold water on the industry. They publicly warned:

Robotic applications in real environments are by no means easy. Whether it is construction, robotic surgery, or warehouse automation, conquering each field requires pouring in the resources of an entire company.

It does sound a bit piercing, but it is absolutely the truth.

Behind the physical deployment of every segmented field are extremely complex dynamics laws and long, painful engineering validation.

For a long time, the development of the robotics industry has relied heavily on specialized development for different devices and different tasks. The progress of general foundation models is now offering the entire industry another possibility:

Let robots' intelligence gradually become independent of specific hardware and become a foundational service that different companies can use.

Only when intelligence can be continuously reused like software can the robotics industry truly break out of the old model of custom-building each unit and developing each item one by one, and usher in an era of large-scale application.