XPeng CEO He Xiaopeng and XPeng s humanoid AI robot IRON visit XPeng booth during the 21st Shanghai International Automobile Industry Exhibition
Artificial intelligence first became a software boom. Then more and more capital poured into chips, data centers and power supply. Now the investment story is moving further into the physical world. Xpeng is raising more than $900 million for its robotics division in its first external funding round, giving the business a valuation of more than $6.3 billion. Figure AI is already valued at $39 billion, Apptronik at more than $5 billion. And within just a few days on the stock market, Unitree showed how quickly enthusiasm can turn into speculation. Behind those billions is a term investors are likely to hear much more often in the coming years: Physical AI.
Xpeng Gets Billions for a Business That Is Only Just Starting
The latest trigger comes from China.
Electric-vehicle maker Xpeng has raised more than $900 million for its robotics subsidiary Dogotix. The first external financing round values the business at more than $6.3 billion after the transaction. The round is led by IDG Capital, with Tencent and Alibaba also among the strategic investors. According to Xpeng, it is the largest single private funding round to date in China’s so-called Embodied AI sector.
The money will go into hardware and software, training proprietary Physical AI models, data collection, production facilities and international expansion.
Xpeng plans to begin mass production of its humanoid robot IRON by the end of 2026. Initial deployments are expected in its own stores and on company premises. Deliveries to external customers in China and abroad are scheduled to begin in 2027.
In other words, there is no broadly established robotics business yet. The multibillion-dollar valuation already exists.
The Carmaker Itself Is Posting Higher Losses Again
The fact that investors are nevertheless assigning such a high value to the robotics business becomes even more interesting when viewed against Xpeng’s own financial results.
The group generated revenue of 19.74 billion yuan in the second quarter and achieved a gross margin of 20.7 percent. At the same time, its net loss widened to 1.34 billion yuan, compared with around 480 million yuan in the same quarter last year. Research and development expenses rose by more than 32 percent to 2.91 billion yuan, partly because of new vehicle models and AI technologies.
Capital markets are therefore valuing two different time horizons at once.
The existing automotive business needs to deliver better results today.
The robotics division is receiving capital for a market whose major returns are expected tomorrow.
Goldman Sachs analysts pointed out that the robotics subsidiary’s $6.3 billion valuation already represented around 53 percent of Xpeng’s total market capitalization as of August 21. Morgan Stanley described the valuation as high compared with many other privately held humanoid robotics companies.
A development project has therefore suddenly become a significant part of the entire group’s equity story.
Physical AI Is Bigger Than the Humanoid Robot
The term should not be equated with humanoids alone.
Physical AI describes AI systems that do more than process information and generate digital outputs. They perceive their physical environment, draw conclusions from it and act within it.
That includes humanoid robots, but also autonomous vehicles, mobile industrial robots, drones and other independently operating machines. Nvidia describes Physical AI as systems capable of perceiving, understanding, reasoning and performing complex actions in the real world.
This is what distinguishes it from the first major wave of generative AI.
A language model has to understand which words meaningfully follow one another.
A robot also has to understand that a cup can break, that a person may cross its path or that an object may behave differently from the way it did in a training simulation.
The real world is less predictable than a chat window. That raises the requirements for sensors, data, models, computing power and hardware all at the same time.
Why Carmakers Are Getting Involved
Xpeng’s presence in this field is less surprising than it may initially seem.
A modern autonomous vehicle is itself a form of Physical AI. It perceives its surroundings through cameras and sensors, processes enormous amounts of data, predicts the movements of other road users and then controls a physical system. Many of those building blocks can also be transferred to robotics.
Xpeng develops its own Turing AI chips and so-called World Foundation Models, while also bringing experience in batteries, drive systems, sensors, manufacturing and supply chains. The company now describes itself as a »Physical AI company« and places intelligent vehicles, robotaxis and humanoids on the same technology platform. That is corporate marketing, of course, but it also makes the strategic connection visible.
Xpeng is not alone. Tesla is developing humanoid robots with Optimus. Chery is building its own systems through AiMOGA and is even preparing an IPO for its robotics division.
The race between carmakers could therefore increasingly become a race to build autonomous machines.
$39 Billion for Figure
A look at the United States shows that capital has already discovered this market.
Figure AI raised more than $1 billion in a Series C funding round in September 2025. Its post-money valuation: $39 billion.
Its backers include Nvidia, Intel Capital, Salesforce, Qualcomm Ventures and Brookfield. Figure intends to use the capital to expand humanoid production, further develop its Helix AI system and collect larger volumes of real-world training data.
The company is already testing its robots in real industrial environments, including deployments at BMW.
Apptronik is also attracting substantial amounts of capital. The Austin-based company has now expanded its Series A funding to a total of $935 million. According to TechCrunch, its latest valuation was around $5.3 billion. Investors include Google, Mercedes-Benz and B Capital.
The capital rush is even more pronounced in China.
Unitree Showed How Quickly Future Expectations Can Turn Into Speculation
When Unitree went public, its shares surged by more than 460 percent on the first day of trading. The closing price implied a market capitalization of around $50 billion; at one point, the robotics company was valued at roughly $66 billion.
The enthusiasm did not last.
By August 25, the stock had fallen around 45 percent from its post-IPO level. Roughly $30 billion in market value disappeared again. Reuters described the move as a trigger for an increasingly open debate over whether a speculative bubble may already be forming in China’s robotics sector.
For investors, the example is particularly instructive. A technology can be highly promising and a stock can still be too expensive.
Unitree is not a company without products or revenue. Its quadruped robots are already commercially successful, and the company is one of China’s most prominent robotics manufacturers. But expectations surrounding humanoid robots extend far beyond its existing business.
The market is therefore not paying only for what Unitree already sells. It is paying for what the company might one day become.
The Market Is Pricing In the Breakthrough Before the Breakthrough Exists
Ironically, Unitree CEO Wang Xingxing himself is urging more patience when it comes to technological progress.
He expects a “ChatGPT moment” for robotics, when robots will be able to understand and perform unfamiliar tasks more independently. But he places the possible timing of that breakthrough somewhere within a range of two to ten years.
Today’s humanoids, he says, are still inferior to humans in many real-world tasks. That captures the central tension in this stock-market story.
Investors are already valuing companies in the billions even though leading figures in the industry themselves do not know when the decisive technological leap will arrive.
That does not necessarily mean those valuations are wrong. Capital markets exist to anticipate future developments. But the further expected profits lie in the future, the greater the uncertainty becomes — both about how large those profits will be and about which companies will ultimately earn them.
China Is Already Producing — but Not All Demand Comes From End Customers
At the same time, the industry is further advanced than a pure future vision.
China has built a leading industrial position in humanoid robotics in a remarkably short period of time. Reuters recently reported that Chinese manufacturers delivered more than 40,000 humanoids in the first half of 2026 alone.
But that figure also requires context.
A significant share of current demand does not yet come from factories, hotels or private households where robots are already replacing human labor economically.
The Financial Times reported on state-backed training centers in China that buy humanoid robots in order to generate real-world movement and training data with them. Some of that data is subsequently sold back to robotics companies. Nearly 370 robotics startups are said to have emerged in China within just two years.
The system can accelerate technological development. But it also makes it harder to assess how large genuinely independent commercial demand already is.
At the World Robot Conference in Beijing, the focus therefore shifted increasingly away from spectacular dances and backflips toward a much simpler question: Can a robot perform a task reliably enough to save its buyer money or generate additional revenue? That is where a sustainable business model begins.
Five Trillion Dollars — but Not Until 2050
The potential size of the market helps explain why investors are willing to deploy billions so early.
Morgan Stanley estimates that the market for humanoids, including associated supply chains, maintenance and services, could exceed $5 trillion by 2050. By then, close to one billion humanoid robots could be in use worldwide. The bank expects around 90 percent of them to operate in industrial and commercial applications.
But the same outlook contains another figure that may be at least as important for investors: Morgan Stanley does not expect adoption to accelerate significantly until the second half of the 2030s.
That leaves many years between multibillion-dollar valuations in 2026 and a potential mass market.
And plenty of room to be wrong.
The Physical AI Boom Is Not Just a Bet on Robots
For investors, the potential value creation also extends far beyond the companies that manufacture the machines.
Physical AI needs processors, sensors, memory, motors and actuators, batteries, simulation software, data centers for training and factories capable of eventually producing complex machines at very large scale.
A similar pattern could emerge to the one already seen in generative AI. The companies behind the best-known chatbots have not necessarily been the biggest immediate stock-market winners. A significant share of the value creation has gone to the manufacturers and operators of the underlying infrastructure — above all semiconductor companies. Physical AI could produce the same effect.
The eventual winner does not necessarily have to be the company whose robot looks most human today. It could be the manufacturer of the motors in its joints, the producer of its sensors, the developer of its world model or the company capable of manufacturing these machines most cheaply by the millions.
The Next AI Story Is Beginning
This is also changing the way investors think about the AI boom itself. Artificial intelligence was initially valued primarily as a software story. Then it became clear just how much infrastructure would have to be built to support it.
Now capital is increasingly moving toward the next step: AI is expected to leave the digital environment and begin acting within the physical economy itself. The potential is considerable. A system that writes a text faster can increase productivity. A machine that independently transports materials, assembles components, picks goods or drives people intervenes directly in real-world production.
Physical AI could therefore eventually open up a market significantly larger than software alone. But stock markets have an inconvenient habit of paying for a possible future long before that future arrives.
Unitree demonstrated within a matter of days how expensive that can become. Xpeng now shows the other side: investors remain willing to provide enormous amounts of capital before industrial scaling has even begun.
The technology is real — and so are the progress and the capital. The large commercial market that some valuations already anticipate still has to emerge.
Physical AI could become the next major stage of artificial intelligence.
For investors, however, that does not only mark the beginning of the next growth story. It also raises the next question: How much of the future should be paid for today?
SK
