What Is Physical AI? Why It Matters Now, the US-China Race, and Japan's Path to Win

Hello, this is Ryuta Hamamoto from TIMEWELL.

Across 2025 and 2026, the centre of the technology conversation shifted a little. On top of the talk about generative AI that writes text and draws pictures, a new phrase started showing up often in the news: "Physical AI," meaning AI that moves things, walks, and drives in the real world. The Japanese government, too, in the growth strategy it put together in the summer of 2026, placed this field at the centre and signalled a plan to invest a large sum in it.

That said, I suspect Physical AI is still an unfamiliar phrase for many readers. This article lays out, in a way you can read through with no specialist knowledge, what Physical AI is, why it has surged into attention at this particular moment, how the United States and China are moving, and where Japan is trying to compete. Rather than the fine details of how the technology works, my aim is to give executives and people on the floor a handle for thinking about "how does this touch our own work." If the distance between your company and AI is on your mind, checking your starting point first with our free AI literacy self-check will make the second half of this article far more concrete.

What Physical AI is, and why the "body" is hard

Let me start with what the words mean. Physical AI is also called embodied AI. Roughly put, it is technology in which AI, acting as a brain, perceives and acts on the real physical world through a body such as a robot or a vehicle. Generative AI like ChatGPT handles text and images strictly inside a screen. Physical AI, by contrast, takes on real-world actions too: gripping the part in front of it and fitting it into place, walking around an obstacle on the floor, driving while watching the cars around it. Robots, autonomous driving, and the base technology for training their brains all belong together in this field.

Here is the point many readers find surprising: why is moving a body so hard? If AI that handles text has become this clever, surely moving a body is the easy part. In fact it is the reverse, and moving a body in the real world is far more formidable. The reason is data. Generative AI that handles text grew clever at a startling pace by studying the near-infinite text on the internet. But data on physical movement, how firmly you must grip an object so as not to drop it, or whether this floor is slippery, barely exists in the world. The only way to gather it is to actually run robots one machine at a time, which costs enormous time and money. This scarcity of real-world data has long been called the single biggest bottleneck for Physical AI.

There is another difficulty. An AI that moves a body is not easily forgiven for failing. With text generation, if a slightly off answer comes out, you can rewrite it and move on. In the real world, if a robot misjudges its force it breaks something, and if a car misjudges a decision it can cause an accident. What is demanded is not just a clever brain but, at the same time, a reliability and safety that obey the laws of physics. This is exactly why Physical AI stayed, for many years, a challenge confined to the laboratory. That situation began to move in a big way in 2025, which is the backdrop to the attention it draws now.

Why Physical AI suddenly started moving now

So what changed? The key lies in two ideas for going around the wall of data scarcity. One is called the "world foundation model," the other the "robot foundation model," and both advanced to the practical stage together in 2025.

The world foundation model, in English the World Foundation Model, is at its core an AI that generates, inside a computer and in large volumes, video and situations that obey the physics of the real world. Instead of running a physical machine tens of thousands of times to train it, you build a physically correct virtual world and let it try an enormous number of times inside, filling in the missing data artificially. The company with a striking presence in this area is the chip giant NVIDIA, which at CES in January 2025 (one of the world's largest consumer-electronics and technology trade shows) announced the platform for a world foundation model called "Cosmos."¹ Around that time, the company's CEO Jensen Huang said that "robotics is having its ChatGPT moment." The idea is that the same turning point that once made text AI spread all at once is now coming to robots. Cosmos was later extended in March 2025 into variants for specific uses, adding functions to predict video and to reason about the results of actions.² In June 2026 a next-generation version was formally announced, and NVIDIA describes it as a model that can handle text, image, video, sound, and action while keeping physical consistency.³ I would add that expressions like "world's first" are NVIDIA's own descriptions, so it is wise to take them with a grain of salt.

The other one, the robot foundation model, is a brain that connects sight, language, and action. It reproduces in AI the same flow a person follows, seeing, understanding an instruction, and moving the hands, and NVIDIA released a foundation model for humanoids in March 2025.⁴ A distinctive feature is its two-layer structure: a fast, reflex-like layer and a slower, deliberate thinking layer. And the small, high-performance computers needed to load such a trained brain onto the robot body have come together too. The robot-oriented hardware NVIDIA began offering generally in August 2025 is described as greatly raising AI performance over the previous generation.⁵ Training in the cloud, trials in simulation, execution on the computer mounted in the robot. With this three-part setup in place, Physical AI finally started to move from research toward deployment. Researchers, too, have said one after another that intelligence which understands space is the next main battleground.

A three-way contest: how the US and China are moving

The competition around Physical AI is now unfolding across three big poles. One is the platform layer, like the NVIDIA we just saw, which tries to hold the very foundation of development. Rather than finishing robots itself, it supplies world foundation models and computers to many companies and tries to become the base for the whole industry. The remaining two are the American startups trying to put humanoid robots into the world, and China, pushing with the power of mass production.

In Silicon Valley, the development of human-shaped robots, so-called humanoids, has heated up. One leading company announced its third-generation machine in October 2025, saying it features a parts configuration built with mass production in mind and hands equipped with tactile sensors for gripping objects. According to that company's own announcement, its previous-generation machine was deployed at a German automaker's plant and, over about ten months of operation, was involved in producing a cumulative total of more than thirty thousand cars. The automaker, for its part, is advancing a trial deployment of humanoid robots at another plant in Germany and has announced it will set up a dedicated base for Physical AI in manufacturing. Another major electric-vehicle company is reported to be planning small-scale production of its own humanoid in 2026. That said, figures such as the number of joints in the hand, the target price, and the timing of production start are numbers from reporting, and that company's own leader has reportedly said, in effect, that it "has not yet reached the stage of doing genuinely useful work in a factory." It is more accurate not to take these as settled facts but to see the work as still in progress.

China, on the other hand, is competing in a completely different way. Its hallmark is growing the field with the whole nation behind it, a whole-of-nation approach. The relevant department of the Chinese government set out a policy for humanoid development at the end of 2023, calling for mass production by 2025 and, by 2027, for reaching a world-leading level and making it a new engine of economic growth. It is also advancing industry standard-setting. Its strengths are the mass-production capacity and supply chain built up in electric vehicles, and low parts costs. The move underway is to carry all of that straight across to robots, in effect running the winning EV playbook on humanoid robots. Some industry surveys and reporting say that many of the humanoid robots sold worldwide in 2025 were made in China, but these are estimates, and the figures for unit numbers and the like carry a wide range. It is certainly true that the depth is increasing, with Chinese startups putting out machines at relatively affordable price bands and major EV makers officially acknowledging their own humanoid development. Three poles, each bringing its own strengths of brain and body, of mass production and capital, have all started running at the same time. That is where we stand.

Some readers may feel these torrents in the wider world are events far removed from running their own business. Yet what is really being tested there is less the advanced robotics itself than the implementation skill of putting AI to work in the tasks of the floor. What we work on with executives through our AI consulting service WARP is exactly this part: pinning down where in your own company AI actually pays off. Not treating the world's movements as someone else's affair, but translating them into your own next step, is what I think will separate the outcomes from here.

Japan's strategy: what 17 fields and 10.5 trillion yen mean

So where is Japan trying to compete? In the growth strategy it assembled in the summer of 2026, the government made a large push for public-and-private investment into "17 strategic fields" including AI and semiconductors. According to reporting and the government's meeting materials, the public-and-private investment goal for these 17 fields is put at more than 370 trillion yen by fiscal 2040, and Physical AI sits at the head of them. And within Physical AI, in the field of AI robots especially, the plan reported is to invest 10.5 trillion yen of combined public and private money. It expects the market for multi-purpose robots such as AI robots to reach roughly 60 trillion yen by 2040, in a structure aimed at easing the structural labour shortage and lifting the country's supply capacity. These figures rest on a draft investment roadmap presented at a government meeting in June 2026, and it is worth also keeping in mind that the fine breakdowns and the wording of definitions are the kind of thing to confirm finally against the government's own primary materials.⁶

The important thing here is why Japan is betting on this field, the strength at the root of that calculation. Japan has companies that stand at the top of the world in the component technologies on the body side of Physical AI. Japanese makers hold a high world share in the industrial robots used in factories; and in machine tools, in motors, in the reduction gears that produce smooth motion in a robot's joints and other drive parts, and in the sensors that accurately measure the position and shape of objects, Japanese technology is trusted worldwide. Even if overseas players hold the robot's brain, if you cannot build the body that brain inhabits, the robot cannot take a single step. On top of that, Japan holds an accumulation of operational know-how, keeping machines running without stopping on production floors that demand exacting quality. The head of one major robot maker has spoken, in effect, of Japan's strength lying precisely in its power to implement things in society and in its hardware.

Still, there is no looking away from the weak points. First, there are not yet many companies in Japan that make finished humanoid robots themselves. Second, in developing the foundation models that serve as the brain and in securing the large-scale computing resources, the so-called GPUs, needed to train them, Japan lags the United States and China. Third, there is the scale of capital. Against leading US and Chinese startups that raise tens to hundreds of billions of yen and concentrate investment into computing resources, the amounts Japan puts in look modest as things stand. Even so, movements have started to appear at home: a major automaker advancing AI research and the deployment of outside robots, a heavy-industry maker developing its own humanoid, and a telecom-and-investment giant accelerating its Physical AI strategy with a large acquisition of a leading overseas robotics business, among others. Precisely because the strengths and weaknesses are so clear, deciding where to compete comes to matter decisively.

Japan's path to win: borrow the brain, win on the body and the floor

With all this in mind, how should we think about Japan's path to win? My own view is the pragmatic one: rather than overreaching to assemble the entire brain in-house, capture value reliably in the positions where Japan is already number one in the world, the body's component technologies and the implementation skill on the floor. Squaring up to the US and China head-on in the competition over foundation models and computing resources is not easy given the scale of capital. But the body that carries that foundation model, the hardware that does not break, that is precise, and that keeps running for a long time, is not so easily imitated. Take in the world's best where the brain is concerned, while competing on the body and on the skill to use it well on the floor. This is not admitting defeat; it is the idea of going for the win with your own strongest card.

And this idea is not only a matter for large corporations. For mid-sized and small manufacturers, the implication sits very close at hand. The essence of Physical AI lies not only in the flashy robot part, but in turning the work and judgements of the floor into data and reshaping them into a form AI can inherit. The knack a skilled worker reads off a drawing, the sequencing grasped through long experience, if such tacit know-how can be arranged into a form AI can handle, it becomes a weapon for everyday design, estimating, and floor improvement, well before it is ever loaded onto a robot. The reason we are honing our enterprise AI product ZEROCK for design and floor-knowledge use in manufacturing comes from exactly this idea of turning the wisdom of the floor into a foundation for AI. Placing servers within the country and putting a company's internal knowledge into a form it can handle safely is, I believe, the footing for a Japan that wins on the body.

One thing not to overlook is the lens of economic security. Technologies such as robots, semiconductors, and drive parts are important fields tied directly to national competitiveness, and they cannot be separated from practical work like export control and checking counterparties. The more seriously you step into Physical AI, the more the question of how to manage the flow of technology and parts takes on real weight. For the overall shape of the policy and the angle of implementation in the regions, our explainer on the regional growth strategy decided alongside the Basic Policies is worth a look as well. For Japan to hold its ground and go on the offensive amid the world's torrents, a view that considers technology, the floor, and the rules together as one is indispensable.

Summary

This ran long, so let me organise the key points.

• Physical AI is technology in which AI, acting as a brain, perceives and acts on the real world through a body such as a robot or a vehicle. It is also called embodied AI
• The biggest wall was that data on real-world movement is scarce. Efforts to get over that wall gained real momentum in 2025, using world foundation models that create synthetic data in a virtual world and robot foundation models that connect sight, language, and action
• The competition is unfolding across three poles: platforms, Silicon Valley's humanoids, and China pushing with mass production. Some of the humanoid figures are at the reporting stage, and it is best to discount them given that the work is still in progress
• Japan places Physical AI as a priority within the 17 strategic fields of its growth strategy and is reported to plan 10.5 trillion yen of combined public and private investment. It is strong in body-side component technologies such as industrial robots, reduction gears, and sensors, while lagging the US and China in foundation models, computing resources, and capital scale
• The realistic path to win, I believe, is not to hold the entire brain in-house but to capture value with world-class body technology and floor-level implementation skill

Physical AI is not a story of some distant future; it is drawing steadily closer to the manufacturing floor. Before your eyes are pulled to the flashy robots, start by arranging the wisdom asleep on your own floor into a form AI can handle. That, I think, is the surest first step for a Japanese company that wins on the body. If you are unsure where in your own company to begin with AI, talk to the WARP team. Specialists who led DX and data strategy at major companies work alongside you month by month, helping you embed AI into management. Dropping in AI does not solve everything, but let us start from working out where it pays off and where people should stay in the loop, and go from there together.

References and primary sources

Footnotes

1. NVIDIA, "NVIDIA Launches Cosmos World Foundation Model Platform to Accelerate Physical AI Development" (January 2025, announced at CES) https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-world-foundation-model-platform-to-accelerate-physical-ai-development ↩

2. NVIDIA, "NVIDIA Announces Major Release of Cosmos World Foundation Models and Physical AI Data Tools" (March 2025, GTC) https://nvidianews.nvidia.com/news/nvidia-announces-major-release-of-cosmos-world-foundation-models-and-physical-ai-data-tools ↩

3. NVIDIA, "NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI" (June 2026) https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-3-the-open-frontier-foundation-model-for-physical-ai ↩

4. NVIDIA, "NVIDIA Isaac GR00T N1, Open Humanoid Robot Foundation Model" (March 2025) https://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks ↩

5. NVIDIA, "NVIDIA Blackwell-Powered Jetson Thor Now Available" (August 2025) https://nvidianews.nvidia.com/news/nvidia-blackwell-powered-jetson-thor-now-available-accelerating-the-age-of-general-robotics ↩

6. Cabinet Office, joint meeting of the Council on Economic and Fiscal Policy and the Japan Growth Strategy Council, distributed materials (June 2026) https://www5.cao.go.jp/keizai-shimon/kaigi/minutes/2026/0624_shiryo02.pdf ; Cabinet Secretariat, Secretariat of the Headquarters for the Creation of a New Regional Economy and Living Environment, materials https://www.cas.go.jp/jp/seisaku/nipponseichosenryaku/kaigi/dai4/sankou1.pdf ↩