When AI Begins to Have a Body: Will Physical AI Become the Next Mainstream Technology?

Written by: Jim, MSX Maitong
Edited by: Frank, MSX Maitong

In the past two years, the AI traded in the capital market has primarily been the "brain" of AI.
From ChatGPT, large models to GPUs, HBM, data centers, optical communication, and power infrastructure, almost all core lines revolve around how to make models larger, training faster, and inference cheaper.
However, these AIs can generate text, images, code, and videos, but most still operate within screens and the digital world.

Therefore, as the capabilities of large models and computational infrastructure gradually mature, the market naturally begins to ask the next question: Can these increasingly intelligent models eventually step out of the screen and enter cars, factories, warehouses, hospitals, and the real world?

This is precisely why Physical AI is beginning to come to the forefront of the industry.

1. From "Thinking" to "Acting": Why is Physical AI Important?

According to NVIDIA's definition, Physical AI allows AI to step out of the screen, enabling autonomous systems like robots, cameras, and self-driving cars to perceive and understand their surroundings, complete reasoning, decision-making, and complex actions.
In other words, if generative AI addresses "how machines think," then Physical AI attempts to solve how machines can act correctly, safely, and cost-effectively after thinking, thereby enabling machines to truly interact with the real world.
From Huang Renxun's recent public speeches, it is evident that NVIDIA is continuously strengthening its product lines such as Isaac, GR00T, Cosmos, Omniverse, and Jetson. The goal is not merely to bet on a single robot but to build a comprehensive underlying platform that covers training, simulation, reasoning, and deployment for machines to enter the physical world.
Because true Physical AI is not as simple as connecting a large model in a robot; it also needs to understand spatial relationships and physical laws, requiring world models, training data, simulation environments, edge computing, machine vision, sensors, and motion control, along with extensive safety testing before deployment.
In the market context, Physical AI overlaps significantly with "embodied intelligence," but the former has a broader connotation, encompassing not only humanoid robots but also self-driving cars, industrial robots, drones, smart factories, warehouse systems, and intelligent spaces driven by cameras and sensors.
Of course, Physical AI is not a suddenly emerging new concept.
Self-driving, industrial robots, machine vision, and warehouse automation have been developing for many years. What has truly changed is that large models, world models, simulation technologies, and edge computing are connecting these previously relatively fragmented technological routes.
Many traditional industrial robots rely on pre-written programs to repeatedly execute standard actions in relatively fixed environments; the goal of Physical AI is to enable machines to adjust their judgments and behaviors based on real-time information when facing different objects, unfamiliar environments, and unexpected situations.
This means that the AI industry chain is extending from "brain" to "body."
In the past two years, the market first reassessed the GPUs, storage, servers, networks, and power needed to train and run AI. Next, funds may further seek carriers that can undertake this computational power and transform model capabilities into real productivity: robots, self-driving cars, drones, industrial automation equipment, and visual and sensing systems spread across factories, warehouses, and cities.
Thus, Physical AI is not a single-point concept that can simply be equated with "humanoid robots"; it truly opens up an entire industrial chain from computational power to action.

2. From Computational Power to Robots: The Five Layers of the Physical AI Industry Chain

To facilitate understanding, MSX Research Institute roughly divides the Physical AI industry chain into five key links.

1. First Layer: Computational Power Layer

Whether training robot models, building virtual environments, or completing real-time reasoning on cars and robots, computational power is indispensable.
It encompasses data center GPUs, edge AI chips, onboard computing platforms, and low-power processors, with the main targets including:

• NVIDIA (NVDA.M): covering training computational power, Jetson edge computing platform, and robot development ecosystem;
• TSMC (TSM.M): manufacturing base for AI chips, automotive chips, and edge computing chips;
• Arm (ARM.M): low-power computing architecture widely used in cars, robots, and smart devices;
• Qualcomm (QCOM.M): layout in automotive AI, edge reasoning, and smart terminals;
• AMD (AMD.M): potential beneficiaries of AI computational power and embedded computing;
  The logic of this layer is similar to the generative AI market in the past two years, continuing the "selling shovels" logic. Regardless of which robot company ultimately wins, the foundation requires chips, computational power, and computing architecture.

2. Second Layer: Model Layer

This is also easy to understand; Physical AI requires not only language models but also foundational robot models, world models, and visual-language-action models.
Language models can understand human instructions, visual models help machines recognize environments, and action models are responsible for translating judgments into specific actions; world models go further, attempting to enable AI to understand relationships between objects, predict what might happen next, and simulate before acting.
This layer is currently mainly driven by large tech companies and platform enterprises, including NVIDIA, Tesla, Google, and some robotics startups.
Compared to large language models, the biggest challenge faced by robot models is data; although there is a vast amount of text, images, and videos on the internet, truly high-quality robot operation data is scarce. Generating enough training data will become a key hurdle in the development of Physical AI.

3. Third Layer: Simulation Layer

Due to the high cost, slow speed, and significant risks of real-world training, robots need to learn in a virtual world first. Therefore, digital twins, synthetic data, and virtual training environments constitute a very important layer of Physical AI.
NVIDIA has built a relatively complete toolchain at this layer: Omniverse is used to build digital twins and simulation environments, Isaac Sim and Isaac Lab support robot training, testing, and validation, while Cosmos provides world models and data generation capabilities.
The value of this layer lies in its ability to transfer the expensive, dangerous, and slow trial-and-error process in the real world to a virtual environment, allowing developers to run numerous scenarios simultaneously, testing different lighting, weather, terrain, and unexpected events, and then deploying the validated models to real devices.
Ultimately, training a robot in reality may take several minutes, while in a simulation environment, it can run thousands of times in parallel.

4. Fourth Layer: Perception Layer

When robots enter the real world, the first step is often not to have flexible hands but to be able to stably "see" and understand the surrounding environment.
They must identify objects, judge distances, understand environmental changes, and complete positioning in complex spaces. After making judgments, they also need to translate decisions into real actions through controllers, motors, robotic arms, and joint modules.
This layer includes machine vision, cameras, LiDAR, sensors, control chips, motion control, and various execution components:

• Cognex (CGNX.M): industrial machine vision and recognition systems;
• Ouster (OUST.M): LiDAR and perception platforms;
• Qualcomm, NVIDIA: providing onboard and edge visual computing platforms;
  Ouster has integrated the next-generation digital LiDAR into NVIDIA Jetson and Isaac ecosystems, advancing applications in industrial robots, inspections, and autonomous systems; Cognex continues to deploy AI vision systems in manufacturing inspection and automation scenarios.
  Compared to humanoid robots, the imaginative space for machine vision and sensors may not be as vast, but it is closer to real orders and existing customers.
  As for execution ends like motors, reducers, and joint modules, pure targets in the US stock market are relatively limited, with relevant opportunities more dispersed among industrial automation, simulation chips, and specialized component companies.

5. Fifth Layer: Application Layer

As the top layer of the industry chain, this is also the most familiar market segment for robots, self-driving cars, drones, and industrial automation equipment, corresponding to targets including:

• Tesla (TSLA.M): Optimus, FSD, and Robotaxi;
• Alphabet (GOOGL.M): laying out self-driving through Waymo;
• Amazon (AMZN.M): warehouse robots, logistics automation, and Zoox;
• Teradyne (TER.M): collaborative robots and mobile robots;
• AeroVironment (AVAV.M), Kratos (KTOS.M), Ondas (ONDS.M): drones and unmanned systems;
• Palantir (PLTR.M): software platform connecting data, decision-making, and unmanned devices;
  Among these, Palantir is not a robot manufacturer but leans more towards a software platform connecting data, decision-making, and unmanned devices; Uber may become a traffic entry point for different Robotaxi fleets to acquire users, schedule orders, and complete transactions, both of which belong to indirect beneficiary directions.
  This is also the segment where Physical AI is most likely to generate high elasticity; once a certain robot, Robotaxi, or drone enters large-scale production, the market will quickly revise its revenue and valuation space.
  However, at the same time, the application layer is also the most competitive and hardest to realize.

3. Who Will Make Money First: Selling Shovels or Building Robots?

From the perspective of industry realization order, the incremental revenue and profit brought by Physical AI may not necessarily appear first in the most sci-fi humanoid robots.
Instead, the more likely path is to first sell underlying platforms, then enter closed scenarios; first solve standardized tasks, then challenge the open world; in short, the certainty of "selling shovels" remains the highest.
Therefore, if the biggest beneficiary in the first phase of generative AI is NVIDIA, then the early development of Physical AI still finds it hard to bypass NVIDIA. Regardless of whether Tesla, Amazon, or a robotics startup ultimately wins, they all need model training, simulation testing, real-time reasoning, and edge deployment.
NVIDIA's advantage is not just its GPUs but that it is integrating chips, models, simulation software, and edge computing platforms into a complete development system, which also means it does not need to produce every robot itself but only needs to enable more robots to use its computational power and software ecosystem.

From this perspective, the clearer benefit direction in the first phase of Physical AI may still be the "shovel sellers" providing computational power, simulation, chips, and development tools; however, "clear benefit paths" do not mean that stock prices are without risk. Whether the market has already priced in growth expectations, whether the software ecosystem can form sustainable income, and whether competitors can provide alternatives still need observation.
Next, factories and warehouses may run through the commercial closed loop earlier, meaning that the scenarios where Physical AI first enters financial reports are likely to appear in manufacturing, warehousing, and logistics.
These scenarios have relatively closed environments, and the routes and tasks are more standardized, making it easier for companies to calculate return on investment—after investing in a robot, how much labor can be reduced, how much efficiency can be improved, and how much loss can be reduced can all be directly quantified.
Amazon has already used robots on a large scale in its warehousing network and optimized scheduling and routes between devices through AI models; Teradyne's Universal Robots and MiR cover collaborative robotic arms and autonomous mobile robots, respectively, and have entered actual production environments in manufacturing, logistics, and semiconductors.
The common characteristic of these companies is that they are not just showcasing what actions robots can perform but have already started placing robots in factories and warehouses to solve real production problems. In contrast, allowing robots to enter homes for cooking, cleaning, and caring for the elderly faces more complex environments and safety responsibilities, and the commercialization cycle may be significantly longer.
Finally, humanoid robots undoubtedly possess the greatest market imagination; theoretically, they can enter factories, warehouses, hospitals, and homes designed by humans, directly using existing roads, tools, and workbenches.
Tesla Optimus has thus become one of the most watched directions in the Physical AI market, but this does not mean that large-scale commercialization has arrived. For humanoid robots, what truly needs observation is not whether the actions at the press conference are smooth but whether the unit cost, continuous working time, and the value it creates can cover procurement and maintenance costs.

In contrast, Robotaxi is already in a more advanced position. Self-driving cars are essentially "Physical AI on wheels"—vehicles perceive the environment through cameras, radars, and LiDAR, models make judgments, and then the car completes the actual actions.
Tesla, Waymo, and Zoox represent the integration of vehicle software and hardware, self-driving systems, and dedicated Robotaxi routes, respectively; Uber attempts to become the platform entry connecting different autonomous driving fleets with passengers; Waymo has begun to promote the complete unmanned operation of its sixth-generation autonomous driving system, with its latest models equipped with this system having completed over 20 million fully unmanned driving trips, indicating that Robotaxi is clearly ahead in commercial validation compared to general humanoid robots.
In addition, drones and defense robots are more likely to obtain order validation. After all, defense customers have clearer demands for autonomous, low-cost unmanned systems and anti-drone equipment, with companies like AeroVironment and Kratos showing revenue and order growth in their autonomous and unmanned systems business, while Ondas continues to receive orders for anti-drone, loitering munitions, and autonomous defense systems.
However, these small companies typically come with higher project concentration, financing, and execution risks.
Therefore, determining whether a Physical AI company is worth tracking continuously ultimately returns to three questions:

• Is it an irreplaceable core link in the industry chain?
• Does it have real customers, orders, and application scenarios?
• Can technological progress ultimately be reflected in revenue, profit, and cash flow?

In Conclusion

Physical AI will not realize its potential overnight.
From an industrial perspective, it is more likely to follow a path that gradually advances from certainty to high elasticity: first computational power, simulation, and edge platforms, then warehouses, factories, and specialized robots, followed by Robotaxi, drones, and general humanoid robots.
What truly determines how far this main line can go is not how many actions robots complete at press conferences but whether they can enter factories, warehouses, roads, and real businesses after stepping off the stage and create value that can be verified by financial reports.
When this happens, AI can be considered to have truly stepped out of the screen and into reality.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.