Written by: Jim, MSX Maitong

Edited by: Frank, MSX Maitong

In the past two years, the capital market's focus on AI has primarily revolved around AI's "brain".

From ChatGPT and large models to GPUs, HBMs, data centers, optical communication, and power infrastructure, almost all core lines of development have centered on how to make models larger, train faster, and reduce inference costs.

However, while these AIs can generate text, images, code, and video, they mostly still operate on screens and in the digital world.

Therefore, as the capabilities of large models and computational power infrastructures gradually mature, the market will naturally begin to question the next issue: Can these increasingly intelligent models eventually step out of screens and enter cars, factories, warehouses, hospitals, and the real world?

This is exactly why Physical AI has started to come to the forefront of industries.

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

According to NVIDIA's definition, Physical AI enables AI to step out of the screen, allowing autonomous systems like robots, cameras, and self-driving cars to perceive and understand their surroundings, completing reasoning, decision-making, and complex actions.

In other words, if generative AI addresses "how machines think", then Physical AI aims to solve how to act correctly, safely, and at low cost after machines think, thereby enabling machines to truly interact with the real world.

From Jensen Huang's recent public speeches, NVIDIA is continuously strengthening product lines like Isaac, GR00T, Cosmos, Omniverse, and Jetson, not merely betting on a single robot, but building a comprehensive underlying platform that covers training, simulation, reasoning, and deployment for machines entering the physical world.

Because true Physical AI is not as simple as connecting a large model inside 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 completing extensive safety testing before deployment.

In the market context, Physical AI highly overlaps with "embodied intelligence," but the former has a broader connotation, including not only humanoid robots but also autonomous driving, industrial robots, drones, smart factories, warehousing systems, and intelligent spaces driven by cameras and sensors.

Of course, Physical AI is not a suddenly emerging concept.

Self-driving cars, industrial robots, machine vision, and warehouse automation have been developing for many years; the real change is that large models, world models, simulation technology, and edge computing are connecting these previously relatively fragmented technology routes.

Numerous traditional industrial robots rely on pre-written programs to execute standard actions repetitively in relatively fixed environments; the goal of Physical AI is to enable machines to adjust judgments and behaviors based on real-time information even when facing different objects, unfamiliar environments, and unexpected situations.

This means that the AI industry chain is extending from the "brain" to the "body".

In the past two years, the market first reassessed the GPUs, storage, servers, networks, and power needed to train and run AI. Subsequently, funds may further seek to invest in platforms that can harness this computational power, transforming model capabilities into real productive forces: robots, self-driving cars, drones, industrial automation devices, and visual and sensory systems spread across factories, warehouses, and cities.

Thus, Physical AI is not a concept that can be simply equated with "humanoid robots"; it truly opens up an entire industry chain from computational power to action.

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

For easy understanding, the MSX Research Institute has roughly divided 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 in cars and robots, computational power is essential.

This encompasses data center GPUs, edge AI chips, onboard computing platforms, and low-power processors, with the main targets including:

• NVIDIA (NVDA.M): Covers training computational power, Jetson edge computing platform, and robot development ecosystem;
• Taiwan Semiconductor Manufacturing Company (TSM.M): Manufacturing base for AI chips, vehicle chips, and edge computing chips;
• Arm (ARM.M): Low-power computing architecture widely used in vehicles, robots, and smart devices;
• Qualcomm (QCOM.M): Layouts in vehicle AI, edge reasoning, and smart terminals;
• AMD (AMD.M): Potential beneficiary of AI computational power and embedded computing;

The logic of this layer is similar to the generative AI market in the past two years, following the "selling shovels" logic, regardless of which robot company ultimately prevails, the basics require chips, computational power, and computing architecture.

2. Second Layer: Model Layer

This is also easy to understand; Physical AI requires not just language models but also foundational robot models, world models, and vision-language-action models.

Language models can understand human instructions, vision models help machines recognize environments, and action models are responsible for converting judgments into specific actions; world models take it a step further, attempting to make AI understand relationships between objects, predict what might happen next, and perform reasoning before actions.

This layer is currently mainly driven by large tech companies and platform enterprises, including NVIDIA, Tesla, Google, and some robot startups.

Compared to large language models, the biggest challenge facing robot models is data, as there is a wealth of text, images, and videos on the internet, but truly high-quality data for robot operations is scarce; how to generate sufficient training data will become a critical hurdle in the development of Physical AI.

3. Third Layer: Simulation Layer

Due to high training costs, slow speeds, and significant risks in reality, robots need to learn in virtual worlds first; thus, digital twins, synthetic data, and virtual training environments form a very important layer of Physical AI.

NVIDIA has built a relatively complete toolchain in this layer: Omniverse is used to construct digital twins and simulation environments, Isaac Sim and Isaac Lab support robot training, testing, and validation, while Cosmos provides world model and data generation capabilities.

The value of this layer lies in that it can transfer expensive, dangerous, and slow trial-and-error processes in the real world to a virtual environment, allowing developers to run numerous scenarios simultaneously, testing different lighting, weather, terrain, and unexpected events, then deploying the validated models to real devices.

Ultimately, training a robot in reality may take a few minutes, but it can run thousands of times in a simulation environment simultaneously.

4. Fourth Layer: Perception Layer

As robots enter the real world, the first step is often not about having dexterous hands but about being able to "see" and understand their surroundings stably.

They must recognize 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: Provide onboard and edge visual computing platforms;

Ouster has integrated a new generation of 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 of machine vision and sensors may not be as large but is closer to real orders and existing customers.

As for the execution end with motors, reducers, and joint modules, the pure stocks in the US are relatively limited, with related opportunities more dispersed among industrial automation, simulation chips, and specialized component enterprises.

5. Fifth Layer: Application Layer

As the top layer of the industry chain, this is also where the market is most familiar with robots, autonomous driving, drones, and industrial automation devices, corresponding to the following targets:

• Tesla (TSLA.M): Optimus, FSD, and Robotaxi;
• Alphabet (GOOGL.M): Entering autonomous driving through Waymo;
• Amazon (AMZN.M): Warehousing 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 platforms connecting data, decision-making, and unmanned devices;

Among these, Palantir is not a robot manufacturer but leans more towards a software platform that connects data, decision-making, and unmanned devices; Uber may become a platform for different Robotaxi fleets to acquire users, schedule orders, and complete transactions, both of which belong to indirect beneficiaries.

This is also the part of Physical AI where high elasticity is most likely to occur; once a certain robot, Robotaxi, or drone enters mass production, the market will quickly revise its revenue and valuation upwards.

However, at the same time, the application layer is also the most competitive and the part with the highest difficulty in realization.

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

From the perspective of industry realization order, the incremental revenue and profit brought about by Physical AI may not first appear in the most sci-fi humanoid robots.

Instead, a more probable path is to first sell underlying platforms, then enter closed scenarios; first solve standardized tasks, then challenge open worlds; in short, the certainty of "selling shovels" remains the highest.

Therefore, if the biggest beneficiary of the first phase of generative AI is NVIDIA, the early development of Physical AI still finds it hard to bypass NVIDIA. Whether it is Tesla, Amazon, or a certain robot startup that ultimately prevails, they all need model training, simulation testing, real-time reasoning, and edge deployment.

NVIDIA's advantage lies not just in GPUs, but in its integration of chips, models, simulation software, and edge computing platforms into a complete development system, which also means it does not need to produce every single robot itself, but rather allow more and more robots to utilize its computational power and software ecosystem.

From this perspective, the clearer beneficiary direction in the first phase of Physical AI may still be the "shovel sellers" providing computational power, simulation, chips, and development tools; but "clear beneficiary paths" do not equate to stock price risks; it still requires observation whether the market has already factored in growth expectations, whether the software ecosystem can form sustainable income, and whether competitors can provide alternatives.

Next could be factories and warehouses, which might more quickly establish a commercial closed loop, meaning the scenarios where Physical AI first enters financial reports will likely be in manufacturing, warehousing, and logistics.

These scenario environments are relatively closed, with routes and tasks more standardized, making it easier for companies to calculate ROI—after investing in a robot, how much labor can be reduced, how much efficiency improved, and how much loss decreased can all be directly quantified.

Amazon has begun to use robots extensively in its warehousing network and optimized scheduling and routes between devices through AI models; Universal Robots and MiR, subsidiaries of Teradyne, have respectively covered collaborative robotic arms and autonomous mobile robots that have entered real production environments such as manufacturing, logistics, and semiconductors.

The common feature of these companies is that they not only exhibit what robots can do but have already begun to place robots into factories and warehouses to solve real production problems. In contrast, getting robots to enter homes to cook, clean, and care for the elderly requires facing more complex environments and safety responsibilities, with commercialization cycles likely to be significantly longer.

Lastly, 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 closely watched directions in the Physical AI market, but this does not equate to large-scale commercialization already being here; for humanoid robots, what truly needs observation is not whether the actions at the launch event are smooth, but the unit cost, continuous working time, and whether the value they create can cover procurement and maintenance costs.

In contrast, Robotaxis have already taken a more advanced position. Autonomous vehicles are essentially "Physical AI on wheels"—vehicles perceive their environment through cameras, radars, and LiDAR, make judgments based on the model, and then take actual action.

Tesla, Waymo, and Zoox respectively represent fully integrated hardware and software for vehicles, autonomous driving systems, and dedicated Robotaxi routes; Uber is trying to become a platform connecting different autonomous fleets and passengers; Waymo has begun to promote fully unmanned operations of its sixth-generation autonomous driving system, and it has disclosed that its latest models equipped with this system have completed over 20 million fully unmanned rides, indicating that Robotaxis are clearly leading commercial validation over general humanoid robots.

In addition, drones and defense robots find it easier to obtain order validation. After all, defense clients have a clearer demand for autonomous, low-cost unmanned systems and counter-drone devices, with companies like AeroVironment and Kratos already reflecting revenue and order growth through their autonomous and unmanned systems business, and Ondas continues to secure orders for counter-drone, loitering munitions, and autonomous defense systems.

However, these types of small companies often come with higher project concentration, financing, and execution risks.

Therefore, determining whether a Physical AI company is worth following continuously ultimately boils down to three questions:

• Is it a core link in the industry chain that is difficult to replace?
• Does it have real clients, orders, and application scenarios?
• Can technological progress ultimately reflect in revenue, profit, and cash flow?

Final Thoughts

Physical AI will not be realized overnight.

From the perspective of industrial规律, it is more likely to advance along a path that transitions from certainty to high elasticity: starting with computational power, simulation, and edge platforms, followed by warehouses, factories, and specialized robots, then to Robotaxis, drones, and generic humanoid robots.

What truly determines how far this main line can go is not how many actions robots take at the press conference, but whether they can enter factories, warehouses, roads, and real businesses after stepping down from the stage and create value that can be validated through financial reports.

When this happens, AI can truly be said to have walked from the screen into reality.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。