Woofun AI reports that Micron delivered a record-breaking earnings statement for the third quarter of fiscal year 2026, posting total revenue of $41.5 billion, which represents a staggering 346% year-over-year increase. The company's Non-GAAP gross margin climbed to 84.9%, marking the fifth consecutive quarter of new revenue records. Beyond these headline figures, the earnings call revealed a critical strategic signal often overlooked: Micron asserts that AI-driven storage demand is expanding far beyond data centers to encompass smartphones, PCs, automobiles, industrial applications, and robots. Specifically, the company highlighted that humanoid robots and advanced autonomous driving systems will generate long-term storage requirements. Micron provided a direct quantitative comparison to illustrate this shift, noting that L2+ and higher-level autonomous vehicles require more than five times the memory and storage capacity of ordinary cars.

Furthermore, the storage capacity needed for humanoid robots is approximately ten times that of L2+ autonomous vehicles. Crucially, Micron predicts that starting in the second half of the 2020s, before 2030, these markets will enter a phase of large-scale storage demand lasting for several decades.

The rationale behind a storage giant, whose primary growth engine remains AI data centers, emphasizing robots and autonomous driving lies in a fundamental industry evolution. The AI value chain is moving past the singular focus on "the larger the model, the more GPUs and HBM are needed" and is shifting its growth trajectory for the next decade or more from digital AI to physical AI. This transition was underscored in January 2026 when Jensen Huang stated on the CES stage in Las Vegas, "The time for physical AI, similar to ChatGPT, has arrived." While this declaration contains promotional elements, the underlying trend is already manifesting in real-world metrics. By the end of 2025, Waymo's autonomous vehicles had completed 450,000 paid trips per week.

Additionally, in June 2025, Amazon deployed its 100th robot to assist in handling 75% of its global customer order deliveries. These instances demonstrate the core problem physical AI addresses: enabling machines to determine their next action within a real-world environment characterized by uncertainty. Physical AI is indispensable in scenarios requiring independent machine execution, including robotics, autonomous vehicles, and industrial control systems.

The transition from large-scale AI models to physical AI represents not merely a technical conceptual shift but a leap in economic scale. If AI applications extend beyond digital offices into automobiles, factories, warehouses, hospitals, energy facilities, and households, the addressable market will encompass a significantly larger portion of the real economy. A report titled "Global Physical AI Market 2026-2040" by the British consulting firm Future Markets projects that the global physical AI market will grow from approximately $383 billion in 2026 to $3.26 trillion by 2040. This trajectory suggests one of the largest expansions in the history of technology markets. At the beginning of the year, Deutsche Bank predicted that 2026 would serve as a turning point for autonomous driving, marking the move from testing phases to large-scale implementation, and for humanoid robots, signifying the shift from laboratories to small-scale mass production. Consequently, major corporations are increasingly focusing resources on physical AI as technology giants collectively seek a second growth trajectory for the AI industry chain.

To understand the impact of physical AI on the current value distribution within the AI industry chain, one must first determine if it will generate new product categories. In the short term, physical AI will rely on the existing infrastructure of the AI industry chain. Training models requires GPUs, while training and inference tasks depend on HBM, cloud computing relies on data centers, and robot manufacturers need cloud services, chips, storage solutions, and software tools. Thus, Micron's HBM, NVIDIA's GPUs, and the computing resources and modeling capabilities of various cloud providers remain essential.

However, in the long run, physical AI will create new demands within the existing industry chain that did not previously exist.

Woofun AI data shows that the logic of storage is being fundamentally rewritten as the industry moves toward mobile multi-sensor AI systems. While large-scale AI model storage needs are primarily focused on cloud training, humanoid robots require local storage and processing for video streams, maps, trajectories, local models, task memories, failure cases, sensor caches, and control logs. Low-power, high-bandwidth, and highly reliable memory and storage solutions will become crucial for physical AI, validating Micron's view that humanoid robots represent a "decade-long demand cycle."

At the chip and storage layer, the requirements diverge significantly from current data center standards. Currently, large-scale AI model training relies primarily on GPUs in data centers, but physical AI also necessitates AI chips installed directly on robots alongside cloud-based training. These AI chips operate in distinct environments and must meet specific constraints, including low power consumption, low latency, resistance to vibration, efficient heat dissipation, and the ability to function over extended periods. They must also integrate cameras, radars, IMUs, tactile sensors, and motor control systems. NVIDIA's Jetson Thor, Qualcomm's Dragonwing IQ10, AMD's Embedded solutions, and Arm Holdings' edge architectures will all play key roles in this emerging segment of the industry chain. The storage logic is equally transformative. Humanoid robots are mobile multi-sensor AI systems where video streams, maps, trajectories, local models, task memories, failure cases, sensor caches, and control logs all need to be stored and processed locally. This necessitates low-power, high-bandwidth, and highly reliable memory and storage solutions, which is precisely why Micron believes that humanoid robots represent a "decade-long demand cycle."

At the model and software layer, the distinction between digital and physical AI redefines market boundaries. Large-scale AI models generate Tokens, whereas physical AI generates Actions. The technical foundation of physical AI is no longer language models but world models, which aim to compress the operating principles of the physical world into model parameters to enable AI to understand space, motion, and causal relationships.

However, world models alone are insufficient. A robot capable of acting in the physical world requires a complete suite of models and software services, including grasping models, navigation models, operation models, safety models, robot operating systems, remote maintenance systems, OTA updates, and Fleet Learning platforms. This list of requirements is gradually becoming a reality. Alibaba has released the Qwen-Robot series of embodied intelligence models dedicated to navigation, operation, and simulation of the physical world. NVIDIA has released the open-source basic model Isaac GR00T for humanoid robots, which supports inference, learning, and multi-task behavior. In the future, robot manufacturers may no longer need to develop each component from scratch; they can purchase existing models and platforms. This will create a brand-new market, potentially as large as today's model and SaaS markets, but its target customers will be machines that operate in the physical world.

At the platform and simulation layer, scaling physical AI implementation requires a corresponding expansion in simulation capabilities. Robots need to understand variables such as the weight of a cup, the friction coefficient on wet ground, and whether lighting changes cause vision systems to misjudge distances. This information cannot be derived from text; it requires accumulating vast amounts of real-world data, allowing robots to run millions of scenarios in virtual simulation environments before transferring trained capabilities to real hardware. This necessity has given rise to a new category of products: development, simulation, deployment, and validation platforms. Currently, NVIDIA is working to make this physical AI toolchain the next generation of CUDA. It combines four core products: Isaac for robot development and simulation, Omniverse for digital twins, Cosmos for world models, and GR00T for basic models and data pipelines specifically for humanoid robots. Automakers such as BMW and Mercedes-Benz are already using NVIDIA's Omniverse simulation platform to create digital twins of entire factories. Once this platform becomes essential for training physical AI, NVIDIA's role will shift from selling GPUs to providing the operating system framework for physical AI.

The changes across the previous three layers will ultimately impact factory production lines, warehouse shelves, and the driver seats of intelligent vehicles. For example, Samsung announced at MWC 2026 that it would transform its global manufacturing systems into AI-driven factories by 2030. Caterpillar is using Omniverse to create digital twins of factories for predictive maintenance and flexible production scheduling. These developments signal changes in real capital expenditures, leading to a reevaluation of the entire production infrastructure. Labor costs, factory depreciation, supply chain resilience, production efficiency, safety risks, and maintenance expenses are all being recalculated. This is the true ambition of physical AI: it is not about adding another application scenario for AI but about connecting AI directly to the physical components of the GDP for the first time. When cars transform into intelligent driving robots and factories become collaborative networks of embodied intelligence, the entire valuation logic of the manufacturing industry will be rewritten.

If physical AI truly becomes the next major technological trend, the question arises as to which companies will benefit the most. Initially, one might assume that companies producing robots, such as those manufacturing robot bodies, manipulators, reducers, motors, and sensors, will be the first to benefit.

However, counterintuitively, these companies may not be the first to reap the rewards. The experience of the AI industry in the past two years demonstrates that in the early stages of a technological revolution, it is often the companies providing necessary infrastructure, rather than application companies, that generate the most profit. During the wave of large-scale AI models, it was not AI application companies like OpenAI and Google that profited first but infrastructure companies such as NVIDIA, TSMC, Micron, SK Hynix, and Broadcom. The same pattern is likely to repeat with physical AI. Before robots are widely used in factories, warehouses, homes, and the service sector, manufacturers will need to train robot models, build simulation environments, collect real-world data, purchase edge chips, configure local storage systems, establish remote maintenance systems, and repeatedly test the safety and stability of these systems. Therefore, in the early stages of physical AI, it is not necessarily the companies that produce robots that will benefit the most but those that provide the necessary infrastructure for the robot era. Only this time, the "shovels" required are far more complex than those in the era of large-scale AI models. They consist of a complete set of infrastructure components, ranging from cloud services to robot bodies, from simulation tools to execution mechanisms, and from models to control systems.

This explains why Micron highlighted the true value of robots in its earnings call: it is one of the first links in this new value chain, responding to these changes at least one or two earnings reporting cycles before robot manufacturers start seeing actual sales increases. Once this value chain takes shape, physical AI will not just be about robots but will represent a new round of expansion for the entire AI industry chain. This marks a definitive shift where infrastructure providers will capture value long before end-user adoption reaches critical mass.