The Rise of Physical AI

Physical AI

Artificial intelligence is moving beyond the digital world and into the physical one. Industrial robotics and physical AI are increasingly being adopted across industries, helping companies streamline operations and improve efficiency. At the end of 2024, data shows that there are currently 4.66 million industrial robots ranging from robotic arms, automated assembly, pick and place, and heavy manufacturing, which is up 9% year-over-year.¹ Until recently, price tags on sophisticated robots were extremely high, but cost declines across actuators, reducers, sensors, and chassis manufacturing have pushed humanoid price points down by nearly 97% in just two years, exemplified by Unitree’s G1 humanoid launching at $16,000, versus comparable humanoids at $500,000 back in 2023.

Physical AI rests on the core backbones of Intelligence Infrastructure, Mobility Systems, Power Systems, and Foundational Platforms. Together, they determine not only what robots can do but also whether they can be produced and deployed at scale economically.

The Intelligence Layer

The Intelligence Infrastructure of physical AI encompasses the advanced hardware and software infrastructure that enables robots to operate. This core backbone includes the semiconductor, cloud, and software ecosystems that train and run the AI models embedded inside machines. Companies such as NVDA, TSMC, AMD, and others sit at the center of this ecosystem. NVDA’s Vision-Language-Action (VLA) framework is effectively providing robots with intelligence powerful enough to justify their energy use. In the past, robotic systems relied on separate components for perception, decision-making, and motion. VLA integrates those functions into a unified multimodal system that combines vision, reasoning, and control, sending real-time movement commands directly to a robot’s joints.

Physical AI operates through a two-layer architecture: cloud training and edge inference. In the cloud, large models are trained inside hyperscale data centers, often using advanced simulation environments that replicate different locations such as factories and warehouses. The computational intensity of this process drives sustained demand for high-performance chips. Once training is complete, the model moves from the cloud to the machine itself. Embedded processors called “edge AI chips” handle real-time sensor data and rapidly translate it into awareness, decision-making, coordinated planning, and physical movement.² This is known as edge inference, and it is critical for autonomy.

As physical AI models grow more complex and deployment broadens, demand for both hyperscale training infrastructure and high-performance edge semiconductors is likely to expand in parallel.

The Mechanical Muscle

The Mobility Systems refer to the mechanical structures, such as motors, gears, joints, casings, and wiring, that allow robots to move and exert force. At the center of this system are actuators, which convert electrical signals, allowing the robot to perform precise physical motion. The cost of smart actuators, which were historically expensive, fell from roughly $1,200 to about $600 in just two years.

High-performance actuators depend on rare-earth magnets and specialty minerals capable of delivering strong torque in compact motors.³ Demand for rare-earth materials in robotics is projected to rise sharply over the coming decade, potentially increasing robotics’ share of global rare-earth demand from under 2% today to around 6% by 2035.⁴

Powering Mobility and Compute

The Power Systems for physical AI come in two different forms. The first being rechargeable lithium-ion batteries. These batteries serve as the primary fuel tank for these systems, providing the rapid bursts of electricity needed for functionality and mechanical actuators to physically move the robots. The performance of these batteries has improved, and energy density has climbed around 50% since 2019, allowing robots to operate longer without increasing battery weight.⁵ By 2035, humanoid robots are expected to require approximately 75 GWh of battery capacity, which is an increase from 0.05 GWh in 2025.⁶

While the robot carries its own power, the intelligence it uses is forged elsewhere. The second form of power comes in training the complex AI models, which allows the machine to understand its environment. This process requires an immense amount of computing power housed in hyperscale data centers and could demand 140 megawatts.⁷

From Builders to Accelerators

Beyond component suppliers, the fourth backbone is the Foundational Platforms. These are companies that either build complete robotic systems or accelerate adoption by deploying them at scale. Automakers are increasingly entering the discussion, leveraging expertise in mass manufacturing, supply chains, and electric drivetrains. At the same time, large-scale operators are demonstrating how robotics can reshape complex supply chains. Amazon has deployed over one million robots across its fulfillment network⁸, using systems designed to improve ergonomics, speed, and efficiency. While robotics don’t show up as revenue for Amazon, it is a cost efficiency and capacity engine powering fulfillment, delivery, and logistics. Amazon has currently implemented Kiva-Style Drive Units, which are ideal for large fulfillment centers, along with another robot, Proteus — which is designed to operate in shared spaces with humans, unlike traditional Kiva robots, which are restricted to fenced areas — and Sparrow, a picking robot designed to automate repetitive picker tasks.

These deployments illustrate a broader point: physical AI is not simply about replacing labor. In many cases, it enhances safety, improves retention, and reallocates human workers toward more complex, customer-facing responsibilities.

Stephanie Link’s TV Schedule:

Sources:

1. International Federation of Robotics: September 25, 2025 ↩︎
2. Barclays Special Report: As of February 23, 2026 ↩︎
3. Barclays Special Report: As of February 23, 2026 ↩︎
4. Barclays Special Report: As of February 23, 2026 ↩︎
5. Barclays Special Report: As of February 23, 2026 ↩︎
6. Intro-act Mobility Monthly: As of February 10, 2026 ↩︎
7. Bloomberg: As of April 15, 2025 ↩︎
8. Barclays Special Report: As of February 23, 2026 ↩︎

Disclosures
Investment Solutions is a group comprised of investment professionals registered with Hightower Advisors, LLC, an SEC registered investment adviser. Some investment professionals may also be registered with Hightower Securities, LLC, member FINRA and SIPC. Advisory services are offered through Hightower Advisors, LLC. Securities are offered through Hightower Securities, LLC. This is not an offer to buy or sell securities. No investment process is free of risk, and there is no guarantee that the investment process or the investment opportunities referenced herein will be profitable. Past performance is neither indicative nor a guarantee of future results. The investment opportunities referenced herein may not be suitable for all investors. All data or other information referenced herein is from sources believed to be reliable. Any opinions, news, research, analyses, prices, or other data or information contained in this presentation is provided as general market commentary and does not constitute investment advice. Investment Solutions and Hightower Advisors, LLC or any of its affiliates make no representations or warranties express or implied as to the accuracy or completeness of the information or for statements or errors or omissions, or results obtained from the use of this information. Investment Solutions and Hightower Advisors, LLC assume no liability for any action made or taken in reliance on or relating in any way to this information The information is provided as of the date referenced in the document. Such data and other information are subject to change without notice. This document was created for informational purposes only; the opinions expressed herein are solely those of the author(s) and do not represent those of Hightower Advisors, LLC, or any of its affiliates.