AI is moving out of the cloud and into the physical world, with industrial robotics emerging as one of the first major adopters of what the industry is beginning to call physical AI.
The shift reflects growing demand for autonomous factories, real-time decision making, and machines that can process data locally rather than relying on distant cloud servers.
At the India AI Impact Summit, Gilles Garcia, senior director and business lead for the wired and wireless group at AMD, said the transition follows a familiar pattern in computing—technologies begin in large centralized systems and gradually move closer to where data is generated and decisions must be made. That shift is enabling what Garcia described as “physical AI,” where AI systems operate inside machines that must sense, interpret, and act in real time.
For more than a decade, much of the industry’s focus has been on cloud AI, supported by large-scale compute infrastructure, but many emerging applications require computation closer to the point of action.
Partner Content
View All
Telink TL322X+ML3228A Launched at Embedded World 2026
Telink TL322X+ML3228A Launched at Embedded World 2026
By Telink 03.20.2026
Integrating Digital Isolators in Smart Home Devices
Integrating Digital Isolators in Smart Home Devices
By Monolithic Power Systems 03.19.2026
Hybrid LIC + BBU: Solving AI Server Power Gaps
Hybrid LIC + BBU: Solving AI Server Power Gaps
By Shanghai Yongming Electronic Co.,Ltd 03.19.2026
“What we are seeing is exactly the same pattern that we saw in computing before,” Garcia said. “The big generative AI models that require a lot of compute will remain in the cloud. But other models that need less interaction with the cloud can run at the edge or far edge.”
Robotics leads early adoption
Industries including robotics, autonomous vehicles, healthcare, and telecommunications are beginning to adopt such systems, Garcia said.
Industrial robotics is currently one of the fastest-moving sectors adopting physical AI. Manufacturers are seeking greater autonomy from robots operating on production lines, allowing them to identify defects, adjust processes, and maintain safety with minimal human supervision.
Companies also want greater flexibility in their factories. Production lines increasingly need to switch quickly from manufacturing one type of product to another. Garcia said local AI processing can enable such changes more efficiently than relying on centralized systems.
Humanoid robots represent another area where the technology could expand. Garcia said AI-powered machines designed to interact directly with people could emerge within the next decade in roles such as healthcare assistance, hospital support, and elderly care.
“I believe that in the next ten years humanoid robotics could become one of the biggest transformations to support people,” he said.
Robots on factory floors, for example, cannot wait for cloud processing before taking action.
“The robot needs to be able to act immediately based on perception, sensing, touching, and vision,” Garcia said. “If it has to send information to a data center and wait for an answer, the robot will be useless.”
Systems used in all these environments increasingly rely on embedded processors designed for long operating lifecycles and predictable latency.
Power and latency shape physical AI systems
Running AI models inside physical machines requires strict power limits.In such environments, latency, power consumption, and responsiveness can take precedence over raw computing scale.
As factories adopt more automation, vision systems, and edge analytics, industrial computing platforms face higher performance demands while still needing predictable latency and controlled power consumption. Many systems must operate on batteries or in thermally constrained environments.
“Complex models do not always mean you need large GPUs,” Garcia said. “At AMD, we can run a Llama 3.2 one billion parameter model on a device that is under 25 W.”
Garcia said AMD has spent years developing low-power embedded processors capable of running complex models within those limits. He added that some systems operate in the 5 W to 15 W range, while larger robotic platforms may operate between 10 W and 50 W.
Advancements at the edge
AMD recently introduced the Ryzen Embedded 9000 Series processors for industrial PCs, robotics controllers, and machine-vision systems. Based on the Zen 5 architecture, the processors offer up to 16 CPU cores and configurable power ranges from 65 W to 170 W.
AMD is also developing processors designed specifically for heterogeneous AI workloads.
One example is the Ryzen Embedded P100, which integrates x86 CPU cores, GPU compute capability, and a neural processing unit (NPU) in a single device.
Garcia described the design as a heterogeneous computing platform that distributes tasks across different processing elements depending on performance and efficiency requirements.
“Some tasks will run better on the CPU, others on the GPU, and others on the NPU,” he said. “The objective is to run each function on the technology best suited for it.”
Modern physical AI systems must integrate networking, sensing, and computing. Connectivity allows machines to coordinate with other devices or infrastructure, while local AI accelerators process sensory inputs such as images and signals.
Chipmakers compete to power edge AI
“We know what we know today about AI,” Garcia said. “Six months ago, we did not know what would happen today, and we do not know what will happen in six months.”
Due to this uncertainty, Garcia said AI platforms must remain flexible enough to support multiple models and changing workloads. Today, flexibility, rather than fixed hardware function, is becoming a priority for customers deploying AI systems.
AMD’s strategy spans AI systems ranging from large GPU clusters in data centers to embedded processors used in edge devices. “One size does not fit all,” Garcia said. “The model running in an autonomous car will not be the same as the model in a humanoid robot or an industrial robot.”
AMD is not alone in this space.
Industrial computer makers have long relied on processors from Intel, whose Atom, Core, and Xeon families power many factory PCs and edge systems. These chips compete directly with AMD’s embedded Ryzen parts because they run the same x86 software and operating systems used in industrial environments.
A different group of suppliers approaches the problem from the Arm side. Companies such as NXP Semiconductors, Texas Instruments, and Renesas Electronics build processors aimed at control systems, robotics and industrial gateways where power use and real-time response often matter more than raw computing speed.
In practice, system designers choose between these approaches depending on whether they need a high-performance industrial PC or a smaller controller that runs closer to the machines on the factory floor.
Future implications of physical AI
Looking ahead, Garcia said the spread of physical AI could open new business models across industries.
“I believe physical AI will open completely new business models for startups, industries, and companies,” he said. “Companies may develop new services based on machines that can perceive their environments and make decisions locally.”
However, he said adoption will depend on how naturally AI integrates into everyday activities and industrial processes. “We cannot force someone to use AI,” Garcia said. “It has to become integrated into what people are doing every day.”