Attributed to Steven Fong, Corporate Vice President, APJ Embedded Business AMD
Artificial intelligence is entering a new phase—one that extends beyond digital productivity into the physical world.
For the past several years, AI has largely been defined by systems that understand, generate, and recommend information. Large language models (LLMs), AI assistants, and copilots have fundamentally changed knowledge work by helping people search, summarise, create content, and improve decisions. This first wave of AI has transformed digital productivity.
The next wave will be fundamentally different. Rather than simply understanding information, AI is increasingly expected to interact with the physical world.
Robots will work alongside people on factory floors. Autonomous vehicles will continuously perceive and respond to their surroundings. Medical systems will assist clinicians with real-time diagnosis and decision-making. Intelligent infrastructure will monitor, optimise, and respond without constant human intervention. This transition represents far more than another AI application. It marks a fundamental shift in what AI is expected to do.
As AI moves from the cloud and digital environments into physical systems, intelligence must no longer simply know. Physical AI systems like autonomous robots must also sense, decide, and ultimately act in real-time, with functional safety and while operating within strict power, thermal and space constraints. The need for this transition is being driven not only by advances in AI technology, but also by structural changes across the global economy.
This is precisely where physical AI begins. Physical AI is already reshaping industries, and AMD believes this is the defining transition of the next decade of AI. It is the transition from systems that know to systems that do.
“Doing” Fundamentally Changes The Compute Problem
Once AI begins interacting with the physical world, the compute problem fundamentally changes. Unlike cloud AI, which has traditionally been optimised for scale, throughput, and centralised processing, physical AI introduces a fundamentally different set of requirements.
A chatbot can tolerate occasional latency. However, a robot operating on a factory floor cannot. An autonomous vehicle cannot afford inconsistent system behavior. A medical device cannot simply retry a decision when timing is critical.
These systems must deliver:
- Real-time responsiveness
- Deterministic behavior
- Functional safety and reliability
- Operation under power and thermal constraints
- Resilience when connectivity is limited or unavailable
In the physical world, intelligence only matters if systems can act within bounded time constraints. Success is no longer determined solely by AI model performance.
Physical AI extends beyond running AI models. Instead, intelligence must flow seamlessly from sensing and perception through inference, planning, control, and ultimately physical action. The conversation is no longer just about building smarter AI models. It is about building intelligent systems that can safely and reliably act in the real world.
This is why AMD believes physical AI is fundamentally a systems engineering challenge, not simply an AI model challenge.
No Single Processor Can Power Physical AI
Understanding physical AI becomes much easier when we compare it with the human body.
The human brain performs reasoning and decision-making. Our nervous system enables rapid responses. Our muscles and joints execute precise physical actions.
Physical AI systems operate in much the same way. Complex reasoning and perception require high-performance compute. Real-time responses require low-latency processing close to where decisions are made. Precise movement requires accurate and deterministic control.
No single processor architecture can efficiently perform all of these functions. Instead, different types of compute must work together.
High-performance CPUs and GPUs provide the “brain,” responsible for complex reasoning, perception, and AI workloads. Adaptive SoCs and NPUs function like the nervous system, enabling low-latency, power-efficient processing close to where decisions must be made. FPGAs support precise control of sensors and actuators, allowing intelligent machines to safely and accurately interact with the physical world.
This is why physical AI requires heterogeneous computing. Rather than relying on a single compute engine, physical AI systems can combine CPUs, GPUs, NPUs, adaptive compute, and FPGAs, with each optimised for a different role across the complete sensor-to-action pipeline.
The Physical AI Era Plays To AMD’s Strengths
The emergence of Physical AI is not changing AMD’s technology strategy. It is validating a strategy the company has been building for decades.
Physical AI requires heterogeneous computing, and AMD delivers the industry’s broadest compute portfolios, spanning CPUs, GPUs, NPUs, Adaptive SoCs, and FPGAs, to proviode the right compute for the right workload.
Rather than optimising a single stage of AI processing, these technologies work together to support the complete physical AI pipeline—from perception and reasoning to real-time control and physical action.
Equally important is AMD’s long heritage in embedded computing. For decades, AMD technologies have powered systems across automotive, industrial automation, telecommunications, healthcare, aerospace, and other embedded markets where deterministic operation, long product lifecycles, functional safety, harsh operating environments, and customer co-development have always been essential.
These are precisely the same characteristics now becoming critical as physical AI moves from research into real-world deployment.
But physical AI requires more than silicon alone. Building intelligent physical systems demands AI models, robotics software, middleware, sensors, simulation, validation, and production-ready platforms working together.
That is why AMD is extending its long-standing open ecosystem strategy through initiatives such as the AMD Robotics Partner Network and AMD Kria™ AI Solutions, helping customers move more quickly from prototype to production.
Rather than adapting to the physical AI era, AMD is extending technologies, software, and ecosystems it has already been building for many years.
The Next Decade Of AI Will Be Defined By Systems That “Do”
Physical AI is not simply the next stage of AI. It represents a fundamental shift in how intelligence creates value.
The first era of AI was largely defined by systems that could understand information. The next era will be defined by systems that can safely sense, decide, and act in the physical world.
This shift is particularly meaningful for manufacturing-driven economies facing structural demographic change. Australia is one such example. Ageing populations, tightening labour markets, and growing pressure to improve industrial productivity are accelerating investment in automation, robotics, and intelligent infrastructure.
For these economies, physical AI is not simply another technology trend. It is becoming an important tool for sustaining long-term productivity, industrial competitiveness, and economic resilience.
Physical AI will not solve these challenges on its own. But by enabling intelligent systems to safely interact with the physical world, it has the potential to fundamentally reshape how industries operate.
AMD believes the next decade of AI will be defined not only by systems that know, but by systems that do.
