Why Physical AI Is So Hard to Engineer: Predictability Becomes the New Optimization Target for System Design
Senior executives at Synopsys, Cadence, Renesas and other semiconductor companies argue that the core challenge of physical AI is not insufficient compute, but the need to maintain predictable behavior across the full sensing-compute-actuation chain. This requires hardware and software to be co-designed from the outset, and is changing how chips are designed and used.
Traditional AI infrastructure is often optimized for throughput, utilization, and scale. But when AI enters the physical world, the optimization target undergoes a fundamental shift. William Chen, product marketing group director for design IP at Cadence, frames the difference: 'Traditional AI infrastructure is often optimized for throughput, utilization, and scale. Physical AI systems must satisfy latency, power, reliability, safety, and environmental requirements while responding to continuously changing inputs. A correct answer delivered too late may be operationally equivalent to a wrong answer.'
This is not a gradual increase in engineering difficulty but a change in design paradigm. Alex Burlak, executive vice president of engineering and customer success at proteanTecs, illustrates the consequences with a humanoid robot: a robot may have a powerful central compute system serving as its 'brain,' but throughout its body there can be dozens of MCUs and other devices controlling motors, joints, sensors, power and communications. A timing issue or degradation in one of those devices may not simply mean an application runs a little slower. It could mean a joint responds differently than expected, a sensor input arrives late, or an actuator behaves outside its intended parameters.
Anthony Matarazzo, head of physical AI, digital twins, simulation-driven AI, and accelerated computing at Synopsys, captures the magnitude of this difference in one sentence: 'We are moving from generative AI to agentic workflow orchestration to physical AI. That is literally making intelligence actionable. But the consequences of being wrong fundamentally change. Life or death, plus or minus 1%.'
This predictability requirement leads to a judgment with substantive implications for the semiconductor industry landscape: hardware and software cannot be optimized separately. Brad Rex, head of physical AI and UX at Renesas Electronics, states: 'Because physical AI systems have to reconcile AI inference with deterministic, safety-critical control, the silicon and the software cannot be optimized separately and expected to integrate successfully. The system architecture, the real-time software, and the underlying compute and power hardware all have to be designed together from the outset.'
Chen at Cadence further explains why single-point improvements cannot solve system-level problems: 'Control-loop behavior emerges from interactions between sensing, memory access, inference execution, communication, and actuation. Improving software alone cannot overcome infrastructure bottlenecks, and faster hardware does not automatically improve system behavior if data delivery, synchronization, or timing remain constrained. The architecture must be considered end-to-end.'
This means the industry opportunity in physical AI cannot be understood solely from a chip-performance perspective.
Satish Radhakrishnan, head of semiconductor and electronics at Vinci, describes the unevenness of adoption: 'Each category is going to be separate and see slow growth, but there will be many sectors growing in parallel. Each requires new-age AI, new compute, and this will impact the speed at which it's going to develop. One area may take longer to grow deep, but before it fully penetrates one market segment, it will enter multiple new market segments. Together it's going to grow.'
This uneven adoption has specific implications for investment judgment. Frank Schirrmeister, executive director of strategic programs for system solutions at Synopsys, identifies where the growth inflection point lies: 'The growth inflection comes when you have this full loop of sensing, reasoning, physical action, and then autonomy. It's really a platform transition where everything is being elevated up to the system level.'
Another dimension of this platform transition is product lifecycle. Unlike mobile phones that are quickly replaced when a new model arrives, cars, robots, and drones have much higher initial costs and are expected to last many years. This requires systems to be upgradable—not just in terms of software or AI models, but also hardware, sensors, and actuators. Chen at Cadence identifies the key constraint: 'The challenge is that many infrastructure decisions become difficult to change once they are embedded into the system. Memory architecture, communication fabrics, and interface choices frequently determine the practical limits of future expansion.'
From an industry investment perspective, the core judgment this article presents is that the competitive dimension for physical AI chips is shifting from single-chip performance to system-level predictability assurance. This affects EDA companies (needing new verification methodologies), IP suppliers (needing foundational modules that support co-design), and system architecture companies (needing cross-domain integration capability) in different ways. Robert Day, director of go-to-market for physical AI at Arm, estimates that manufacturing, logistics, and transportation represent an approximately $200 billion annual compute opportunity in the 2030s, but whether this figure materializes depends on whether the engineering constraints described above are actually resolved.
It is worth noting that all interviewees' judgments come from their respective companies' product positioning perspectives—Synopsys sells design tools, Cadence sells IP, Arm sells architecture licensing, and proteanTecs sells on-chip monitoring and analytics solutions. Their consistent emphasis on the necessity of system-level co-design is partly because that very necessity creates demand for their respective products. The article provides no engineering validation data independent of these stakeholders, nor does it quantify a realistic timeline for the co-design methodology to reach production deployment.
Matarazzo at Synopsys raises a fundamental question that remains unanswered: 'We know that the Internet trained generative AI, but what's going to train physical AI? This is the big question that companies are asking. Reality has rules.' The answer to this question is not yet clear, and it may be one of the critical obstacles that must be resolved before physical AI can be deployed at scale.