Physical AI unmistakably dominated CES 2026, emerging as a dominant frontier of applied AI. It came to life in the form of robots (making cocktails, kickboxing, and folding origami), wearables, sensors, machines and environments.
Between Serve Robotics scaling from 100 to 2,000 autonomous delivery robots deployed across U.S. cities in a single year, Boston Dynamics’ Stretch robot unloading over 20 million boxes in warehouse operations globally, and Hyundai announcing production plans for 30,000 Atlas humanoid units annually by 2028 for automotive manufacturing, the appetite for production-scale physical AI is accelerating rapidly. And with this comes a demand for a fundamental rethink of AI architecture.
Why is physical AI different?
Large language models, centralised training, and cloud inference have dominated the recent AI wave, with human interaction through text and voice.
Physical AI is different in that it generates continuous, high-frequency sensor data, fuses multiple modalities in real time, and operates under hard constraints around latency, power, and safety. Critically, physical AI acts, rather than just responding.
Inevitably, this means that AI systems take on agent-like characteristics. They must maintain state, make decisions over time, and balance short-term reactions with longer-term goals. This is a very different notion of “agentic AI” than cloud-based language agents. In physical systems, agency is shaped by real-world constraints – like physics, latency, energy, and safety – and emerges from tightly orchestrated perception, control, and planning loops rather than open-ended autonomy.
Why is synthetic data important in physical AI?
The challenge with physical AI is that it quickly runs into a data ceiling. Real-world data is slow to collect, expensive to generate, risky to explore, and often impossible to capture at the scale or diversity modern models require. Increasingly, training happens in simulation (also referred to as ‘dreaming’), where physics-based worlds, sensor emulation, and domain randomisation expose models to failures, edge cases, and dangerous scenarios that may occur very infrequently, and where the consequence of failure is expensive and destructive. These synthetic environments don’t replace real data so much as amplify it, powering cloud-scale foundation model training, on-prem digital twins for site-specific validation, and edge-level adaptation in the field. In physical AI, the advantage is moving from who has the biggest models to who has the most accurate simulated worlds, and the tightest feedback loop between simulation and reality.
As Paul Heathcote, Stelia’s VP of Applied AI, observed at CES, “We’ve spoken before about the downsides of using synthetic data when it comes to LLMs, but the value of synthetic data in training robots is vastly different. We can create virtual environments using technologies like Unreal Engine with realistic physics, so robots can learn about how to interact with real-world objects, get trained on tasks, and learn how to interact with other robots without causing damage in the real world, so when they’re deployed in the real world, they’ll be safe.”
Edge AI: no longer just shorthand for cars and phones
Edge is becoming the default execution layer for physical AI, with intelligence being embedded directly into physical systems that can’t afford the latency, bandwidth, or privacy trade-offs of sending all data to the cloud. Edge AI is showing up in robots that navigate human spaces, where perception and control must happen in milliseconds; in wearables where personalisation and privacy are vital; and inside factories and warehouses where uptime is a key success marker.
The resurgence of on-prem
As edge deployments scale, the need for on-prem AI is increasing as a practical response to data gravity and operational complexity – alongside the need for low latency, especially in safety-critical scenarios. When hundreds or thousands of edge systems are generating high-volume sensor data, robust coordination is required.
On-prem systems increasingly serve as that coordination layer, aggregating local data, running simulations and digital twins, managing model updates, and enforcing compliance. In physical AI architectures, on-prem is critical for keeping the whole system coherent.
That said, the cloud does remain indispensable. But it’s being used differently.
Cloud platforms continue to excel at training foundation models, running large-scale experiments, and optimising learning across geographies. They are where long-term learning happens and where insights from many physical environments are distilled into shared capability. Combining edge, on-prem and cloud ensures real-time perception and control, coordination, and learning can be combined to form a powerful AI system for physical AI.
A hybrid approach was demonstrated by Fujitsu at CES, where sensor data is collected and shared with a centralised world model which makes predictions and then shares these predictions with AI in physical devices like robots and drones or cars. This offloads the really complex processing to big infrastructure, and lets the local AI deal with how to react to the prediction.
Physical AI and the future of AI architecture
Physical AI is reshaping AI architectures. Latency, energy, safety, privacy, data and governance have become the primary design constraints for building and scaling deployments successfully. The organisations that succeed in physical AI won’t just have the best models, they’ll build better systems that span edge, on-prem, and cloud in ways that enable the safe and effective deployment of physical AI in real-world environments.
