The Physical AI Stack: Understanding the Full Opportunity

Physical AI is not just robotics with better software. It is a complete reimagining of how machines interact with the physical world — from how they sense and model their environment, to how they plan and execute actions, to how they learn from experience and compound that learning across deployments. The venture opportunity is correspondingly large. But mapping it requires understanding the full stack: not just the headline-grabbing robot companies, but the infrastructure, tooling, and enabling technologies that will determine which of them actually scale.

I've spent the past eighteen months building conviction across this landscape — meeting the founders, talking to the customers, and watching how the technology performs outside of demo environments. What I've come to believe is that the Physical AI stack has five distinct layers, each with its own competitive dynamics and investment characteristics, and that the value distribution across those layers over the next decade will look very different from where the current conversation focuses.

The Stack, Layer by Layer

The foundation is silicon and sensing. This includes the compute platforms for edge inference, the sensor hardware — cameras, LiDAR, force-torque sensors, tactile arrays — and the low-level signal processing that converts raw sensor data into structured inputs for higher-level systems. Players range from semiconductor giants to specialized startups like Hailo and SiMa.ai in inference compute, and Photoneo and Ouster in three-dimensional sensing. The moat here is technical: physics-limited sensor performance, TOPS/watt efficiency, and interface compatibility with industrial control systems are not easy problems to replicate quickly.

The second layer is perception and world modeling: the algorithms that transform sensor data into a structured representation of what exists in the environment and where. Computer vision, 3D reconstruction, scene understanding, and object recognition all live here. This is where most of the deep learning research of the past decade has been concentrated, and performance has improved dramatically. The remaining challenges are in unstructured environments and at the reliability thresholds that industrial deployment demands.

Above that sits planning and control: the systems that decide what to do given a world model and execute those decisions through physical actuators. Motion planning, manipulation planning, task planning, and real-time control belong here. This layer has historically been dominated by classical algorithms — optimal control, model predictive control, trajectory optimization — but learned approaches are rapidly taking over for complex manipulation tasks where the combinatorial space of possible actions is too large for classical search.

Where the Value Concentrates

The fourth layer is learning infrastructure: simulation environments, data pipelines, imitation learning from human demonstration, reinforcement learning from deployment feedback, and the training and evaluation tooling that supports continuous improvement. This is where the data flywheel lives. A robot company that can learn from every deployment and compound that learning across its fleet is a fundamentally different business from one that ships a fixed-capability system and moves on. This layer is underinvested relative to its importance.

The top of the stack is application software and deployment tooling: programming interfaces, remote monitoring platforms, fleet management, and system integration tooling. This layer is where most of the commercial activity has been concentrated historically, but where technical differentiation is hardest to maintain. It's also where the most acute near-term pain is — the firmware problem, the observability gap, the integration cost that keeps mid-market manufacturers locked out of automation adoption.

The Foundation Model Bet

Physical Intelligence (Pi), founded by Sergey Levine and Chelsea Finn from the Berkeley robotics and machine learning community, is building at the intersection of learning infrastructure and general-purpose robot control. Their π0 model — a large vision-language-action model trained on diverse robotic manipulation data — is an early attempt at a foundation model for robot actions. The thesis is that the same scaling dynamics that produced large language models will, with the right architecture and enough diverse training data, produce robot controllers that generalize across tasks and environments without per-task retraining.

Covariant, founded by researchers from OpenAI and UC Berkeley, is pursuing the same general direction but with a stronger emphasis on the logistics market and a more aggressive real-world data strategy. Their RFM-1 model is trained on data from hundreds of millions of picks across their deployed fleet. The flywheel compounds: more deployed robots generate more training data, which improves the model, which wins more deployments. It's a defensible position if they can maintain the deployment velocity that feeds it.

The Integration Question

One of the consistently underappreciated challenges in Physical AI is that the stack layers don't integrate cleanly out of the box. A perception system optimized for speed may produce outputs too noisy for precise manipulation planning. A motion planner designed for fixed industrial arms may not interface with a mobile platform's control architecture. An edge inference chip may not support the model architecture the ML team wants to deploy. These integration challenges are where much of the actual engineering effort in robotics goes, and they're one reason why vertically integrated companies — those that co-design and co-optimize across multiple stack layers — have tended to outperform best-of-breed assemblers in comparable application domains.

The Investment Angle

The Physical AI opportunity is not a single bet — it's a portfolio of bets across a stack that is being built out in parallel. The companies I find most interesting are those that own a layer with genuine technical moat, understand where their layer interfaces with adjacent ones, and are building toward a position that compounds as the rest of the stack matures. The worst investments in this space will be in companies that are one layer of a five-layer problem and are pricing themselves as if they own all five. The best will be in companies that understand exactly which problem they solve, solve it better than anyone else, and build from there.