The first wave of Artificial Intelligence (AI) systems read, wrote and reasoned over text, images and code. We believe the next wave of AI will move beyond software workflows into the physical world via machines that perceive, decide and act in real environments, from robots and autonomous vehicles to drones and other embodied systems.
This report examines where, in our view, Physical AI is already expanding into labor-, logistics- and asset-intensive operations that have historically been difficult to automate, and where we believe significant disruption and value creation are likely to occur next. In terms of potential economic impact, Nvidia estimates that physical AI has the potential to affect $50T USD of economic activity in manufacturing and logistics alone.
In this paper we analyze why we believe physical AI should be understood as both a new frontier of model capability and a broader stack that includes hardware, data, evaluation, observability, safety, deployment operations and outcome systems alongside the model layer itself.
Access the Report
Report Highlights
1. Why Now
Robotics and physical AI startups raised approximately $27.6 billion across 1,009 VC deals in 2025, more than double the $13.7 billion invested in 2024. In our view, the emergence of physical AI reflects the alignment of four forces:
- capability reaching action rather than just language and perception,
- deployment infrastructure maturing enough to operate outside controlled lab environments,
- data flywheels beginning to form, and
- capital following into Physical AI companies.
We explore why we believe physical AI is entering a meaningful proving period, where the question is no longer whether physical AI can produce compelling demonstrations, but whether physical AI companies can convert them into repeatable, economically attractive deployments.
2. Competing (and Complementary) Technical Paths
We map the different technical bets companies are making on how physical AI systems are built: modular robotics stacks with well-defined perception, planning and control layers; vision-language-action (VLA) models that connect perception, instruction and action end-to-end; world modeling approaches that predict future states to support planning; and skill-based orchestration patterns that sit atop these architectures. We examine why, in our view, the decisive variable is which companies can learn fastest from live deployments in a constrained operating environment, not which architecture looks most elegant in theory.
3. From Demos to Deployments
We analyze why the difference between a physical AI system that demonstrates well and one that deploys well is, in our view, important. We explore why the strongest evidence usually sits one layer beneath the visible surface: fewer interventions, better recovery behavior, faster deployment times and stronger evaluation discipline, and why metrics like intervention rate may be misleading if a company narrows its operating domain rather than improving the underlying autonomy.
4. Enabling Systems and Where Value May Accrue
Beyond robot OEMs and full-stack application companies, we examine where we see the potential for value to accrue as the category matures: robotic foundation models that attempt to provide reusable intelligence across embodiments; data infrastructure and evaluation tooling; observability and fleet operations; and safety and cybersecurity as core infrastructure rather than compliance overhead. We use this analysis to argue why physical AI is an expanding stack with the potential to create multiple points of investment relevance over time.
5. The New Physical AI Stack
We introduce a framework where we believe value is forming across physical AI: hardware and edge compute, which must operate under tight latency, power and safety constraints; data, simulation and evaluation, where measuring task success and real-world robustness may matter as much as the model itself; the operating layer of developer tooling, observability, fleet management and incident response; safety and cybersecurity; and, at the top, the applications and outcome systems that solve concrete operational workflows. We use this stack to examine why physical AI is, in our view, not simply a model market but a systems market.