Our framework for understanding one of the largest technology shifts of the next decade, breaking down the market and where 250+ companies sit within it
Over the past several years, the focus of AI has mostly been about automating the digital world. Models trained to generate text, write code, summarize documents, and answer questions. At defy.vc we have been excited by the next wave of AI and its impact on the physical world as well. Cars are driving themselves. Robots are picking product in warehouses. Drones are inspecting power lines. And it is just the beginning.
It feels like we are at a critical point for Physical AI. Nvidia CEO, Jensen Huang, recently described physical AI as being at the “ChatGPT moment” for the physical world: the point where machines stop just answering questions and start perceiving, reasoning, and acting in the real world. Investors seem to agree. Physical AI and robotics startups have attracted tens of billions of dollars over the last 18 months, having raised $47.4 billion in the first half of 2026 alone, according to Crunchbase. This is nearly 4x what the category raised in the second half of 2025, and more than the entire 2022–2024 period combined.
And yet, despite the excitement, the category is surprisingly hard to define and decompose. “Physical AI” has become an umbrella term for everything from humanoid robots to autonomous vehicles to simulation software to the foundation models underneath it. Companies operating at different layers of the stack are often discussed as though they’re direct competitors. To build defy’s own point of view, we built a market map of the space to test our thinking and decided to share it here. To date, conventions around categorization have yet to be standardized or have not kept up with the speed of change. This post is our attempt to explain both why physical AI is having its moment right now, and the framework we used to organize the 250+ companies on our map.
What Is Physical AI?
We define “Physical AI” to mean AI systems that can perceive, reason about, and act in the physical world.
Hardware automation isn’t new. Industrial robots have existed for decades. Autonomous vehicles have been in development for years. The primary change is the intelligence layer underneath. There is a level of flexibility and adaptability to tackle less structured problems that wasn’t possible previously. Instead of hand-coding a robot with explicit rules for every scenario, we’re increasingly training systems that learn, adapt, and generalize. As a result, the dawn of Physical AI is a shift from machines that execute instructions to machines that perceive, reason, and act. It is the emergence of systems that can generalize.
The same way large language models can answer questions they were never explicitly trained on, physical AI systems are beginning to perform tasks, adapt to new environments, and solve problems without being individually programmed for each scenario. We’re still in the earliest stages of seeing this work in production, but the potential impact on both individuals and industries is already incredibly exciting.
Why Now?
Three things appear to be converging simultaneously.
It’s Now Existential for Companies
There isn’t enough labor to do all the work a customer needs executed at a price point sustainable for the business. Across manufacturing, logistics, agriculture, healthcare, construction, and infrastructure, labor shortages continue to worsen. Many industries face aging workforces and growing difficulty filling skilled roles.
Historically, automation could only address a narrow set of highly structured tasks. Physical AI expands the universe of tasks that can potentially be automated. It is no longer only about productivity or cost savings but also about filling the labor supply gap.
The Models Are Generalizing
Companies like Physical Intelligence and Skild AI are attempting to build models that can transfer skills across tasks, environments, and robot types. Rather than training a robot to do one job, these models are able to operate across embodiments. For the first time, intelligence and embodiment are becoming separate layers.
The Infrastructure Stack Finally Exists
A decade ago, building a physical AI company meant building nearly everything yourself. Today, entrepreneurs can leverage increasingly mature ecosystems of simulation tools, synthetic data platforms, robotics infrastructure, foundation models, and deployment software.
Applied Intuition can help generate and test millions of scenarios before deployment. NVIDIA Omniverse provides realistic simulation environments. Formant and Foxglove help operators deploy, manage, and monitor robotic systems.
Just as cloud infrastructure accelerated software startups, the emerging physical AI infrastructure stack is accelerating robotics companies.
Underpinning all the excitement is a belief that physical AI won’t just automate existing labor but that it’ll expand what’s economically possible to build, staff, and ship in the physical world altogether. We believe the question isn’t really whether physical AI happens. It’s where the value accrues.
How We Drew the Map
We debated a few ways to segment the market: by form factor (humanoids, drones, AVs, industrial arms), by industry (defense, logistics, healthcare, agriculture, manufacturing), or by technology (models, simulation, sensors, software).
Each of those approaches felt incomplete. Organizing solely by industry obscures the horizontal platforms that increasingly power multiple verticals. Organizing solely by technology or form factor ignores the reality that a defense robot, warehouse robot, and surgical robot are fundamentally different businesses with different customers, sales cycles, and economics.
Ultimately, we landed on a framework that breaks the market most broadly across horizontal and vertical lines. Horizontal companies provide capabilities that can be leveraged across many industries and form factors. Vertical companies package those capabilities into solutions customized for customers needs in a specific industry. While there may be companies that could arguably be placed in more than one segment, we did our best to categorize them according to the application or use case we believe they are leading with.
Horizontal
Models – These companies are building the brains or foundation models to build a general intelligence layer for perception, reasoning, and action across many robots and tasks. World Labs, General Intuition, and Dyna are among a new generation of startups attempting to create general-purpose models capable of understanding and acting in the physical world. Meanwhile, NVIDIA, OpenAI, and Google DeepMind are making significant investments in their own physical AI initiatives.
Data + Simulation – unlike language models which had a much richer jumping off point for training thanks to the vast internet, Physical AI is fundamentally more constrained by data. Physical systems need experience. They have to learn from interactions with the real world. This data is expensive, slow, and hard to scale by comparison. Waymo spent billions of dollars and over 15 years collecting real world driving data starting in 2009. Today there is a growing ecosystem around simulation, synthetic data, and digital twins: Scale AI, Mecka, micro1, and Lightwheel are all, in one way or another, betting that if data is the fuel for physical AI, simulation is the most scalable way to generate it.
Inference, Observability, Ops – every technology platform eventually develops tooling. Companies like Reflex, Openmind, Viam, and Intrinsic help developers deploy, monitor, debug, and manage increasingly complex fleets of autonomous systems.
Full-stack – Some companies are building both the brain and the body. Figure, Apptronik, 1X, Tesla, and Boston Dynamics represent variations of the full-stack approach, where the intelligence layer and robotic platform are developed together. Whether the market ultimately favors integrated systems or more modular ecosystems remains one of the most important open questions in the category.
Vertical
Customers don’t buy physical AI. They buy lower labor costs, higher throughput, better safety, or a task done that they couldn’t previously staff for. As a result, we organized vertical companies primarily by the way the buyer actually experiences the market. We believe that some of the largest opportunities exist in industries that have historically been difficult to automate. These sectors combine labor constraints, large economic value, and highly variable environments, making them ideal candidates for physical AI.
We have twelve verticals: industrial/manufacturing/warehousing, delivery, autonomous vehicles, defense, consumer/home, hospitality, healthcare, life sciences/lab, construction, utilities/energy, space, and agriculture/environment.
Of the verticals above, we wanted to expand on a few:
Industrial, Manufacturing & Warehousing – This is one of the most mature markets in physical AI, which is clear visually as well. The environments are structured, the labor shortage is acute, and companies like Symbotic and Dexterity can often point to ROI measured in months rather than years. Facilities can be retrofitted or designed from the ground up to be more compatible with physical AI while actively distancing humans from the automation for their safety if needed. For example, Dexterity’s robots can autonomously manage the movement of different sized parcels across truck loading, unloading, palletizing, and sorting.
Transportation & Autonomous Vehicles – In many ways, autonomous vehicles were the first large-scale deployment of physical AI. Waymo demonstrated that autonomous driving can work commercially at scale. Companies like Aurora, Waabi, Kodiak Robotics, and Zoox are pursuing similar opportunities across trucking, logistics, and mobility. While the market has experienced periods of hype and disappointment, transportation remains one of the largest potential categories.
Defense – Defense has rapidly become one of the most active segments in Physical AI. Autonomy is increasingly viewed as a strategic capability both for safety and speed, making defense one of the fastest-moving adoption environments given the current global geopolitical climate. Anduril helped redefine the category by integrating autonomy, software, and hardware into a unified defense platform. Shield AI, Saronic, and Skydio are applying similar principles across autonomous, maritime, and aerial systems.
Healthcare – Healthcare automation is early but compelling. Intuitive Surgical demonstrated decades ago that clinicians will adopt robotic systems when they improve outcomes and workflows. A new generation of companies is now bringing AI-driven intelligence into healthcare workflows and expanding the range of tasks that can be assisted or automated. For example, Akara is enabling operating rooms to be more intelligent during procedures.
Agriculture, Construction, Energy & Beyond – Agriculture, construction and energy represent some of the most relevant opportunities for automation based on the tailwinds in each market. These industries are growing quickly especially domestically, face persistent labor shortages, operate in challenging real-world environments, and involve repetitive tasks that have historically been difficult to automate.
In agriculture, companies like Carbon Robotics are using AI-powered systems to perform precision tasks such as weed control, helping farmers reduce chemical inputs while improving productivity. Over time, autonomy could expand across planting, harvesting, monitoring, and broader farm operations.
Construction presents a similarly large opportunity. Companies like Built and Bedrock are bringing autonomy to heavy equipment, allowing repetitive and hazardous tasks to be completed more safely and efficiently. As intelligence improves, physical AI has the potential to automate portions of site preparation, excavation, material movement, and infrastructure maintenance.
Where We’re Watching Closest
Most of the public attention in Physical AI is focused on humanoids. They’re the most visible and futuristic expression of what the category could become. But the more time we spent studying the market, the more interested we became in the layers underneath: the data and simulation infrastructure, the inference and observability tooling, and the orchestration software required to deploy and manage autonomous systems at scale.
These are the picks-and-shovels businesses that can benefit regardless of which robot, vehicle, or platform ultimately wins. We’re also excited about enabling technologies across a range of form factors from robotic arms and sensors to other systems that generate and capture data that doesn’t exist cleanly today. In many cases, these technologies can make existing environments more efficient long before fully autonomous systems become ubiquitous.
Like any market map, ours is a snapshot, not a census. We placed companies where we thought they most honored the actual bet they’re making, but plenty of them span more than one box, and we’ve almost certainly missed a few that belong here. We’re also still very early in the evolution of this market – the structure is being written in real time, new companies are launching every week, and whoever looks like the leader today may not be the leader in five years. For example, we initially attempted to categorize vertical players by whether they were software/OS, data, or full stack oriented. While we believe the market is moving in this direction, the market is still too nascent to do so. We’re excited to continue to learn and evolve alongside this rapidly changing market. If you’re building something that should be on this map, or think we got a category wrong, tell us. We’ll keep it updated.
If you’re building in physical AI, we’d love to hear from you. We’ll continue updating the map as the ecosystem evolves.
https://makecents.substack.com/p/from-models-to-machines-the-physical?shem=aimgspc,
