The physical world is emerging as a rare moat and the biggest market opportunity of the AI era.

The most valuable technology companies of the 21st century were mostly built behind a screen. Software ate whatever it could reach through a phone or PC, and venture capital paid the bill. That was the game. Build asset-light, scale fast, and defend with switching costs the database accumulated.

AI breaks the pattern, collapsing the cost of building software and extending capabilities into work that used to require humans. This shift is forcing us to look for new moats and unlocking markets software couldn’t reach. Both point to the same place: the physical world.

The physical economy, including industrials, manufacturing, energy, agriculture, logistics, construction, transportation, defense – represents north of $50 trillion of global GDP. Welding, physical security, shipbuilding, farming, even the actuators that make your seat on the airplane recline. These processes have seen little automation compared to their digital counterparts. A car today is built using more or less the same physical processes as it was four decades ago, while the screen on its dashboard runs computation that would have required a data center at the time.

This gap is closing as a few structural changes are happening at once.

Software can now operate on the physical world. Physical automation has historically been limited to controlled, pre-programmed environments: the robotic arm that moves on the same joint in the same spot, thousands of times a day. Anything unstructured, variable, or messy was out of reach. Foundation models have changed that. Machines can now perceive, reason, and act inside real-world environments, from a cluttered warehouse aisle to a chaotic construction site.

Hardware costs have collapsed to the point where physical automation pencils. Lithium-ion battery costs have fallen nearly 90% since 2013. Edge compute performance per dollar has improved by orders of magnitude; NVIDIA’s Jetson line delivers roughly 100x more AI performance per dollar than it did a decade ago. Industrial robot system costs have fallen substantially over the past three decades and are projected to keep falling through 2030. Automation that once required government-scale capital now pencils for the market.

The demand is here. By 2033, U.S. manufacturing may need 3.8 million new workers, with nearly 1.9 million roles at risk of going unfilled. The labor force is aging out while demand is accelerating, driven by the AI data center buildout and a reindustrialization push to reduce manufacturing dependence on China. Defense is compounding the pressure. Rising geopolitical tension and a battlefield increasingly fought with technology are pushing defense budgets higher and creating space for a new generation of software-first companies to define modern conflict. These structural shifts are pulling the demand curve forward.

The capital structure has caught up. Venture funding into physical AI categories is rising. Policy and government grants are opening more funding to startups. The debt stack has matured as equipment, inventory, and project financing give founders access to non-dilutive capital. Founders building in the physical world have options today that were not available a decade ago.

Physical AI has become necessary for industrial and national prosperity at the exact moment the technology and financing stack has caught up to deliver it. The next frontier will produce a generation of category-defining companies that use software to reinvent the real world.

A Map of the Physical World

Physical AI is a loose term. It gets used to refer to deep tech, robotics, hardware, and industrials. We define it as all of the above: the core processes that act on the real world rather than the digital one. The map below is a rough sketch of where innovation is clustering, and a deliberate oversimplification.Verticals are defined by end market, though plenty of companies play across more than one (defense and space, for instance, with dual-use applications). Altana, for example, is building the first real map of the global supply chain, mission-critical for both logistics giants like UPS and government agencies like U.S. Customs and Border Protection.

Robot brain companies are trying to build the foundation model for the physical world. The bottleneck here is data. LLMs had the internet, a vast and well-structured corpus of data. The physical world has no equivalent. Data here is scarce, expensive to collect, and the world it describes is harder to predict. The cost of getting a prediction wrong is also more severe; a hallucination in a warehouse or on a highway can kill someone. But the prize, if anyone builds it, is the biggest in software’s history.

The enablement ecosystem is early and broad. Micro1 is going directly at the data problem. Nominal is building the engineering backbone hardware teams need to ship. Several companies are emerging here, building the railroads for physical AI.

Lastly, the compute frontier itself is moving. Quantum is particularly exciting as the next major technological inflection following AI, with implications spanning materials discovery, drug design, and AI compute efficiency itself.

Two themes are worth pulling out, both of which would have read as red flags in the last cycle but are now central to the opportunity: hardware and the government.

Verticalization & Hardware as a Trojan Horse

The traditional venture playbook has treated hardware as a risk, but in the physical world it can be a distribution wedge and a better product.

Physical industries are analog and run by buyers who purchase things they can see and touch. Traditional OEMs move slowly and face little competition. Trying to sell software into these markets might not get you through the door fast enough to build a venture-scale business. Here, hardware becomes the delivery mechanism that carries your software into industries that would otherwise never adopt it.

Flock Safety is a good example: selling machine learning to legacy camera OEMs would have meant years of integration work and incremental adoption. Instead, they sold the camera itself, a form factor that public safety agencies already knew how to buy, with intelligence built in. Flock Safety was last valued at over $8B and has now built a unique ecosystem of products spanning multiple form factors across both public and private sectors.

Verticalization can also be an avenue for product differentiation. There is a limit to what off-the-shelf hardware can do; the performance ceiling is set by someone else’s roadmap. In physical markets, owning the stack can lead to dramatically better unit economics, a supply chain you control, and most critically a data flywheel that compounds as real-world operations inform how to innovate across every layer of the stack.

Tesla is the canonical example: by owning the vehicle, the battery manufacturing, and the software that runs both, every mile driven generates data that improves the next model, lowers cost, and widens their gap.

Hardware is not a solution to every problem. Field AI shows that being asset-light in the physical world can itself be a differentiation, as the problem they solve is defined not by a form factor but by a condition: messy, unstructured environments that are hard for robots to navigate. Focusing only on the software and staying asset-light lets them deploy across many industrial sites at the speed needed to build a data advantage that compounds over time.

The degree of vertical integration should work backwards from the problem and the customer. But for investors who haven’t had to think about hardware in a long time, it’s time to reassess.

Sovereignty & The New Industrial Base

Defense and national security have become key macro drivers reshaping the physical economy.

The American defense industrial base was built for a different era. A handful of prime contractors won the bulk of major awards with no real competition, and program timelines were measured in decades. That worked when the threat moved slowly and the U.S. held a clear technological lead. Neither is true anymore.

Ukraine has shown that a drone built for a few thousand dollars can destroy a tank worth millions. Cost, adaptability, and speed of iteration are the strategic pillars that decide who wins today.

Underneath all of this sits a harder problem: four decades of offshoring critical manufacturing left the U.S. dependent on China for the supply chains that make modern defense possible. Reshoring has become an explicit national priority.

Policy is reinforcing the shift. The CHIPS Act put over $50 billion behind domestic chip manufacturing. The Defense Innovation Unit and new contracting mechanisms have opened the system to startups in a way that would have been unthinkable a decade ago. U.S. defense spending crossed $1 trillion in FY2026 for the first time, and venture funding into defense tech roughly doubled last year, to $49 billion. Manufacturing itself is being rebuilt around this: U.S. construction spending on manufacturing facilities roughly tripled between 2020 and 2024 to $235 billion, and the industry is being treated as a core innovation vector for the first time since the Cold War.

A new generation of defense companies is emerging from this. Anduril is a prime example. The company invests its own capital to build finished defense products before any contract pays for them, ships them at fixed prices, and runs the whole portfolio on a single software platform that updates in the field. The primes do not work this way. Anduril was last valued at over $60B.

Another example, Saronic, shows how verticalization can be a major competitive advantage in defense-adjacent verticals, producing autonomous surface vessels in a purpose-built shipyard at a speed and price the Navy’s traditional procurement cycle has never matched. Chariot Defense attacks another layer: the power infrastructure underneath these systems. It builds software-defined battlefield energy for the drones and directed-energy weapons that today’s diesel generators were never designed to support.

National security is not a single vertical that happens to be having a moment. We believe it is a new normal in the American industrial base, and a powerful driver of the next generation of physical AI companies.

Conclusion

As the cost of cognition continues to collapse, we believe moats are being pushed to contextual data, distribution, and trust. The physical world is the next frontier where most of these moats compound and where most of the global economy already sits.

These companies each look different and break the pattern venture capital has relied on. Sales cycles can run longer, milestones are not as clean as a SaaS metric, capex sometimes comes before revenue, with no guarantee the product will work in the end, and the core risk is often technical, not market. But if it works, the products sell for tens of millions instead of hundreds of thousands, the moats are much harder to displace, and the markets are orders of magnitude bigger than anything software ever addressed.

The next generation of category-defining companies will use frontier technologies to drag the physical economy into this century. We are tracking this closely and are proud to be backing companies like Lila (scientific superintelligence), Altana (global network of trusted trade), and Avathon (industrial operations), among many others.

We encourage you to go deeper with recorded sessions from our most recent Montgomery Summit across robotics, defense tech, quantum computing, and adjacent themes.