Why the next era of AI advantage will depend on whole-system competence
Most of the AI breakthroughs that have shaped public imagination happened in environments designed for information. A prompt goes in, and text, code, an image, or a decision comes out. When something fails, we can often retry the task, update the software, or change the workflow.
The physical world gives us less room for abstraction.
A robot operating on a production line has to perceive correctly despite changing light, vibration, dust, people, and variation in the objects in front of it. It has to act within limits set by power, heat, mechanics, safety, cycle time, and cost. A capable model matters enormously, but the factory ultimately experiences the performance of the entire machine and everything connected to it.
BMW’s 2025 deployment of Figure 02 at its Spartanburg plant illustrates the difference. Over ten months, the humanoid robot supported production of more than 30,000 vehicles, moved more than 90,000 components, and accumulated roughly 1,250 operating hours. Bringing it into production required early involvement from production IT, occupational safety, process management, and shop-floor logistics. BMW also learned about the need for standardized interfaces, stronger 5G coverage, and revised safety concepts.
The robot’s capability was inseparable from the industrial system that allowed it to work.
This is why I have become convinced that the movement of AI into the physical world changes more than the applications of AI. It changes where competitive advantage lives.
In physical AI, the unit of performance is the system.
When intelligence meets physics
The scale of automation is already substantial. The International Federation of Robotics reports that 542,000 industrial robots were installed globally in 2024, more than twice the number ten years earlier, and annual installations exceeded 500,000 for the fourth consecutive year.
At the same time, advances in robotic intelligence are expanding what machines can perceive, reason about, and do. Yet the research itself exposes how different the physical world is from the digital one.
Google DeepMind’s Open X-Embodiment project brought together more than 20 research institutions to assemble experience across 22 different robot embodiments, more than 500 skills, 150,000 tasks, and over one million episodes. DeepMind described gathering sufficient diversity from the physical world as too resource-intensive for a single laboratory.
That observation is easy to overlook.
Digital AI benefited from enormous quantities of existing information. Physical intelligence often has to create its experience through robots, equipment, environments, supervision, simulation, and repeated interaction with reality.
Deployment adds another layer to this difficulty. ISO 10218-2:2025, the international safety standard covering industrial robot applications and cells, addresses integration across design, commissioning, operation, maintenance, decommissioning, and the interaction of robots with other machines and components.
These are not peripheral concerns. They help determine whether an intelligent system can create value at all.
Sensing determines what the system knows. Materials and mechanics determine what it can withstand. Power and thermal management constrain what it can do and for how long. Manufacturing determines whether it can be reproduced economically. Reliability, safety, integration, and maintenance determine whether customers can trust it enough to keep using it.
Model performance remains essential. Physical AI makes visible how much value depends on everything surrounding the model.
Industrial know-how is compressed knowledge of reality
This shift should cause industrial leaders to look differently at capabilities their companies have accumulated over decades.
An experienced industrial organization knows which tolerances actually matter. It knows how a component behaves after thousands or millions of cycles. It knows where heat accumulates, how vibration propagates, what contamination can do to a sensor, which supplier variation creates problems, and which rare failure mode can stop an entire production line.
It also knows something harder to document: how the real workflow differs from the process map.
That knowledge is distributed across engineering specifications, quality systems, service histories, test protocols, factories, supplier relationships, customer conversations, field failures, and the judgment of experienced operators and engineers.
I think of industrial know-how as compressed knowledge of reality.
For much of the software era, these capabilities could appear less strategically exciting than software, cloud platforms, digital networks, and data. Physical AI changes their strategic relevance because every intelligent physical system eventually encounters realities that industrial organizations have spent decades learning to manage.
Incumbency by itself offers little protection. Knowledge trapped inside organizational silos has limited value. Experience can also harden into assumptions that make a better architecture harder to see. Installed infrastructure can become a platform for learning, or a reason to preserve the past.
The opportunity is to make differentiated industrial knowledge usable.
That can mean translating field experience into better test conditions, exposing new technologies to operating environments earlier, turning failure histories into validation requirements, or building interfaces through which intelligent systems can safely interact with existing infrastructure.
The strategic asset is therefore larger than data alone. It is the combination of data, context, physical infrastructure, engineering judgment, customer access, and environments where a technology can learn what reality requires.
Frontier companies bring the complementary advantage
Entrepreneurs approach the same problem from the other direction.
A frontier company can question an architecture that an established industry has spent years optimizing. Its engineers can combine new models, sensors, compute, controls, actuators, and software without having to preserve every interface inherited from previous generations. They can move quickly because they have fewer historical assumptions to protect.
Industrial teams bring deep knowledge of which constraints are real. Startups bring the freedom to question which of those constraints can now be redesigned.
Neither capability is sufficient on its own.
An industrial company can understand the workflow in extraordinary detail and still miss a discontinuity in what technology now makes possible. A startup can build something technically remarkable and discover late that a seemingly minor issue in reliability, integration, serviceability, safety, or economics prevents adoption.
Working at the intersection of a global industrial company and frontier technology investing has made this complementarity increasingly visible to me.
The strongest collaborations give entrepreneurs access to difficult workflows, operating expertise, physical environments, and pathways to scale while preserving their ability to challenge the existing system. Industrial partners, in turn, gain access to architectures and capabilities that would be difficult to develop through incremental improvement alone.
The value comes from creating a productive interface between frontier innovation and industrial reality.
When the system becomes the investment thesis
This shift also has consequences for investors.
If physical AI changes the unit of performance, it should change what we diligence.
If the unit of performance is the system, the unit of diligence should become the system too.
A strong model or impressive demonstration remains important evidence. Investment judgment also has to reach further into sensing, hardware-software integration, reliability, manufacturability, power, safety, field service, supply chains, customer workflows, and deployment economics.
The central diligence question becomes broader: what has to be true for this technology to perform repeatedly in the environment where customers actually need it?
This also changes how we should think about defensibility.
A physical-AI company’s advantage may sit partly in algorithms or proprietary data. However, it can also accumulate through hardware-software co-design, knowledge of a difficult operating environment, years of failure data, certification, manufacturing processes, customer integration, or access to physical environments where the system can continue learning.
Some of these advantages take longer to build precisely because they cannot be created entirely in software.
Capital therefore deserves a more nuanced lens. Physical-AI companies may need funding for hardware iteration, testing, manufacturing, inventory, certification, deployment, and service infrastructure. Capital intensity by itself tells us relatively little about investment quality.A more useful question is what each increment of capital de-risks and whether that spending compounds an advantage that becomes harder for the next company to reproduce.
Capital intensity can become a moat or a tax. The difference is whether the capital builds differentiated capability.
That has implications for financing strategy. Milestones should increasingly demonstrate reductions in system risk, not simply growth in model capability. A technically impressive prototype, a reliable production deployment, and a repeatable deployment architecture represent very different stages of de-risking.
Physical AI can also change how we think about strategic investment and M&A.
Hyundai Motor Group’s acquisition of a controlling stake in Boston Dynamics offers an earlier example of this logic. Hyundai explicitly described the combination in terms of complementary strengths across robotics, manufacturing, logistics, construction, automation, and a broader robotics value chain. The transaction was therefore about more than owning a robotics technology in isolation. It was also about what those capabilities could become when combined.
That principle may become increasingly important.
In physical AI, strategic value can depend not only on what a company owns, but on what a partner or acquirer can combine with it.
An industrial company may contribute manufacturing scale, engineering expertise, customer environments, distribution, field service, or domain knowledge that materially changes a frontier technology’s trajectory. A startup may bring an architecture that allows those industrial assets to create entirely new value.
For investors, that means understanding the ecosystem around a company can become part of understanding the company itself.
Industrial leaders should inventory reality
This perspective also changes the questions executives should ask about AI.
Many strategies begin by inventorying technologies: Which models should we use? Which processes can we automate? Where should we deploy agents or robots?
Industrial leaders should also inventory reality.
Where does the company possess operating knowledge that outsiders would need years to reproduce? Which workflows expose technologies to unusually difficult conditions? Where do sensing, power, materials, reliability, manufacturing, or integration expertise create an advantage? Which field environments could help an intelligent system learn faster? Which customer relationships provide visibility into problems that remain poorly solved?
Those questions can reveal AI assets that never appeared on an AI roadmap.
They also change how deployment should be managed. Engineering, operations, manufacturing, safety, IT, procurement, and service cannot always be invited after a technology has demonstrated promise. In physical AI, their constraints often define the problem itself.
The World Economic Forum’s Global Lighthouse Network now includes 238 advanced manufacturing and supply-chain sites. In its June 2026 assessment, the Forum described AI moving from isolated pilots toward a core operating capability and found that the greatest impact comes when technological innovation is combined with operational fundamentals, workforce engagement, and clear strategic objectives.
The BMW example makes the same point at the level of one deployment. Moving from a robot demonstration to long production shifts touched communications infrastructure, safety systems, production processes, logistics, interfaces, and employees. The achievement belonged to the system.
This has an important implication for how industrial companies engage frontier innovators.
The objective is to expose the hard constraints early without forcing a startup to inherit the incumbent’s architecture. Give entrepreneurs access to reality while preserving enough freedom to redesign what reality no longer requires.
That balance is difficult. It may also become one of the most valuable capabilities an industrial company can build.
The next era of industrial technology leadership
There was a period when “technology company” and “industrial company” seemed to be moving further apart.
Physical AI may bring them closer together.
As intelligence becomes embedded in machines and infrastructure, the boundaries among AI models, software, semiconductors, sensors, materials, energy, mechanics, manufacturing, and operations become harder to separate. The performance that matters emerges from how those layers work together.
This gives industrial companies an opportunity, but no entitlement.
Their accumulated knowledge becomes strategically valuable only when they can translate it into an advantage for the next architecture rather than a defense of the previous one. Frontier companies face the mirror-image challenge. Technical brilliance becomes enduring value only when it can survive customers, factories, economics, safety requirements, maintenance, and physics.
Investors will increasingly have to understand both worlds. The strongest companies may not always have the most impressive intelligence in isolation. They may be the companies that assemble a system in which intelligence, hardware, industrial knowledge, economics, and access to reality reinforce one another.
The first chapter of modern AI taught machines to work with a world we had already digitized. This next chapter increasingly asks intelligence to operate inside a world that cannot be reduced to tokens, pixels, or APIs.
That is why I believe industrial know-how is becoming strategic again.
For industrial leaders, the important question is no longer whether decades of accumulated knowledge can remain relevant in the AI era. The more consequential question is whether that knowledge becomes part of the reason AI can enter the physical world at all.
In physical AI, intelligence earns its value as a system, in reality.
