Physical AI and the rise of robotics 

We believe “Physical AI” is arguably marking the next stage in the evolution of artificial intelligence (AI) – systems that can perceive, reason, and act in real-world settings, from warehouses and factories to hospitals and logistics networks. While much of AI’s early impact has been largely defined in digital environments – generating text, interpreting images, writing code, and supporting decision-making on screens – the shift to physical AI could be more demanding. It brings AI off the page and out of the browser, embedding intelligence into the physical systems, assets and infrastructure that underpin the economy.

In our view, this shift is drawing investor attention for good reason. Real-world deployment brings AI closer to parts of the economy where outcomes such as productivity, labour efficiency, precision, throughput, and operational resilience, can be easier to observe. It also introduces a very different set of technical and commercial requirements. A system working in the physical world must contend with force, motion, spatial awareness, timing, and uncertainty. Those demands widen the opportunity set, while also making the path to adoption more complex than many software-led AI use cases.

At its simplest, physical AI refers to AI systems that must operate within the constraints of the physical world. That means not only understanding language or images, but also friction, gravity, inertia, force, and cause and effect. As Nvidia has framed it, these systems require a “world model”, an ability to interpret how the world works before taking action within it.¹ That is what separates a large language model from a robot operating in an unstructured environment. One predicts likely sequences of tokens, the other must make judgments in real time, with consequences if it gets them wrong.

That challenge helps explain why physical AI has only now become a more serious discussion for investors. Recent progress in foundation models has improved machine perception and reasoning. At the same time, the supporting infrastructure for robotics is becoming more capable. Nvidia’s own description of the problem is useful here, not because it settles the matter, but because it highlights the scale of what is required. In Jensen Huang’s framework, physical AI depends on three linked computing layers: a training layer in the cloud, a simulation layer in a digital environment, and an inference layer on the robot itself. In practical terms, that means intelligence has to be trained, tested in simulation and then executed in real time.

For investors, the significance of that framework is twofold. First, it suggests that the opportunity is broader than any single robot manufacturer. If physical AI develops as expected, value is likely to accrue across the stack: semiconductors, high-performance computers, simulation software, sensors, actuators, control systems, and robotics original equipment manufacturers. Secondly, it underlines why commercial adoption may be slower and more uneven than the enthusiasm around AI headlines suggests. Physical deployment is difficult precisely because the real world is variable, unpredictable, and expensive to test in.

This is one reason simulation has become central to the investment case, with robotics developers increasingly using digital twin environments to train and refine behaviours before machines are deployed physically. Nvidia’s Omniverse is one example of this approach. The advantage is not only speed, but economics and safety. A robot can attempt to grasp an object millions of times in simulation before doing so in the real world, helping developers refine how much force is needed and how to respond to variation. That does not eliminate deployment risk, but it materially improves the training process.

We believe the near-term commercial relevance of physical AI is therefore likely to be strongest in industrial and logistics applications, where environments are relatively structured and the economic case can be clearer. Advanced robotics and automation are already more established than many other frontier technologies, with a stronger base of real-world use cases and revenue generation. For investors seeking frontier technology exposure, robotics – currently the most mature area – can provide a practical anchor alongside themes like fusion, quantum computing, or space, which offer significant upside potential but typically involve longer time horizons and less certain commercialization timelines.

The geographic dimension also matters. A growing body of research argues that China’s robotics expansion reflects more than a cyclical recovery. Sands Capital and Invesco both characterise it as a structural development, shaped by industrial policy, domestic manufacturing ambition, and an increasingly deep engineering talent base. That matters because China is not only a large end market, but also a major place of production capability and component development.

At the same time, some of the most durable opportunities may potentially sit away from the most visible end products. As Harding Loevner argues, humanoid robots may attract attention, but the more robust business models may lie in the enabling technologies beneath them: pneumatics, actuators, precision motion, sensing, and other industrial components.² For investors, that is an important distinction. The eventual winners in humanoid platforms remain uncertain, but the demand for critical sub-systems could prove more durable across use cases and geographies.

None of this removes the need for caution. Adoption timelines in robotics are frequently overestimated, especially when extrapolating from demonstrations to scaled deployment. The gap between a robot that can complete a task in a controlled setting and one that can do so reliably, safely, and economically in varied environments remains significant. Dexterity is one example. One challenge, the “hand-grasping” problem, highlighted by Nvidia sounds narrow, but it illustrates a broader truth: many tasks that appear simple to humans remain technically difficult for machines. Cost is another constraint. For physical AI to move beyond high-value industrial niches, the hardware stack, including sensors, actuators, and energy systems, may need to decline materially in cost.

There are also portfolio-relevant geopolitical and supply chain risks. US-China tensions may complicate access to some parts of the technology stack, and the conflict in Iran is adding further pressure, disrupting supply chains for critical minerals with downstream consequences for semiconductor manufacturing, which remains heavily concentrated in Asia. Logic, graphics and memory chips represent foundational infrastructure for AI development, making stable supply a strategic dependency for the sector's continued expansion. Meanwhile, the broader buildout of robotics and advanced automation could place additional strain on supply chains for critical minerals and specialized components. Those risks do not invalidate the theme, but they do argue for selectivity and for a value-chain approach rather than an overly narrow focus on a handful of headline names.

For investors, then, physical AI is best understood not as a single theme but as an emerging industrial ecosystem. Some parts of that ecosystem, such as industrial automation, simulation, and enabling hardware, are investable today. Others, particularly humanoid robotics at scale, remain earlier-stage and more speculative. The opportunity is potentially real, but it is unlikely to be linear. 

For investors with a 5-10 year horizon, we believe physical AI and robotics deserves a considered allocation within a diversified thematic sleeve.