Physical AI is the infrastructure, intelligence, workflows and machines through which artificial intelligence performs physical work.

Where judgment lives

For a century, machines couldn't judge, so industry engineered around it.

The thinking happened up front. It was frozen into fixed layouts, jigs, tight tolerances, and steps simple enough for a person to repeat. Whatever that rigid process couldn't handle got caught afterward, by someone at an inspection gate.

That arrangement is expensive in both directions. A rigid line can't handle variation, and the inspection gate makes quality a headcount.

Power had the same shape once. Factories ran on central line shafts, one engine's power carried by belts to every machine on the floor. Electrification did not change what power was. It changed where power could afford to be. Once electric motors got cheap enough to put one inside each machine, the line shafts came out and the floor opened up.

Physical AI is doing that to judgment now. It moves the decision inside the machine. Each unit gains the ability to perceive variation, decide what matters, and adjust as it works.

Source note: the electrification history draws on Paul A. David, American Economic Review 80(2), May 1990.

Why now

Three things arrived at the same time.

Labor is getting harder to source, and the shortage looks structural rather than cyclical. Aging populations, tighter immigration and reshoring arrived in the same decade. When a farm cannot hire a crew, the crop stays in the field. A robot bought to fill an empty shift is a different purchase from a robot bought to cut a wage bill, and it is the purchase we now see being made.

Inference is getting cheap enough to run inside the machine. Measured as intelligence per watt, meaning task accuracy per unit of power, local inference improved 5.3 times between 2023 and 2025. Over the same period, the share of real-world queries a local device could answer rose from 23.2 percent to 71.3 percent. Those figures measure laptops answering chat queries, not robots. We read them as the electric motor moment for judgment: the thinking is now cheap enough to put one inside each machine.

Policy has started to treat productive capacity as strategic. In July 2026 the Federal Communications Commission added foreign-produced advanced robotic devices to its Covered List. New models of those robots can no longer obtain the equipment authorization they need to be imported, marketed or sold in the United States, short of a case-by-case conditional approval. The measure looks forward only: models authorized before it keep their authorization, and fixed industrial arms sit outside it.

Source note: the intelligence per watt figures are from Jon Saad-Falcon and others, Intelligence per Watt: Measuring Intelligence Efficiency of Local AI. The Covered List additions were announced on 28 July 2026; the Commission’s own FAQ on the robotics and power inverter listings sets out what is and is not covered.

What the market gets wrong

Four objections, and what each one misses.

That conclusion extrapolates from a handful of headline deals, humanoid rounds and foundation-model rounds, to the entire category.

But valuations and commercial maturity vary enormously across robotics, and the loudest deals aren't representative of the field. A focused machine built to solve one valuable problem can be less visible than a humanoid demo. It can be less capital-intensive, and closer to commercial readiness than the headlines suggest.

Robotics isn't one market. The overvaluation story comes from treating it like one.

Many improvements in models and components will commoditize. That part is probably right.

But the real moat sits somewhere else: in what a working deployment accumulates over time.

• Customer knowledge
• Integration
• A certified safety case
• Service infrastructure
• Field data recorded under real conditions
• Technicians and spares within driving distance
• A machine calibrated to that plant

None of that lives in the model, so an upstream vendor can't reach it by shipping a better version. A model vendor can match a capability in a release. It cannot acquire ten seasons of weeding under real weather.

This objection is often right. A lot of the robotics market genuinely is too slow, or needs too much capital, to work as a venture return.

So we test capital strategy and timing before investing. What makes a deal work anyway is the path in: a focused deployment can enter through one paid workflow, doing one job well enough that a customer pays for it now. From there it can expand into a recurring service or a larger product suite, and create a large outcome by capturing a small share of an enormous physical-work budget.

Specialists can commercialize earlier than a general-purpose platform, because they only have to solve one workflow well, not every workflow. That lets them own valuable workflows and build the kind of exit value a strategic buyer pays for, before the market converges on one form factor, if it ever does.

That's the case for not calling the race now. General-purpose bodies still have a place, but they earn it: where the form factor is genuinely necessary for the job, and where the added complexity is worth what it costs.

What would prove us wrong

We'd be wrong under a few specific conditions.

The thesis fails if deployment produces no proprietary learning, if marginal integration costs don't fall, or if most of the value flows upstream to model and component makers instead of staying with deployment.

It also fails if technically impressive companies can't reach reliable commercial deployment within 10 to 12 years. That timeline means a poor return, even from real technology. It fails if manufacturing, service and financing eat the ownership and margins venture returns need. And it fails if the labor shortage eases, through migration policy, wage deflation or demographic change, because necessity, not cost, is what makes deployment urgent.

A machine that works eventually can still be a bad venture investment today.

What we don't invest in

Saying no is as useful as saying yes, so here's the list.

We don't invest in generic software or AI, the kind that plays no direct, meaningful economic role in an actual physical system. We don't invest in funds of funds, in SPVs as investment targets, or in public-market strategies. We pass on businesses that don't have a specific physical workflow, a real customer, or a credible path to a venture return, however interesting the technology looks.

And a program that exists only for defense, with no credible commercial use or civilian use by an allied government alongside it, is outside what we do right now.