Here’s the problem with re-shoring humanoid robots:

A robot that costs about $46,000 to build using today’s global supply chain costs roughly $131,000 if you remove Chinese suppliers from the bill of materials, according to a new report by OpenMind.

That’s a 2.85X cost increase.

Until recently, that was mostly an interesting scenario to model. Now it matters a lot more.

On July 28, following a national security determination from the White House one day earlier, the FCC added foreign-produced advanced robotic devices to its Covered List. Their definition is fairly broad: warehouse AMRs, delivery robots, inspection quadrupeds, humanoids, and essentially other network-connected autonomous ground robots over two kilograms.

New foreign-produced robots covered by the rule can no longer simply receive FCC authorization for the US market.

Which means the supply-chain thought experiment is becoming an industrial strategy problem.

“It’s not very useful for the US to have robots that cost 3x more than anywhere else,” OpenMind founder and CEO Jan Liphardt told me on the Humanoid Daily podcast. “That immediately puts us at a disadvantage.”

Liphardt is a Stanford bioengineering professor and the founder of OpenMind, which is building an open, hardware-agnostic software layer for robots — essentially an operating and coordination layer for physical AI.

OpenMind has also just published its Physical AI Readiness Index, ranking 58 countries and 22 metropolitan clusters on their ability to build, supply, power, and operate advanced robots at scale.

And the results make the US problem pretty clear.

• China scores 82.9
  
• Japan is second at 69.6
  
• The United States is third at 69.0
• South Korea sits at 66.9
• Germany is at 66.1
  

China’s 13.3-point lead over Japan is larger than the gap between second place and fourteenth. That matters because physical AI isn’t just AI.

It's also physical.

America has the brains. China has the parts.

This is an overgeneralization and therefore not entirely true, but it's directionally accurate. The US dominates two important categories.

It scores 91.6 in compute and 78.7 in human capital, leading the field in both.

Then you get to materials and supply chain.

The US scores 48.5, but China scores 95.

That gap may ultimately matter more than who has the best foundation model.

“Many people think when they think of physical AI is that physical AI is all about AI,” Liphardt said. “There’s this notion of if you have the best models, then you’re also going to win physical AI. But based on what we can tell, that’s not true.”

There’s another problem with betting everything on AI leadership: models diffuse. China is increasingly open-sourcing advanced AI models, including models relevant to robotics and physical AI. They are getting better and better, quickly approaching top foundation models for quality.

That compresses the software advantage.

And the hardware advantage is much harder to replicate.

The bottleneck isn’t chips

It’s tempting to frame robotics supply chains through the semiconductor lens, but humanoid robots aren’t primarily a semiconductor supply-chain problem.

Actuators can represent 40% to 60% of a humanoid robot’s bill of materials. Semiconductors are closer to 10%. And many of those chips aren’t leading-edge 2nm or 3nm processors. They’re mature-node components built on processes such as 28nm, 65nm, or 90nm.

The difficult components are much more mechanical:

• strain-wave reducers

• cycloidal reducers

• planetary roller screws

• frameless torque motors

• precision bearings

• encoders

• rare-earth permanent magnets

These are the things that turn AI decisions into precise physical movement, and this is where China has built enormous industrial depth.

China produces about 94% of the world’s sintered neodymium permanent magnets and refines roughly 91% of global rare earths, according to OpenMind. The report identifies magnets as perhaps the tightest single constraint in the physical-AI supply chain.

Chips are pretty widespread. Battery tech and even production is possible in many places. But magnets are tough.

Japan doesn’t solve the problem either

There’s another wrinkle : the FCC restrictions aren’t simply a China ban.

The FCC's domestic-content calculation is tied to the Buy American Act standard: 65% domestic component cost through 2028, rising to 75% in 2029.

That means sourcing critical components from Japan, Germany, South Korea, or Canada doesn’t necessarily solve the domestic-content problem.

And Japan is extremely important to robotics.

OpenMind estimates that Harmonic Drive Systems controls somewhere between 71% and 85% of the global strain-wave reducer market. Nabtesco says it supplies roughly 60% of medium and large industrial robot joints. Those two Japanese companies sit deep inside the global robotics supply chain.

So one obvious strategy — manufacture the robot in America while continuing to import world-class actuators and reducers — may still leave manufacturers short of the required domestic-content threshold.

That makes reshoring much harder than simply putting a robot assembly plant in Texas.

Shenzhen’s even bigger advantage: everything is close

Possibly the most useful finding in OpenMind’s report is geography. Or, more precisely, proximity.

Eight of the ten highest-ranked robot-building metro clusters are in East Asia. Shenzhen–Dongguan–Guangzhou ranks first with scores nearing the 90s. The US's top region is the San Francisco Bay Area, and it ranks ninth — and scores just 32 on materials.

Why?

Density of companies. Think supply web, not just supply chain.

“One of the things that China has done very well is spatially concentrate everything that you might need,” Liphardt said. “Within about 10 kilometer radial distance, you have everything you could possibly need from carbon fiber to CNC to bearings to sensors to PCBs.”

That has a direct effect on engineering velocity.

If a robotics team needs a redesigned gear, a new motor housing, a revised PCB, a different bearing, a machined bracket or a new cable harness, the supplier may be nearby. The engineer can visit the factory, inspect the part, change the design and run another iteration quickly.

In the United States, many of those same capabilities exist, but they are often separated by hundreds or thousands of kilometers.

"We certainly have a lot of the pieces potentially here in the U.S., but they're typically in very different places," Liphardt said. "And those pieces that we do have are typically very small and there aren't a lot of choices."

That distinction is important. America’s problem is not simply manufacturing capacity. It is supply-chain topology.

A robotics ecosystem works better when motors, reducers, bearings, machined parts, electronics, sensors, tooling companies and engineering teams are concentrated in the same region. That reduces lead times, lowers logistics costs and, most importantly, increases iteration speed. Liphardt compares the model to Detroit in the middle of the twentieth century, when automakers and huge networks of specialized suppliers developed in the same industrial region.

Today, OpenMind ranks Detroit-Ann Arbor 21st in physical AI cluster readiness. 

Austin is 16th. 

Boston is 12th.

Ukraine shows why iteration speed matters

Ukraine offers a very different example.

It ranks 51st out of 58 countries in OpenMind’s overall capability score. It has little semiconductor manufacturing, no major rare-earth processing base and an electrical grid operating under wartime conditions.

And yet Ukraine has become exceptionally good at producing and iterating unmanned systems. Necessity, as they say, is the mother of invention.

"Ukraine is building millions of drones in very simple conditions, and they are able to iterate the design of these drones on time scales of days or weeks," Liphardt said. "I heard one example where experiences from the front lines in Ukraine are being transmitted to people building the drones that same day."

The important lesson for commercial robotics is not military. It is the speed of the feedback loop.

Robots improve when they are deployed into real environments, encounter failures, generate data, return that information to engineering teams and then go back into the field with changes. Shortening that loop can matter as much as having a better model at the start.

That suggests another policy lever for the United States: make it easier to deploy physical AI in real operating environments.

Hospitals, warehouses, construction sites, utilities, logistics hubs, agriculture and infrastructure maintenance can all generate the edge cases robotics companies need. Every real deployment creates data around manipulation, contact, force, localization, navigation, human interaction, error recovery and long-duration reliability.

Physical AI needs that data because physical intelligence is not learned entirely in simulation or in a GPU cluster. It is learned by interacting with the world.

China also has a data advantage

This is where China’s manufacturing scale becomes relevant for AI itself.

"China has one great advantage, which is country scale collection of data," Liphardt said. "Even though we have a razor thin margin in AI in general, the special case of models for physical AI, we may be at a disadvantage there simply because of our smaller manufacturing base."

That is an important distinction.

The United States may continue to lead in frontier AI models while falling behind in the specialized datasets required for robotics. A model controlling a humanoid needs data about friction, objects, collisions, grasp failures, joint loads, contact dynamics, deformable materials, humans, tools and thousands of other physical variables.

A country deploying large numbers of robots into factories and other real-world environments gets more of that data automatically.

Manufacturing scale therefore becomes both a hardware advantage and a software advantage.

Two dates now matter

There are also two important clocks running.

The first is January 1, 2028, when companies face the next major milestone in the FCC process and potential Pentagon approval requirements. Companies seeking relief will need to provide detailed bills of materials and multiyear capital commitments.

The second is November 2026.

China’s October 2025 rare-earth export-control package, including its 0.1% de minimis extraterritorial rule, is currently suspended rather than repealed. If those restrictions return, access to rare-earth materials and permanent magnets could become an even more serious constraint for Western robot manufacturers.

That would hit exactly the part of the supply chain where the United States is already weakest.

The gap is large, but it is not permanent

There is some good news in the numbers.

The United States scores 69.0. Japan scores 69.6. That difference is negligible. China’s 82.9 is much more significant, but industrial capability can be built.

"The gap is definitely surmountable," Liphardt said. "And so we shouldn't be afraid. This is, if anything, it's information that helps us prioritize."

That is probably the most useful way to read OpenMind’s report.

The United States does not have an AI talent problem. It does not have a compute problem. It has world-leading universities, capital markets, software companies, AI labs and robotics startups.

It does have a physical stack problem.

Motors, reducers, bearings, magnets, precision machining, components, suppliers and factories all matter. So does having enough of them located close enough together that engineers can iterate quickly.

If America wants to manufacture robots at scale, building the intelligence layer is only half the job.

The harder part is rebuilding the industrial system underneath it.