Key Points
- The AI and Robotics Strategy explicitly moves the axis of competition toward “integration and operational capability.” That reading is correct. The problem lies in the shape of the market the strategy assumes.
- On July 28, the United States, using place of production as its test, blocked authorization of new models of foreign-made mobile robots. This origin-based designation is the third such case, after drones and routers.
- Even if robots run on Japanese soil, the “memory” will not remain unless data ownership is settled at the procurement stage. The yardstick must be rewritten now—from unit count to memory.
*Dollar equivalents are approximate, converted at ¥155/$ as of September 2026, and will shift with the exchange rate.
The Government Is Looking at the Right Axis
Let me begin with a caveat: the government’s assessment is not off the mark.
The AI and Robotics Strategy finalized on March 26 explicitly states that the center of gravity of competition is shifting from the “scale” of computing resources and web data toward the “integration and operational capability” of taking in on-site data and continuing to run improvement cycles even after deployment. It calls for not merely increasing the number of units deployed but building a cycle that accumulates on-site data and operational know-how and feeds them back, through evaluation and verification, into model improvement. The priority fields were initially 16. Then, on June 30, the Ministry of Economy, Trade and Industry (METI) released a revised version, adding food service and food manufacturing as well as healthcare to bring the total to 18, and setting a deployment target of roughly 10 million units by 2040. The sector-by-sector deployment roadmap also refers, as measures to prevent the outflow of real-machine data, to the use of closed networks, access control, and the management of storage and transfer.
“On-site data is the real asset.” “Move first on domestic deployment.” These arguments are already written into the national documents. There is no point in this essay repeating them.
The question to ask lies beyond that: which market was this strategy written for?
The Line Was Drawn at the Place of Production, Not the Company
On July 28, the U.S. Federal Communications Commission (FCC) added foreign-produced “advanced robotic devices,” along with network-connected foreign-made power inverters, to its “Covered List”. Within scope are devices meeting certain requirements—humanoid, quadruped, and autonomous mobile robots, among others—that move across the ground and are equipped with sensors, network connectivity, and control software. New models that qualify cannot, in principle, obtain FCC equipment authorization, and thus cannot be imported into or sold in the United States. The measure does not apply retroactively to models authorized before July 28. There is a path of exception through conditional approval, but for robots this is a sole review by the Department of War (formerly the Department of Defense). The involvement of the Department of Homeland Security that is allowed for inverters and drones is not available for robots.
What is decisive is that the test was placed on the place of production, not the nationality of the company. The determination uses the domestic end-product test of the Federal Acquisition Regulation, which requires, in addition to manufacture within the United States, that domestic component costs exceed 65 percent of total component costs. Whether a company is Japanese is not asked. Whether it was made in Japan is.
It is also worth noting the layer at which the line was drawn. To view Physical AI through the lens of economic security, it helps to divide the field into three layers. The component layer is reducers, motors, sensors, bearings, and power semiconductors. The control layer is base software, motion control, the integration of vision and force sensing, safety control, and remote updating. The deployment layer is workflows, exception handling, maintenance histories, records of failures, and operational data.
The United States drew its line at the entrance to the second and third layers. Where Japan is strong is the first layer. This asymmetry governs everything that follows.
This Is Not a Matter of Bad Timing
One is tempted to file this away as “the strategy came in March, the regulation in July—a mismatch that could not be helped.” I once wrote exactly that. But the facts do not allow it.
A blanket, origin-based designation of an entire category is not new. This is the third case in under eight months, following drones in December 2025 and home routers on March 23, 2026. The drones came more than three months before the strategy was finalized; the routers, three days before. Missing the routers is forgivable. The drones are another matter.
There is a weightier fact still: the strategy was subsequently revised. On June 30, it widened the fields from 16 to 18 and newly set out a figure of roughly 10 million units by 2040. By that point, the U.S. template of an origin-based test had been used twice and was public information anyone could read. There was an opportunity to revise. What was revised was the number of fields and the unit target—not the assumption about the market. I do not put this down to the negligence of the officials involved. Japanese policy documents are good at adding demand-side fields and poor at questioning the very contours of the market where that demand is supposed to arise. Add fields, and the number of ministries and industry associations involved grows, widening the area of agreement. Question the premise, and the agreement breaks, at least for a time. With no one willing to take on the work of breaking it, the documents grow thicker. This structure has not changed one millimeter since my days in the Diet.
Premise One—Make at Home, Sell to the World
The strategy’s first goal is to capture more than 30 percent of the global market by 2040—worth some ¥20 trillion (~$129 billion). Its supply-side measures, too, are built around cultivating domestic makers of multi-purpose robots and constructing the whole chain—from development to mass production to maintenance—within Japan.
But the largest advanced-economy market has closed to new models of mobile robots produced in Japan. The Japanese government’s strategy says make it at home; the U.S. market says do not make it at home. There are only three ways to reconcile the two: produce more than 60 percent within the United States, seek conditional approval, or designate markets other than the United States as the destination.
What I would first ask of the government is not a countermeasure but disclosure. Of that ¥20 trillion, how much was expected to come from exports to the United States? If the answer is close to zero, then the target was built from the start around the domestic and non-U.S. markets, and the center of gravity of the investment to be supported shifts accordingly. If a substantial amount was expected, the target must be redrawn. Leaving the figure standing while saying neither is the most harmful thing of all for corporate investment decisions.
It should be added that Japanese players’ shipments to the United States are likely very small at present. If so, what has been lost is not revenue but future options. The loss of options does not show up as a single yen on an income statement or in trade statistics. Because the numbers do not worsen, no alarm sounds; because no alarm sounds, the premise is never rewritten. What Japan has experienced many times is precisely this quiet kind of retreat: with no memory of ever deciding to withdraw, one looks up to find the market gone.
Premise Two—The Components Layer Is Safe
In industrial robots, Japanese firms remain strong. But the way share is read calls for care. The frequently cited “roughly 70 percent global share” is a rounding of a value-based figure for 2022 (65.9 percent) drawn from a NEDO survey, and it differs from volume-based shares such as units installed or produced.
The substance of the strength lies in the drive train. As of 2022, Nabtesco held a roughly 60 percent global share in the precision reducers used in the joints of mid- to large-sized industrial robots, and Harmonic Drive Systems likewise occupies an important position in strain wave gearing. I have called such positions “invisible monopolies.” The company name never appears on the finished product, but if supply stops, the production line stops. It is Japan’s quietest and strongest card.
This layer lies outside the July line. But “outside, therefore safe” does not hold. There are two reasons.
First, the place where the line is drawn moves. Six days before this designation, on July 22, the FCC adopted a rule barring authorization of equipment that incorporates logic-bearing components made by entities on its Covered List. Japan’s reducers are not covered now. But that the line can move toward the component side has already been demonstrated. Moreover, because the test is a ratio of component costs, the component layer is, arithmetically, already inside the line: the more a robot assembled in the United States carries expensive Japanese-made reducers, the higher its foreign-content ratio rises. The better and more expensive the component, the more likely the customer’s robot is to be rejected.
Second, a monopoly does not carry across generations. What conventional reducers were asked to deliver was high rigidity and precision and long-term durability—specifications for repeating the same motion millions of times, fixed in place on a factory floor, in a predictable environment. In humanoids, with their many joints, the weighting changes: lightness, torque density, efficiency, and the safety of coexisting with people all come to the fore. In July 2026, Nabtesco announced it would add two compact precision-reducer products for collaborative robots and humanoids. The very party that has long held the “joints” now positions humanoids as a market it is “about to enter.”
Japanese players are moving. According to the Nikkan Kogyo Shimbun, Harmonic Drive Systems plans ¥27.5 billion (~$177 million) in capital investment over the three years from fiscal 2024, allocating about a third of it to humanoid applications. What is in question is not direction but scale and speed. According to JETRO, a vice minister of China’s Ministry of Industry and Information Technology said in July that the country’s humanoid-robot output is expected to rise from about 20,000 units in 2025 to more than 40,000 in the first half of 2026 and to exceed 100,000 for the full year. It is an official estimate, and a considerable share of shipments is surely bound for research, demonstration, and exhibition. Even so, between the side making 100,000 units a year and the side stuck at a few thousand, the number of chances to learn from defects differs by an order of magnitude. Overtaken in volume while still ahead in quality—how many times have we watched Japan lose this same way?
Premise Three—Deploy, and the Memory Stays
First, how to read the FCC’s measure. Some see it as a foolish move by a United States that cannot build humanoids. I do not.
The reason the FCC gave is the risk that the records of work environments and operations that connected foreign-made mobile robots collect within the United States could flow abroad. It does not say it is enclosing data as industrial policy. But the effect is to narrow the paths by which data generated on U.S. sites flows to foreign suppliers. The timing matters, too. Drawing the line before installation advances is cheaper than peeling it off afterward. The United States has the experience of having had to set up a compensation program for carriers in order to remove equipment made by a certain country from its communications networks. It is not foolish because it is early; it is cheap because it is early. What the United States loses this time is not the memory it would have gained. It is its own speed of adoption.
Of course, success is not guaranteed. Protecting the demand side does not ensure that supply-side capability will grow. There is a precedent in small drones: even after shutting out those made by a certain country, substitutes could not catch up on cost, and U.S. sites were left to choose between using expensive machines and giving up on automation.
Turn, then, to Japan. Running robots at home while also using foreign-made machines, and storing up memory—this path itself is one the strategy has already drawn. The most severe labor-supply constraint in the advanced world is, in the context of Physical AI, also one of the world’s largest sources of deployment demand. In this one respect alone, our weakness becomes an asset.
What is not written is whose the accumulated memory becomes.
What the strategy deals with is data “security”: closed networks, access control, the management of storage and transfer. All of these are designs for preventing outflow abroad. But keeping data within the country and Japan being able to own it are separate questions. Operational records do not automatically remain as the asset of the company that used them unless a contract says so explicitly. The maker, the cloud provider, the remote-maintenance operator, the systems integrator—who among them acquires how much, stores it where, and how far they may use it for secondary purposes, for training, or for transfer abroad—that is decided at the procurement stage. Deploy without deciding anything, and the memory accumulates not with the side that operated the robots but with the side that holds the records. Even when it is kept within the country.
What is needed is to write, in advance, the specifications for capturing records and the ownership of the data into the conditions for public subsidies and procurement, and for the state to provide a standard contract template.
But this single line is not free. Make data ownership a requirement, and the maker loses its greatest form of compensation—training data. Either that feeds through into price, or the maker’s entry into the Japanese market is pushed back. If one still insists it should be written in, then unless one also designs who bears the excess cost, the template will be quietly crushed on the ground. That China’s Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission jointly announced, on June 8, a special initiative on the field training of humanoid robots, explicitly naming the accumulation of high-quality real-machine data as its aim, is likely the flip side of the same judgment.
It should be noted that if the premise of this essay is to break down, it is from here. If advances in simulation, public datasets, and synthetic data come to substitute for much of real-machine data, the arena of competition returns once again to computing resources and models. Even so, long-term operational data—including site-specific objects, work procedures, exception handling, and records of failure—is hard to substitute for completely.
What Is Halting Deployment Is Not Technology but Reimbursement Design
And Japan is also blocking this winning hand with its own hand.
The strategy placed elderly care among its priority fields. But it does not touch the reimbursement system. Because long-term care fees for facility-based services are set according to the level of care required and similar factors, becoming more efficient does not increase revenue. Because staffing standards set a floor on the number of employees, it is also hard to translate labor-saving directly into lower personnel costs. The fiscal 2024 revision created a new add-on for promoting productivity-improvement systems, and the incentive to claim it has been strengthened in stages. But what moved was the add-on side. As long as the base fee remains tied to the level of care required, the structure in which the gains from labor-saving do not translate into revenue will not change. Both unit count and memory stop here.
Whether Japan’s largest source of deployment demand can be turned into numbers depends not on the performance of robots but on the design of reimbursement.
To Keep Holding Is to Keep Renewing
One thing should be added. A generative-AI accident is information leaking; a robot accident is things moving. A loss of control from malfunction or cyberattack produces physical consequences—a runaway or a shutdown. The scope to be protected extends beyond the machine itself to remote maintenance, ID management, networks, software updates, and the supply chain. The FCC’s concern points to this same scope. Separately from information security, one must design shutdown procedures for the case where control is seized.
The government has two things to do. First, redraw the premise of the ¥20 trillion target for a market in which the origin-based test is now entrenched, and disclose the breakdown. Second, change the yardstick of deployment from unit count to memory. The strategy set out a cycle of data feedback, but its indicators lean toward the number of units deployed. The hours and number of tasks safely completed; the volume of data captured in reusable form; the total cost and accident rate per task—set these three as national indicators, and insert ownership clauses into the requirements for subsidies and procurement.
Japan took hold of the joints. But that is a victory in one generation—industrial robots. The United States has already drawn a line at the layer that constitutes the next generation. The winning hand drawn in March was correct. There was even an opportunity to revise it in June. And yet the premise alone remains unmoved.
What it takes to keep holding is not only making the components more precise. It is to hold, thickly and at home, the sites where those components operate, and to make the “memory” generated there one’s own asset through a single line in a contract. A monopoly is not something to defend but something to renew.
*The views expressed in this article are solely those of the author and do not represent the views of any organization, including the institution with which the author is affiliated.
(c)Alamy Stock Photo/amanaimages
