The Physical AI Deficit

IThe Illusion of the Choke Point

On March 12, 2026, China's legislature approved the 15th Five-Year Plan. The plan lists robotics, intelligent vehicles, drones, and machinery as AI priorities and calls for "extraordinary measures" to achieve decisive breakthroughs in embodied systems — AI that acts in the physical world through robots, vehicles, and machines.¹ China has a coordinated national strategy for this technology. The United States has world-leading companies and laboratories, and no architecture to connect them.

This is not a story of American failure. By its own metrics, the current U.S. strategy is succeeding. TSMC's first Arizona fab is operational and shipping advanced silicon for Apple and Nvidia — including Blackwell-class AI processors, the first cutting-edge chips TSMC has produced outside Taiwan. Its second Arizona fab finished construction ahead of schedule, with three-nanometer volume production pulled forward a full year on AI demand.² Export controls have denied Beijing the most advanced compute. American labs hold the frontier in foundation models, and on its second day in office the second Trump administration announced a $500 billion private buildout of AI data centers.³ The choke point is real, and it chokes. Edward Fishman's account of American economic warfare makes the definitive case for the chokepoint as doctrine — and concedes its central vulnerability: chokepoints erode as adversaries route around them, the way Beijing now promotes the yuan to escape the dollar's reach.⁴

But a choke point controls what an adversary cannot make. It says nothing about what an adversary can. While the United States consolidated its grip on the digital substrate — outspending China on computing power by roughly twelve to one — China built structural advantage in AI's physical expression: the actuators, reducers, sensors, batteries, and assembly ecosystems through which intelligence will touch the world. China outspends the U.S. on robotics by roughly forty percent, and the gap is widening.⁵ These are two different races. American policy is optimized for the one it is winning.

The deficit is not inventive. American venture capital funds robotics startups lavishly; the design talent is here. The deficit sits at a specific stage — the unglamorous passage between a working prototype and volume manufacturing, where capital intensity rises, margins compress, and American capital markets, structured over forty years to reward asset-light business models, systematically decline to follow. China's system is built to dominate precisely that stage. The argument of this essay is that closing the gap requires restructuring American demand and capital incentives — not escalating a subsidy contest, or a tariff contest, that neither system can actually win.

IIThe Anatomy of Party-State Capitalism

To see why China's position in physical AI is structural rather than circumstantial, start with what the Chinese system has become. Margaret Pearson, Meg Rithmire, and Kellee Tsai have given it the most precise name available: party-state capitalism — a system in which the Chinese Communist Party expands its authority inside firms, public and private alike, through party committees, golden shares, state-linked investment vehicles, and informal discipline, until the boundary between state and private capital stops carrying reliable information.⁶ The category matters for policy. Washington keeps searching for the line between "Chinese companies" and "the Chinese state," and the search fails because the system is designed to blur it — not to erase it, which would be command economics, but to make it permanently ambiguous.

Ambiguity is not harmony. Rithmire's comparative work on business-state relations in authoritarian Asia identifies the Chinese pattern as mutual endangerment: the party can destroy any firm, and large private capital — through flight, offshore structures, and the leverage that comes with employing millions — can damage the party's project in return.⁷ Each side holds the other hostage. The consequences are visible in sustained capital outflows, in entrepreneurs holding foreign passports, and in Rithmire's own skepticism that this system can ever build what it most lacks: large, efficient, trustworthy financial institutions.

The pathologies run deeper than finance, and they are well documented: access money greasing allocation toward the connected, a coalition structure built on selecting for weakness over threat rather than competence (leaders who posed no rival ambitions rose; capable figures were sidelined), unforced errors on the scale of the zero-COVID endgame.^{8, 9} None of it has prevented dominance in solar, batteries, or electric vehicles. The finding that matters most for American policy is fiscal: Rhodium's accounting shows Chinese subsidies nearly universal — more than 90 percent of listed firms since 2010 — with median amounts stagnating since 2020 as the local governments who fund them run out of room.¹⁰

A reasonable reader will ask the obvious question: why mobilize against a system this compromised? The answer is the solar industry. Chinese solar manufacturing destroyed its Western competitors and its own margins simultaneously — a decade of bankruptcies, price wars, and consolidation inside China that nonetheless ended with Chinese firms controlling the overwhelming share of every stage of the global supply chain. The pattern is now repeating in batteries, stage by stage down the value chain.

The same dynamic runs across the economy: more than a hundred electric-vehicle makers of which a handful turn a profit, subsidized steel mills running at a loss, a food-delivery sector so margin-starved that a 2026 crackdown uncovered tens of thousands of unlicensed ghost kitchens funneling orders through the apps.¹¹ The bloodbath that produces world-beating cost structures and the one that produces cockroach-kitchen scandals are not two phenomena. They are one.

Profitless dominance is still dominance. For statecraft purposes, what matters is not whether Chinese robotics champions earn returns for their shareholders. What matters is whether, ten years from now, anyone else can make the machines. The brittleness is real, the fiscal strain is real, and neither is a reason for American complacency, because the system's wastefulness and its capacity to eliminate foreign competitors are the same process viewed from different shores.

IIIThe Two-Sided Dilemma of Statecraft

Clayton Christensen's disruption mechanism, stripped to its core, describes an asymmetry of motivation. Incumbents retreat upmarket because each retreat is individually rational: the low end has thin margins, demanding customers, and inferior products, so ceding it improves the income statement every single quarter — until the entrant's products, refined under brutal cost pressure, become good enough to take the high end too. The incumbent dies of a thousand sensible decisions.

Disruption explains the motivation. It does not, by itself, explain why the retreat cannot be reversed once the error becomes obvious — and the answer to that lies in the economics of path dependence. The foundational work of W. Brian Arthur and Paul David argued that under increasing returns, early allocation choices become self-reinforcing and lock in trajectories that are costly or impossible to escape later.¹² Two mechanisms from that literature do the work here. The first is agglomeration: as a manufacturing base concentrates — Shenzhen, the Pearl River Delta — it draws the skilled workers, the specialized suppliers, and the tacit process knowledge that make it more efficient still, until competitors elsewhere face not a gap but a lock-out, the condition in which a filled niche becomes nearly impossible for an outsider to re-enter.¹³ The second is the capital-market mechanism the lock-in literature names directly: because financial markets are biased toward short-term profitability and incumbents are already heavy users of capital-intensive assets, large new investments in a capability-rebuilding trajectory are, for those incumbents, individually counterproductive.¹⁴ The disruption is the choice; path dependence is the trap that springs shut behind it.

That trap has a famous skeptic, and the objection is worth meeting head-on rather than stepping around. Stan Liebowitz and Stephen Margolis spent the 1990s taking apart the canonical lock-in stories — QWERTY above all — and their case against the famous examples has largely carried the field: most celebrated lock-ins were either never inefficient or were abandoned the moment a clearly superior alternative arrived. The architecture that looked permanent fell faster than the theory predicted. Intel's x86 was the textbook barrier to entry at the chip level for forty years; ARM went from a rounding error to roughly a third of cloud capacity in five, and the default for new workloads, on price-performance alone. Windows lost the desktop's primacy without any antitrust remedy doing the work, because the computing surface that mattered simply moved. If physical-AI manufacturing were a lock-in of that kind — one architecture, one standard, unseated by a better one — the optimist would be right, and American capital would route around China the way it routed around Intel.

It is not that kind. The lock-ins that survive the skeptics' test are the ones that persist for reasons unrelated to technical merit: COBOL still runs the banks and Excel still runs corporate finance, not because nothing better exists but because switching requires moving thousands of mutually dependent actors at once, and no single one of them can move first. A manufacturing base is that case in its purest form. The supplier who would relocate, the process engineer who would retrain, the toolmaker who would retool, the financier who would underwrite the move — each is rational to wait for the others, so none of them moves, on either shore. The distinction is not academic. A single-standard lock-in dissolves itself once the gap grows large enough; a coordination lock-in does not, because the force that would break it is exactly the collective action that markets, left alone, will not organize. This is why the live debate has stopped asking whether such lock-ins exist and started asking the only question policy cares about — whether they are remediable, and at what cost. For a coordination failure the answer is fixed in advance: remediable, but only by an actor large enough to move the agents at once. That actor is not the venture market, which is the whole reason the chapters that follow are about demand and patient capital rather than exhortation.

The United States is running this pattern at national scale. Its industrial strategy concentrates on the most sophisticated, highest-margin tiers of the technology stack: advanced-node semiconductors, frontier model training, leading-edge chip design. Each is a genuine strength and a genuine win. The CHIPS and Science Act is the policy expression of the upmarket march — $39 billion to anchor the most advanced manufacturing process on earth onto American soil.¹⁵

China is disrupting from below, and the mechanism is documented in unusual detail. Barry Naughton's testimony to the U.S.-China Economic and Security Review Commission in February 2025 identified where Chinese industrial policy actually works: sectors with modest entry barriers and a deep, competitive seedbed of firms, where state support funds a domestic bloodbath from which a handful of champions emerge — and the cost discipline learned in that bloodbath then powers decades of decline down the experience curve.¹⁶ Batteries followed this pattern. Solar followed it. Electric vehicles followed it. Humanoid robotics fits the profile exactly: dozens of competing Chinese entrants, a vast domestic lead market, and a manufacturing base in which every required component — motors, gearboxes, sensors, castings, batteries — already exists at scale for adjacent industries.

The pattern's substrate is what Gary Pisano and Willy Shih named the industrial commons: the shared pool of suppliers, process engineers, and skilled technicians on which any manufacturer draws.¹⁷ Their 2009 warning — that offshoring decisions rational for each firm were collectively destroying capabilities America would someday need and could not quickly rebuild — was prophetic and ignored for fifteen years. Commons are know-how, not buildings. They erode invisibly and rebuild on decade timescales — the lock-out condition made physical — which is why the skeptical literature on Chinese industrial policy actually sharpens the warning rather than softening it. Ilaria Mazzocco's assessment for CSIS shows Chinese industrial policy failing expensively where commons and competition were absent — commercial aviation, advanced logic chips — and succeeding precisely where lead-market demand met manufacturing depth.¹⁸ Robotics is not COMAC. Robotics is BYD.

And this time the pattern is not a forecast; the instruments are already deployed. Beijing's National Venture Capital Guidance Fund has earmarked roughly CNY 1 trillion over twenty years with embodied intelligence as a named target; provincial governments are subsidizing up to 30 percent of project costs for automation; and MIIT has stood up a standardization committee for humanoid robots — China is not only building the machines, it is writing the standards the machines will speak.^{19, 20} The solar and battery playbook, run again, in public.

Two honest caveats about the disruption frame, because it breaks in instructive places. Firms can exit markets and die quietly; nations cannot. There is no bankruptcy court for states, no acquirer for a hollowed industrial base — which makes the cost of ceding the low end higher for a country than for a company. And in this case the low end shoots back: the cheap, good-enough machines climbing the curve are dual-use by nature.

Naughton's warning completes the dilemma. China's current phase is explicit self-reliance — written into the 14th Five-Year Plan as science and technology "self-reliance and self-strengthening" — which means American export controls are effective and self-defeating at once: effective on today's timeline, because the denied technology is genuinely hard to replicate, and self-defeating on the longer one, because every control validates Beijing's substitution drive and guarantees that the eventual Chinese alternative will owe nothing to American suppliers.¹⁶ Both clauses are true. Policy built on only the first is a wasting asset. The instruments Washington has deployed — tariffs, grants, export controls — all aim at what China cannot make. None of them addresses what America has stopped making.

IVWhy Current U.S. Industrial Policy Is Misfiring

Set the two systems side by side and the mismatch is architectural. Rhodium's 2026 autopsy of Chinese strategy catches the system mid-mutation: Beijing is not picking sectors anymore — it is attempting to swallow the entire stack of production, from upstream inputs to frontier applications, and doubling down under pressure rather than retreating.¹ The report's strategic conclusion deserves quoting nowhere and paraphrasing everywhere: the window for an effective response is narrowing, and fragmented Western reactions — duplicative national subsidies, trade defenses, allies competing against allies for the same reshored supply chains — risk leaving the underlying imbalance untouched while adding costs of their own.

The American approach, by contrast, is a portfolio of discrete sectoral bets — and since January 2025, an argument about whether even those bets should exist. The second Trump administration arrived with a theory: tariffs could substitute for subsidies. The America First Trade Policy memorandum of January 20, 2025 made tariffs the primary instrument of industrial repatriation; by March the president was urging Congress to "get rid of" the CHIPS Act, citing TSMC's expanded $165 billion commitment as proof that tariff pressure works better than grants.^{21, 15} What followed was the most compressed natural experiment in the history of American trade policy.

The experiment returned three results. Tariffs proved legally fragile — the Supreme Court voided the entire IEEPA regime in Learning Resources v. Trump in February 2026.²² They proved strategically reversible — zero to 145 percent and back to 10 within fourteen months, with Beijing absorbing a thirty-nine-day near-embargo and outlasting it.^{21, 23} And most instructive: the case the administration cites as vindication refutes the substitution theory. TSMC expanded its Arizona commitment under tariff threat while keeping its $6 billion in CHIPS grants — Commerce renegotiated the agreement rather than canceling it.¹⁵ The carrot and the stick worked together; neither worked alone. Congress, meanwhile, declined to repeal CHIPS, and the One Big Beautiful Bill Act of July 2025 spared the IRA's 45X manufacturing production credit from repeal while layering foreign-entity restrictions onto it and accelerating other phaseouts.²⁴ Production-linked incentives survived their most hostile political environment in a generation — barely, and with their investment-horizon value discounted by the visible nearness of the miss.

Now state the structural claim carefully, because its loose version is false. It is not true that American capital refuses to fund physical AI. Venture capital has poured billions into U.S. robotics firms; private AI investment ran to $285.9 billion in 2025, twenty-three times China's.²⁵ Manufacturing construction spending has run at record levels since 2022. The loud counterevidence is real, and it proves the narrow claim rather than refuting it: American capital funds the invention of physical AI and refuses its capital-intensive scale-up. Venture economics can absorb a $500 million bet on a robotics lab; they cannot absorb a $5 billion bet on a robot factory with thirty-percent gross margins and a seven-year payback, and public markets punish exactly that bet — the structural reason America's most successful companies are asset-light and its factories are someone else's problem. The record construction spending makes the same point from the other direction: it is overwhelmingly semiconductors and batteries, the two sectors where federal policy explicitly de-risked the scale-up stage.^{15, 26} This is the capital-market channel of the lock-in introduced earlier, named precisely: evaluation horizons and the asset-light bias of public markets push capital toward quick-cycle, high-margin bets and away from long-payback capability, reinforcing the incumbent trajectory even when the abandoned capability is systemically valuable.²⁷

The two most-cited American counterexamples sharpen rather than break the argument. Leading surgical robotics platforms have scaled on procedure-linked instrument margins running near seventy-five percent — closer to software economics than manufacturing, with hospitals as the guaranteed high-price buyer substituting for a federal demand backstop. Amazon Robotics scaled its warehouse systems against Amazon's own internal procurement — a captive demand guarantee that is, structurally, an advance market commitment run by a single corporation. Both prove the rule: where demand was certain and premium, American capital followed into physical systems at scale. The missing instrument is a mechanism that provides comparable demand certainty to firms that do not have Amazon's balance sheet or access to captive premium demand.²⁸ The newest entrants make the pattern almost too vivid. In March 2026, OpenAI's former chief research officer raised $70 million at a $700 million valuation, backed by Founders Fund and Accel, for Arda — a startup whose explicit purpose is to reduce Western reliance on Chinese manufacturing. Arda builds a video model that learns from factory footage and orchestrates machines and people across a production line. It does not build a single actuator, reducer, or robot.⁵¹ It is the asset-light, high-margin layer the American capital stack reaches for by reflex: the brain for a body manufactured elsewhere. The most credentialed American answer to Chinese manufacturing dominance is an operating system for a robot fleet the United States does not yet make.

The temptation is to read Arda as the cavalry: if software can now learn a factory from video, surely software can also collapse the cost of rebuilding the commons, the way large language models are beginning to collapse the cost of escaping a COBOL mainframe. The analogy breaks exactly where it has to hold. Code lowered the cost of migrating off COBOL because the thing migrated was itself code. The commons is not code. An orchestration model can schedule a supply chain it does not own, route work to suppliers that do not exist, and optimize a line of machines no domestic firm has built; what it cannot do is manufacture the reducer, install the process knowledge, or stand up the equipment base its own instructions assume. Software lowers the cost of running a commons. It does nothing to lower the cost of having one.

Where Washington backstops first-of-a-kind manufacturing, American private capital follows at ten times the public outlay — TSMC's $165 billion Arizona commitment rides on federal grants of roughly $6.6 to $11.6 billion.² Where Washington does not, the capital does not come. Robotics manufacturing has no backstop. That is the deficit, located precisely.

The pattern is old enough to have an empirical literature. Bessen's study of the British cotton industry describes the same dynamic in an earlier idiom: firms delayed irreversible machinery investment even when the new technology was plainly superior, because sunk cost, uncertainty, and short evaluation horizons made waiting privately rational and collectively disastrous — and the delay let foreign competition mature into a shock the industry never recovered from.²⁹ The first-factory gap in physical AI is the same random-walk problem with more zeros on the end.

The honest audit of the two flagship laws confirms both their value and their boundary. CHIPS works as designed: it is winning the advanced-compute race it was built for, and the acceleration of TSMC's Arizona timeline by an estimated two to three years is attributable to it. The same record shows commons erosion inside the flagship — Micron's New York fabs have slipped roughly five years, with DRAM production now expected around 2029 or 2030,³⁰ and the industry projects a shortfall of some 90,000 fab-operations workers by 2030.³¹ America funded the buildings and discovered it lacked the people. The Inflation Reduction Act taught a cleaner design lesson: its 45X production credits — paid per unit of output, durable, rules-based — moved private capital into clean-energy manufacturing faster than any discretionary grant program.²⁶ Both flagship laws survived a hostile administration; neither touches actuators, precision reducers, sensors, or robot assembly. In the Marshall–Krugman language of economic geography, that is exactly the part of the system where agglomeration has already lapped the field: China's manufacturing commons now enjoys the circular causation Krugman modeled — dense supplier bases, thick labor markets, cumulative process knowledge — while latecomers face the mirror-image problem of lock-out, formally free to enter and substantively excluded.³²

One more datum locates the vacuum. In 2025, the U.S. Chamber of Commerce Foundation published a framework, co-authored by Rithmire and David Fagan, teaching American companies to build internal geopolitical risk committees — structured corporate governance for navigating a security-driven Chinese political economy and an American response made of statutes, tariffs that may or may not exist next quarter, and congressional rhetoric.³³ Read it as evidence rather than guidance: when the nation's largest business federation must train firms to privately manage a risk this systemic, the state has not resolved it. This diagnosis is no longer heterodox: writing in Foreign Affairs in April 2026, former National Security Advisor Jake Sullivan and co-author Tal Feldman reached the same conclusion — that American investors "devote far less attention and money to capital-intensive, lower-margin industrial production" and that any strategy relying on market forces alone to allocate capital to strategic hardware manufacturing will fail.³⁴ The essay that follows proposes the mechanism their piece does not. Even Andreessen Horowitz — the firm whose fortune rests on the thesis that software eats the world — now warns that China has "rapidly surpassed global competitors in robotics, building a complete supply chain," that it passed the United States in robot density in 2021 and Japan and Germany by 2024, and that without industrial reform America risks losing the physical-AI race outright; the firm has begun funding domestic actuator and motor production directly.⁵² When the software maximalists start financing factories, the diagnosis has carried the room.

VA New Framework for U.S. Industrial Policy

The diagnosis dictates the architecture. The failure is two-part — capital refuses the scale-up stage; the commons has eroded — so the response must be a structure, not a list: a foundation of capital-market plumbing, and on it three pillars — finance, demand, and commons — each matched to a mechanism the Chinese system cannot replicate and the American system has run before.

The foundation: finish rewiring the plumbing (federal). Congress already proved, mostly by accident, that the rules governing capital intensity matter more than appropriations. The 2022 requirement that firms amortize research costs under Section 174 — a budget gimmick — measurably suppressed R&D-heavy investment until the One Big Beautiful Bill Act restored full domestic expensing in 2025.²⁴ The restoration was correct and incomplete. Two gaps remain. First, first-of-a-kind manufacturing finance: the Export-Import Bank's domestic "Make More in America" authority and the Department of Energy's loan programs show the form, but no instrument exists for non-energy advanced manufacturing — a robot factory cannot borrow against a federal guarantee the way a battery plant can. Second, institutional capital: ERISA fiduciary interpretation keeps the largest patient-capital pools in the country — pension and insurance assets measured in trillions — effectively out of advanced-manufacturing debt. Safe-harbor guidance permitting allocation to qualified industrial credit would move more money than any conceivable appropriation. The objection writes itself: these tools are untargeted and shift risk toward retirees. Untargeted is the design — at the plumbing layer the state should set relative prices, not pick firms — and the risk objection is why the instrument is senior debt with federal guarantee architecture, not equity. The concession that remains: plumbing alone will not close a sector-specific gap. Hence the pillars.

Pillar one: an American scale-up bank (federal). The United States has run state industrial finance before, twice, and both times it worked by being boring. The Reconstruction Finance Corporation, chartered in 1932, financed everything from banks to synthetic-rubber plants through professional underwriting and was dissolved when its moment passed. In-Q-Tel, chartered in 1999, gave the intelligence community an equity instrument insulated from procurement politics. The Department of Defense's Office of Strategic Capital, created in 2022, is the embryonic third instance — but it lends only, and small. The proposal: scale it into a chartered industrial finance corporation with equity and first-loss authority, mandated narrowly to the volume-manufacturing stage of physical AI — actuators, precision reducers, sensors, robot assembly — the stage where the valley of death sits. The strongest objection is the serious one: state equity invites politicized allocation, and Ang's framework for access money applies to American capital as readily as Chinese.⁸ Solyndra is the standing cudgel. The design answer is the RFC's, not a denial: independent underwriting, portfolio rather than project accountability — a venture fund is judged on the fund, and a state industrial bank must be too, or no rational manager will ever take the risks the mandate requires — and a statutory sunset that forces periodic rejustification. The institution must be independently chartered, not housed in the Pentagon. The Office of Strategic Capital is the seed, but the Defense Department optimizes for exquisite, low-volume, high-specification systems — the exact upmarket incumbent behavior this essay diagnoses as the disease — and a bank captured by service politics would finance the next exquisite platform, not the boring volume factory. The independent charter carries its own scar, named here rather than hidden: Congress guards appropriative power jealously, and the RFC itself was dissolved in the 1950s in part because it had grown too independent of the body that created it. The harder political birth is the price of the only design that survives contact with the diagnosis.

Pillar two: buy the curve down (federal, with a state variant). The deepest American precedent in this entire domain is demand-side, and Washington has forgotten it. In the early 1960s, the federal government — NASA for Apollo, the Air Force for Minuteman — purchased nearly the entire output of the infant integrated-circuit industry, absorbing the early costs no commercial buyer would bear. Within six years, the average price of an integrated circuit had fallen by more than ninety-five percent, and a commercial industry existed where none had.³⁵

The mechanism was not subsidy. It was guaranteed volume, which is the one thing the experience curve runs on — the same fuel Naughton identifies under every Chinese industrial success.¹⁶ The modern instrument exists: Operation Warp Speed's advance market commitments demonstrated that binding federal purchase guarantees, milestone-gated, can pull a technology through development at speed. Apply the architecture to robotics across the federal estate — postal logistics, VA hospital systems, defense depot operations, disaster response — as competed, milestone-based advance commitments. And one condition of award is the keystone of the entire mechanism: open interface standards. A credible critic will note that the integrated circuit of 1962 was a relatively standardized product and the robot of 2026 is not — every firm builds its own proprietary architecture, which is part of why capital cannot flow efficiently to the sector in the first place.³⁶ So the government should not buy robots. It should buy robots that meet open interface standards, forcing a fragmented supply chain to standardize and scale in the same motion — and denying Beijing's MIIT committee an uncontested run at writing the global rulebook.²⁰ The political opening is already visible: as of late 2025 the administration was drafting a robotics executive order, with the Commerce Secretary convening robotics CEOs and framing the goal as removing regulatory friction.³⁷ An advance-commitment mechanism slots directly into that conversation — it asks the government to do something it is already considering, with a financial architecture it has already run. The objection that survives is procurement speed: the federal government has never bought robots well, and Warp Speed worked partly because HHS held emergency authorities no civilian robotics program possesses — that authority gap is real, and Congress would have to close it. The claim stays narrow on purpose: not that government buys well, but that guaranteed volume forgives many sins that grants do not, because the vendor still has to ship working machines to get paid. States hold a parallel lever in their own procurement — transportation departments, university and hospital systems — at smaller scale but faster cycle times.

Demand should be pointed one layer deeper than the finished robot. The tightest chokepoint in the physical-AI stack is not assembly but the precision components beneath it — the harmonic and cycloidal reducers, servo motors, and actuators where a handful of mostly Japanese suppliers hold the world's capacity and the United States has almost none. A robot is only as domestic as its reducer. An advance commitment that buys finished robots while their drivetrains are imported simply relocates the dependency; the purchase guarantee should reach explicitly to the component tier, the way the integrated-circuit precedent pulled an entire upstream materials and equipment base into being rather than only the finished chip. The early private market already sees it: when Andreessen Horowitz seeds domestic actuator and motor manufacturing, it is pricing the same chokepoint, one venture round at a time.⁵² Public demand can do at scale what a seed round signals.

Pillar three: rebuild the commons at the scale the word implies (state-led, federally cost-shared). The institutional forms are proven and the American version is homeopathic — and Ohio has already built the prototype. Arsenal-1, Anduril's hyperscale autonomous-systems campus in Pickaway County, went from groundbreaking to first article in roughly fourteen months, financed through a layered stack: some $900 million of private capital; a $310 million grant from JobsOhio, the state-chartered nonprofit that deploys patient capital at private-sector speed; a thirty-year state tax credit; and a $60 million advanced-manufacturing training center at Columbus State Community College that Franklin County voters approved before the factory existed.³⁸ The governor co-architected the deal. High schools in Teays Valley are writing curriculum for it. The commons is not a federal program; it is a zip code, and this one is 43160.

Arsenal-1 also marks the model's current limit, and the limit is the argument's next step: it is one firm's factory running one firm's proprietary platform. The mature form is the same governance architecture — state capital vehicle, governor as co-architect, community-college pipeline, federal cost-share — applied to a neutral, multi-firm institute whose manufacturing platform is an open standard any robotics firm can build to. That is what Germany's Fraunhofer-Gesellschaft actually is: roughly 76 institutes on a budget near €3 billion, a third earned through industry contracts, functioning as the process-engineering memory of an entire industrial ecosystem.³⁹ The American counterpart, Manufacturing USA, runs institutes on single-digit millions of baseline federal funding per year.⁴⁰ Taiwan's ITRI — which spun out TSMC itself in 1987 — is the standing rebuttal to the claim that institutes cannot precede the ecology they serve. And the lever is one states hold today, without waiting for Washington: pension allocation policy and public college systems are state instruments. What JobsOhio and Columbus State are to Ohio, CalPERS and the community-college system are to California.

What California still lacks is the convener — the actor that assembles capital, anchor, site, and pipeline into a single deal the way JobsOhio did. Two organizations are closer to that role than any public agency. The Silicon Valley Leadership Group, whose member companies account for nearly a third of private-sector jobs in the Valley, launched a Physical AI Task Force in April 2026 to keep California "the global epicenter of physical innovation as AI transitions from the cloud to embodied systems"; its CEO, Ahmad Thomas, frames the work as "a framework for innovation that balances competitiveness, workforce readiness, and thoughtful public-private partnerships," and the task force has already begun the unglamorous convening — roundtables, a cross-industry framework, university lab visits — that precedes any factory.⁵³ The Bay Area Council Economic Institute has supplied the other half: a full economic model for a Solano County advanced-manufacturing complex anchored by what would be the largest manufacturing park in the United States, with an explicit focus on aerospace, defense, energy, and robotics, in a state that by the Council's own accounting has added no net new manufacturing jobs since 2010.⁵⁴ Between them sits the lesson and the gap. They have built the convening table and the site prospectus — the two legs Ohio needed most — but neither has yet paired a private anchor tenant with a shovel-ready site the way JobsOhio paired Anduril with Pickaway County. The Solano analysis is a prospectus waiting for its Anduril. The work of the next eighteen months is to close that pairing — and the organizations positioned to do it already exist.

One objection deserves its answer here, because it is the political one: automation reads as job destruction, and any state-sponsored automation program will meet labor resistance — resistance Beijing, managing its own anxiety over a shrinking workforce, suppresses by decree, and Washington cannot.¹⁹ The American answer has to live in the design. A country will either build the machines that automate its factories or import them; the jobs at stake are the jobs of building, programming, and maintaining the machines; and a country that does not build them will not train the workers who keep them running. The 90,000-worker fab shortfall is not an argument against the factories³¹ — it is the proof that the pipeline and the factory are the same investment.

What this framework refuses to do is also part of the design. No tariff escalation as strategy: the past eighteen months demonstrated that tariffs are legally fragile, strategically reversible, and — because they tax the American adopters of automation — pointed in exactly the wrong direction for buying the curve down. No robotics-CHIPS block grants: grants fund announcements; production-linked instruments fund output, and the IRA already ran the controlled experiment.²⁶ And no further reliance on export controls as strategy rather than tactic, for Naughton's reason — every control buys time on one clock and validates substitution on the other.¹⁶

VIThe Geopolitical Stakes

Rush Doshi's account of Chinese grand strategy describes a sequenced project to blunt American order and build an alternative, with technological and economic displacement as instruments rather than byproducts.⁴¹ One need not accept the framework's strongest form — critics reasonably argue it over-reads coherence into a system that improvises more than it plans — to accept its operational implication: Beijing treats industrial position as strategic position, and behaves accordingly.

Run the physical AI race forward ten to fifteen years and the defense-industrial arithmetic does the arguing. The commercial and military bases are not adjacent; they are the same base — the drone and the humanoid draw on the same actuators, the same batteries, the same supplier networks that Section III's experience curve is compounding quarter by quarter. A nation that manufactures autonomy at scale — cheap, attritable, iterating systems in the millions of units — sets the depreciation rate on everyone else's military capital. Ukraine previewed the logic with improvised drones; physical AI industrializes it. The American defense base, optimized for exquisite platforms in small numbers, is the upmarket incumbent of Section III wearing a uniform.

The scale asymmetry cannot be closed by the American market alone. This is where the alliance system stops being a sentiment and becomes a production function.

Japan dominates the precision components — strain-wave and cycloidal reducers — that every humanoid robot on earth requires; Germany and South Korea hold machine-tool and automation depth; Taiwan holds the silicon. Pooled allied demand and pre-committed supply-chain roles — the Combined Production and Resources Board of 1942 is the precedent, the F-35 consortium the imperfect modern echo — could match Chinese lead-market scale in a way no national program can. Rhodium's warning makes the negative case: the current pattern, in which allies subsidize competing national champions for the same reshored supply chains, is the one configuration guaranteed to fail, converting alliance into intra-allied industrial rivalry while the underlying imbalance compounds.¹ The tariff campaign of 2025 — which hit Japan, Germany, South Korea, and Taiwan alongside China before the courts intervened — ran this failure mode in real time. The trust damage is not cost-free to reverse: alliance relationships that function as production functions require the kind of credible, durable commitment that the past eighteen months have specifically destroyed.

The capacity is not in doubt. In December 2025 the United States launched Pax Silica, a coalition now spanning sixteen nations and an investor consortium of sovereign funds — Mubadala, SoftBank, Temasek — that the State Department put at more than a trillion dollars in assets, anchored by a $250 million federal seed fund; the European Union moved to join in June 2026, reportedly alongside a commitment to purchase some $40 billion in American AI chips.⁴² When Washington decides a layer of the technology stack is strategic, it can still convene the allied industrial coordination this essay calls for. But read the scope: critical minerals, energy, semiconductors, AI infrastructure, logistics. Pax Silica is a coalition for the silicon — the upmarket layer the United States already leads. It is the incumbent organizing its allies around the high ground it holds, while the low end of the stack, the actuators and reducers and assembly lines through which intelligence becomes physical, has no equivalent pact. That is the disruption pattern of Section III rendered as foreign policy: the rational concentration of effort on the profitable, sophisticated tier, and the quiet ceding of the unglamorous one. The Iran conflict supplied the warning in literal form — Iranian strikes on data centers in the UAE, and a Strait of Hormuz that reopened only on Tehran's terms, demonstrated that the physical infrastructure of the AI age is a target, and that the chokepoints run through hardware the United States does not make.⁴³

The model-layer counterargument — that America can capture the high-margin intelligence tier while allies or partners handle assembly, the Apple model applied to robots — deserves engagement rather than dismissal. Apple's own trajectory refutes it. Patrick McGee's account of Apple's supply-chain entanglement documents how Cupertino did not simply outsource assembly to China; it built China's precision-manufacturing ecosystem, trained its workforce, and transferred its process knowledge until the assembler's leverage over the brand became structural, not contractual.⁴⁴ The robot is not the iPhone. The robot's manufacturer controls its firmware update cycle, its sensor data, its maintenance dependencies, and its iterative hardware evolution. Ceding manufacture of the physical substrate to a strategic adversary is not the Apple model — it is the worst version of it, without the thirty-year head start that made exit seem possible.

VIIThe Reformulation

The temptation, reading China's subsidy totals, is to match them. The counter-temptation, watching the tariff experiment, is to wall them out. Resist both — and not from restraint, from arithmetic. Subsidy escalation is the one contest both systems lose: China's is fiscally exhausting itself sustaining near-universal support on stagnating local budgets,¹⁰ and America's political economy cannot sustain sectoral subsidy across election cycles, as the IRA's near-death experience demonstrated within three years of enactment.²⁴ The tariff contest fares no better: its legal foundation was stripped by a 6–3 Court, and Beijing demonstrated it could outlast a near-embargo measured in weeks.^{22, 23} A bidding war between an exhausted treasury and an inconstant one enriches neither; a tariff war between a just-in-time economy and a state that absorbs pain by decree favors the decree.

The leverage lies where the systems differ. Demand certainty, capital plumbing, and a living industrial commons are things the American system can build because it has built them before — Minuteman procurement, the RFC, the land-grant pipeline — and things party-state capitalism cannot buy, because mutual endangerment poisons the trust that patient private capital requires.⁷ That is the realism of the moment: not out-spending the Chinese system, and not out-taxing it at the border, but out-structuring it, on terrain where ambiguity between party and capital is a liability instead of a weapon.

The clock that matters is not a fiscal year, and it is not an election cycle. The convergence of cheap intelligence and Chinese manufacturing depth is a compounding process — every quarter of volume production buys cost declines, process knowledge, and supplier depth that no future appropriation can repurchase at any price. Until Washington builds this architecture, the corporate risk committee remains the country's only hedge against a structural deficit — and a risk committee is not a factory. The deeper stakes belong to another essay: the choice of who manufactures the physical layer of intelligence is also a choice about which order succeeds the one now passing, and critical junctures, by their nature, do not announce themselves until they have closed. That argument can wait. This one cannot. Compounding processes are gentle to early entrants and merciless to late ones. America is not yet late — but the price of entry rises with every factory built on the other side of the world, and there is no discount for waiting.