Part III: Assets, Minerals, Routes

40–60%of the entire BoMStrip a humanoid to its bill of materials and movement dominates the cost: actuators (the motor, reducer, and screw assemblies in every joint) run 40–60% of the entire BoM by McKinsey's count. The precision part in each rotary joint is the strain-wave reducer, a gearbox the size of a hockey puck built around a thin flexing steel spline, delivering 50–100:1 reduction with near-zero backlash. Tesla's Optimus Gen2 carries 14 of them plus 14 linear joints on planetary roller screws, body only, before the hands; other platforms run anywhere from 16 to 56 degrees of freedom, so treat the count as a platform choice and use a blend.

542,000robots installed in 2024The world's precision-reducer industry is sized for the industrial-robot market: 542,000 robots installed in 2024 per the IFR, above half a million four years running, on an operational stock of 4.66 million. Against that base, world strain-wave output plausibly runs in the low single-digit millions of units a year. This is the least certain estimate in this part; published figures disagree by multiples. A working midpoint of around two million units is effectively consumed by existing industrial demand.

Even the bearish forecast strains that base. TrendForce, in the same breath as calling 2026 the inflection year, expects global humanoid shipments to merely pass 50,000 units. At a conservative blended ten precision reducers per robot, that one number asks for about 500,000 strain-wave units: another quarter of estimated world output, for a product category that barely existed three years ago, on top of the industrial demand that already consumes the base. China's own figures come next: the formal deployment target is small, about 10,000 humanoids by end-2026, set by MIIT, and MIIT's own officials put 2026 production above 100,000 units. Blend 8–14 precision units across a cheap-skewing mix and that is around a million reducers. One country, one year, roughly half the working estimate of current world output, again on top of the industrial base that already absorbs that output. Bank of America expects 1.2 million humanoids a year by 2030, which lands above ten million reducers. Morgan Stanley has more than a billion in service by 2050; assume a ten-year service life (an analytical assumption; no bank publishes one) and replacement demand alone runs toward a hundred million a year.

The forecast spread runs from IDTechEx's $29.5 billion realizable market to ARK's ~$26 trillion, which ARK itself conditions on humanoids operating at scale. Such a spread indicates extreme forecast uncertainty. The bearish row is enough: its 500,000 reducers do not exceed the estimated two-million-unit gross base, but they ask an already-committed industry to add roughly 25% more qualified output for a new category. The bottleneck is incremental qualified headroom, not gross nameplate.

A supply chain is gated by its scarcest layer. Surplus in one layer never covers a hole in another: a warehouse of motors does not conjure a missing reducer, and a flood of reducers does not license a banned magnet. Capabilities that merely add can average their way past a weak spot. Hardware cannot.

Three circulating claims are wrong. Leaderdrive's "90% share" is 2017 marketing copy; the real domestic figure runs 26–40%. "Actuators are 95% of BoM" fails any full-text audit; 40–60% is the sourced range. And there is no "three-company magnet monopoly"; the magnet industry is fragmented across some 200 producers, which is exactly why the chokepoint sits somewhere else entirely, as the minerals section shows.

Strain-wave. Harmonic Drive Systems of Japan dominates the niche (trade press puts it near 80% of the global harmonic market, with independent trackers bracketing between 60% and 85%), and its financials tell the story of an incumbent meeting a flood: FY2026 revenue up 7% to ¥59.6 billion while net profit fell 53.7%, with guidance for a 180% profit rebound in FY2027 on humanoid volume. The challenger is Leaderdrive (Suzhou Green Harmonic; one company, two names), holding a real 26–40% of China's domestic harmonic market: 2025 revenue up 47%, net profit more than doubled, capacity climbing from 50,000 units a month in early 2026 toward a targeted 100,000–120,000 by year-end, against roughly 425,000–500,000 shipped in all of 2025. Its ¥1.4 billion placement, completed January 2025, funds a project sized at a million reducers and 200,000 actuators a year. The market has already voted on the trajectory: Leaderdrive trades near $9.9 billion, about twice Harmonic Drive's $4.4 billion. Zhejiang Shuanghuan, the auto-gear giant, is standing up a Suzhou plant described at 500,000 reducers a year.

RV and cycloidal. Nabtesco owns the other reducer family, the heavy cycloidal units in big industrial arms, at roughly 35% of the global RV market for articulated robots, and is reported doubling capacity by 2026 for humanoid demand. A separate lane: Optimus uses zero RV units, though heavier platforms' hips and knees may go cycloidal. Europe's Spinea, now Timken-owned, sits adjacent.

$1.2Mper five-axis thread grinder, 18-month backlogPlanetary roller screws, the linear half. Two Western owners: GSA, which acquired Rollvis in 2016 and with it holds over half the global market, and Schaeffler, which bought Ewellix, the number three, in 2022. The constraint underneath them is meaner than the screw itself: the custom planetary thread grinders that hold sub-5-micron tolerances carry reported 18-month backlogs and cost over $1.2 million per five-axis machine, and precision machining runs 25–35% of screw cost. Schaeffler premiered an all-in-one humanoid actuator at CES 2026 (two-stage planetary gearbox, motor, encoder, controller, 60–250 Nm, pitched at 25–30 units per humanoid), aimed squarely at this bottleneck. China is pouring in: Shuanglin bought a grinding-tech firm and cut equipment cost from ¥10 million to ¥3 million, Wuzhou Xinchun plans 980,000 screw sets a year in Zhejiang, Beite is building a ¥1.85 billion base in Kunshan, with tens of billions of yuan of screw capex announced since 2025. One market-research shop sizes humanoid roller screws at $421 million in 2025 going to $7.9 billion by 2032; it is a single-source estimate.

Motors, feedback, and the rest. High-end frameless torque motors are led by Nidec, Kollmorgen, Parker Hannifin, Kinco, and Guangzhou Haozhi, with Germany's TQ-RoboDrive in the precision niche; at the broader servo level the volume leader is Chinese: Inovance, with 28–32% of China's servo-system market. Joint feedback runs through Renishaw's RLS magnetic encoders and Heidenhain's optical ones. Six-axis force-torque sensors for wrists and ankles come from ATI and FUTEK in the West, Hypersen and XJC in China, with Mitsumi's 9.6-millimeter fingertip sensor the smallest and a Bosch MEMS program that wouldn't qualify before 2028. Cross-roller and thin-section bearings run through THK, HIWIN, and NSK; China holds about 20% of the high-end precision-bearing segment; the 40% figure you sometimes see counts commodity bearings. The dexterous hand (tendons, 2–8 millimeter bearings, microdrives) has no settled supplier landscape at all yet.

At humanoid tolerances, the two-firms pattern is literal: strain-wave is effectively Harmonic Drive plus one climbing challenger; roller-screw grinding is GSA/Rollvis plus Schaeffler/Ewellix; RV is Nabtesco's. "N times the combined output of the two firms that make them to tolerance" is not rhetoric: there really are about two per layer. Tesla supplies the revealed preference of the actor with the most design freedom and capital: fourteen harmonic reducers and fourteen roller screws in one robot. It chose the two most supply-concentrated components on the menu because, in the 10–100 Nm band where a humanoid lives, nothing else delivers the torque density in the package. The company best placed to route around the chokepoint routed into it.

The flexspline, the thin steel cup at the heart of a strain-wave reducer, survives by flexing millions of times without cracking. It is a fatigue-critical part, and a robot maker will not put a new supplier's fatigue-critical part inside a warranty until it has life-test history. Warranty history cannot be bought, and announcements do not create it. The industrial world already runs formal gates for exactly this: automotive PPAP under IATF 16949, aerospace AS9145 with AS9102 first-article inspection: months of documentation, statistical process control, and measurement-system analysis per part before volume, and a rotating fatigue part is the hard case. For reducers specifically, no public figure exists for what qualification costs or how long it takes. The defensible bracket runs from an automotive-style two to four quarters at the low end to the multi-year life-test history the incumbent parts actually have. Even the low end sits far above the quarters in which nameplate capacity gets announced.

A million units of unqualified nameplate do not ship inside a warranted, financeable robot; they ship into the grey market or the domestic warm base. Unitree and AgiBot have qualified domestic Chinese reducers for production precisely because they own both ends of the qualification loop. The flood serves the warm base first, and for everyone else it lengthens the queue, because someone has to sort the flood, and sorting is qualification by another name. A glut of unqualified supply raises the value of the stamp that says this piece will not fail inside someone's home.

Buying inventory ahead of the queue creates another timing risk. Robot generations turn over in roughly two years, components are not fungible across generations, and Chinese overcapacity is deflating exactly these parts, so a naked stockpile is a melting asset that strands the day the next platform ships. A spare-parts pool already spoken for under service contracts avoids that trap.

$110per kilogram NdPr price floorThe July 2025 MP Materials deal supplies the first US template, and it is already replicating: a $400 million Department of Defense convertible preferred stake plus a $150 million loan, a warrant, a ten-year price floor on NdPr at $110 per kilogram (roughly twice China spot, structured as a contract-for-difference), and a ten-year DoD offtake for the entire output of the new magnet campus, with a billion dollars of JPMorgan and Goldman financing stacked behind it and a separate $500 million Apple offtake for recycled magnets. The federal government became a shareholder in a minerals company and set a price floor, the clearest available signal that a market has been designated strategic. A price floor is also the specific antidote to the way the last cycle died: after China's 2010 embargo against Japan sent dysprosium from about $90 to over $2,300 a kilogram, the price spike financed Molycorp's $1.25 billion revival of Mountain Pass, and when Beijing relaxed and prices fell, Molycorp went bankrupt in June 2015 under $1.7 billion of debt. The incumbent can end any alternative by relaxing, unless the state removes that lever. Molycorp's estate split into today's MP Materials and Neo Performance Materials.

The capacity behind the template: MP's Independence facility in Fort Worth is producing at about 3,000 tonnes a year, with the $1.25 billion 10X campus in Northlake targeting another 7,000 tonnes from 2028; call it 10,000 tonnes of planned US magnet capacity. A reported $725 million Pentagon loan in June 2026 aims at the metallization gap. And the template is replicating. Vulcan Elements opened its North Carolina facility in March 2025 and landed a $1.4 billion package that November: a $620 million Office of Strategic Capital loan, the largest OSC has made, plus $50 million of Commerce CHIPS equity and $550 million of private capital, with government warrants in Vulcan and in its feedstock partner ReElement. Noveon, the sole incumbent US sintered-NdFeB producer, raised $215 million in January 2026 to expand past 2,000 tonnes. In Europe, Neo Performance opened the continent's first integrated oxide-to-magnet plant in Narva, Estonia, in September 2025, at 2,000 tonnes scaling to 5,000. The Western magnet effort stopped being one company deep sometime in 2025.

The heavy rare earths (dysprosium and terbium, where the April 2025 controls bite hardest) have their own thinner ledger. Lynas became the first heavy-rare-earth producer outside China: dysprosium in May 2025, nine tonnes shipped in a single quarter by late 2025, samarium ahead of schedule, a 5,000-tonne separation expansion underway, and Washington extended it MP-style terms, $96 million of purchases over four years at the same $110 floor, alongside a twelve-year Japanese offtake. Its Texas refinery, though, is stalled pending government funding, and should be described that way. Energy Fuels answered the US question directly: 99.9% dysprosium oxide at White Mesa in August 2025, the first US primary terbium in decades in March 2026, heavy oxides qualified for magnet use that December: demonstrated, with commercial scale still a final investment decision away.

Designing the magnet out is the main substitution route, and the tonnage remains small. Niron Magnetics' iron-nitride plant in Sartell, Minnesota targets 1,500 tonnes a year from early 2027, about a seventh of MP's plan. ZF's magnet-free I2SM motor is production-intent by ZF's own description, with no named vehicle program. Tesla states an intent to cut heavy-rare-earth content in its next drive unit. These are real programs and a rounding error against demand on any published timeline. Recycling (Cyclic Materials, REEtec, Apple's recycled-magnet offtake) is feedstock diversification on the same modest scale.

$12Bcritical-minerals stockpile, Project VaultAs of mid-2026, the reserve covers raw material rather than finished robot parts. The United States keeps a strategic petroleum reserve because 1973 taught it to; the Energy Policy and Conservation Act that created the SPR was signed just two years after the embargo. The rare-earth version has begun: Project Vault, announced February 2026, is a ~$12 billion critical-minerals stockpile, and it stockpiles ore. The defense heavy-rare-earth stockpile covers about six months against the two-to-three-year buffer defense logisticians recommend, and it is dwindling. There is no reserve of finished robot-grade magnets. There is no reserve of precision reducers, though the inventory trap above explains why that one may never make sense as a national stockpile. And there is no crash program qualifying alternative reducers or roller screws anywhere in the Western state apparatus: the state money went to magnets. That absence is the sharpest one on the board.

April 2025. China's export controls put rare-earth magnet shipments under license, and Elon Musk spent part of Tesla's April 22 earnings call explaining that magnet licensing was constraining Optimus production. The controls demonstrated the ability to throttle the exact part counted here. That licensing regime for heavy rare earths was never part of any truce and remains fully in force.

December 2025. MOFCOM issued its first general licenses (multi-shipment, roughly one-year approvals) to major Chinese magnet makers: Jinli, Zhongke Sanhuan, Ningbo Yunsheng. The valve works in both directions, and December showed the open position.

June 2026. China added ten US entities to its own export-control list, including MP Materials and USA Rare Earth: a near-total supply prohibition aimed at the companies building the alternative. The list is now targeting the escape routes by name.

November 10, 2026. The US–China rare-earth truce expires, with a second suspended track expiring November 27. Every magnet-dependent production plan has a scenario branch dated inside the next five months.

January 1, 2027. The Pentagon's ban on Chinese magnets in defense supply chains binds (10 U.S.C. §4872, implemented through DFARS 252.225-7052), covering samarium-cobalt and NdFeB magnets, tantalum, and tungsten if any production stage from mining to fabrication touched China, Russia, Iran, or North Korea. Enforcement runs through contract termination and the False Claims Act, with XRF spot checks. On that date, a cheap Chinese reducer with Chinese magnets inside disqualifies the whole assembly for a defense integrator, whatever its factory-gate price.

January 20, 2027. The EU Machinery Regulation applies. Self-certification ends for machinery with self-evolving behavior; AI safety components move to third-party notified-body conformity assessment; and Europe's notified bodies inherit this work with finite capacity and no robot-ML tooling, on a date with no grace period. How many bodies are actually designated for the AI-safety scope is a number nobody publishes, which is itself informative.

February 28, 2026, already landed. MIIT released its 52-standard system for humanoid robots and embodied intelligence: six domains, 120-plus participating institutions, most standards in drafting, the ISO humanoid-dataset work item pen-held by Chinese institutions. Standards are routes: the definitions of "qualified" decide what may travel to market, and China is writing them first. The EU AI Act's high-risk provisions follow in August 2028.

2.85×times without Chinese supply chainThe price of routing around all of it has already been computed. Morgan Stanley's teardown prices Optimus Gen2 at about $46,000 with its Chinese supply chain and about $131,000 without: 2.85 times, with the actuator line alone at 2.6 times. The same bank's Humanoid 100 puts 73% of the humanoid value chain in Asia and 56% in China, and describes the investable Western exposure as basically Tesla and NVIDIA: two tickers, neither of which makes a reducer. Overcapacity lowers the price of the thing inside the fence. It does nothing to the fence, and the min() rule from the math chapter applies with full force: surplus in the parts layer does not cover a hole in the permission layer.

The 1943 Schweinfurt raids show what happens when a single unglamorous part class is attacked. The Allies decided German industry hung on ball bearings and hit the plants twice: output fell roughly 38% after the August raid and two-thirds after October's, and Albert Speer said afterward that promptly repeated raids would have been disastrous. Germany survived by running every escape route at once (drawing down army bearing reserves, dispersing machinery, importing through neutrals, substituting plain bearings where designs tolerated them, redesigning around the part), and the combination took months to a year while pressure held. Bombs recover on different clocks than licenses, and an export regime does not get tired the way an air force does. Even with that disanalogy, every workaround on the list exists here too, and every one is priced in time. A supply weapon does not need to be permanent to be decisive. It needs to bind longer than the target can bridge, and the Germans bridged with a reserve they had built in advance, which does not exist for these parts.

The flood is probably coming. The last two floods show where the value moves.

$2.35H100 rent, +40% rebound by 2026Solar modules went from about $8.70 a watt in 2010 to under ten cents by 2024. Nearly every Western maker died; First Solar is the rare survivor, and 2025 was its most profitable year ever. The industry got enormous anyway, and the money moved to whoever qualified, financed, and deployed what the flood produced. The first solar securitization ran in 2013, a few years after standardized production data existed to underwrite it. GPUs ran the same movie faster and added a twist: CoreWeave built itself on chip-collateralized debt against take-or-pay contracts while the chips themselves deflated: H100 rents fell from over $7 an hour to a $1.70 contract low by October 2025, then rebounded about 40% to $2.35 by March 2026 on sold-out capacity. Floods overshoot and tighten; the toll moves, it does not vanish. The robot version is already visible at the data layer: a teleoperated demonstration hour fell from about $340 in early 2024 to the $118–136 range by late 2025 (two vendor-reported anchors, direction unambiguous), and the scarce thing became verification.

Substitution is the other route around the reducer, and each branch lands somewhere already on this map. Quasi-direct drive trades the high-ratio gear for a bigger motor, which means more NdFeB, the exact mineral the routes chapter licenses; it relaxes the reducer constraint by tightening the magnet one. Cheap gears plus learned compliance moves the fatigue risk from a steel spline into a software policy that has to hold up under warranty in a stranger's kitchen: a different qualification, on the same clock. Artificial muscle is the genuine multi-chokepoint escape: Clone Robotics' hydraulic Myofiber design uses no reducers, no screws, and less magnet, and as of mid-2026 there is no qualified, deployed muscle-actuated fleet and no published qualification timeline. A muscle-actuated platform qualified for volume before 2030 would invalidate the reducer arithmetic; no such qualification timeline exists.

$13,500Unitree sells a humanoidThe chokepoint map is a schedule. The parts flood eventually; qualification decides who the flood serves and when; the licensing regime decides whose flood crosses which border; and as each layer floods, the toll migrates to the layers a flood cannot touch: qualified capacity inside OEM supply chains, the definition of "qualified" itself, magnet capacity outside one country, and the fleet-utilization records nobody yet collects. The deflation is already visible at the finished-robot layer: Unitree sells a humanoid at about $13,500 while Morgan Stanley's high-income cost curve starts near $200,000, and whoever holds the residual on the expensive fleet eats that slope. When the hardware floods, the scarce assets are the deployments, the records, and the money.