Part VI · 10 min
Part VI: The Open Positions
Seven open positions and the order to take them in.
Each gap is buildable with methods that already exist somewhere else. Aircraft leasing, franchise law, acceptance sampling, price floors: the machinery is old, and robotics just hasn't been run through it yet. Seven positions, in the order the dependencies run; for each one, who is closest to the seat today, what starting it takes, and what it produces for the positions below it.
Most of these positions were bad businesses until recently. Testing and certification grows 4–6% a year, and ratings agencies were mediocre businesses until the 1970s switch from investor-paid to issuer-paid fees turned them into annuities. Their economics improve only if robot fleets reach volume in 2027–2030. If that slips past roughly 2032, builders may exhaust their capital before the positions become valuable.
1. The neutral operator
First in line because it produces what everything below needs: utilization records, trained people, and third-party proof that the work gets done. A company that deploys robots across manufacturers, runs the depots, and owns the customer relationship.
99.30%Formic trailing fleet uptimeThe near-claimants map the vacancy. GXO runs the closest thing, an operational incubator deploying three humanoid brands side by side in live warehouses, the first logistics provider to do it, but the fleet serves GXO's own fulfillment demand. Agility's Digit is the most-deployed platform, over a hundred units across Amazon, GXO, Spanx, Toyota, and Mercado Libre, every one an OEM-direct engagement. Formic has proven the operator economics in fixed automation (zero customer capex, one monthly fee, a stated 20% operating-cost reduction, 99.30% trailing fleet uptime) and says humanoids join its line in 2026–27. Apptronik's Robot Park runs Apollo fleets continuously as a data engine with Google DeepMind, which is the captive version of the same motion, and a demonstration that the OEMs will occupy this seat themselves if no neutral party does. The multi-brand fleet software already ships (InOrbit, across UR, Fanuc, KUKA and custom stacks).
No Western company operates fleets across manufacturers while remaining independent of them. Starting one requires a sales team and depot capex; every component is shipping, and forward deployment with teleop fallback already works. That fallback produces the utilization record required by every financing product below it.
2. The record standard
Telemetry every lender, insurer, and lab can read, shipped as working tooling before any committee votes, because record standards are ratified after the fact. The pattern has run recently and fast: Foxglove's MCAP went from side project to default ROS 2 log format in about two years; LeRobot became the de facto open dataset format on the same clock; C2PA was Adobe's spec first, SLSA was Google's, PCI was Visa's.
$40MFoxglove Series B led by BessemerThe claimants hold pieces. Foxglove is the best-capitalized (a $40 million Series B led by Bessemer in November 2025, with NVIDIA, Amazon, Anduril, Wayve, and Dexterity as customers) and owns the developer substrate. Formant and InOrbit compete on fleet-telemetry ingestion. VDA 5050 hit version 3.0 in March 2026 with OTTO-by-Rockwell certifying against it a month later: manufacturer-neutral fleet communication reaching real multi-vendor adoption. Open-RMF handles heterogeneous multi-fleet traffic. And on the demand side, Koop Technologies already underwrites robotics by pulling hundreds of telemetry fields per second from fleets over API, proof the insurer appetite for a machine-readable record exists and is being fed ad hoc, vendor by vendor.
No one has published a fleet-telemetry schema for the fields lenders and insurers underwrite against: uptime, intervention rates, and residual signals. NVIDIA's Halos took the safety-inspection seat with no dataset scope, so the record and data-quality seats stand open. The party best placed to publish that schema is whoever operates fleets, which is why this position is second and the operator is first.
3. The capital stack
$1.3TUS equipment finance a yearUS equipment finance runs about $1.3 trillion a year, and robots join it the day the records exist. The design is the DSP chassis from Part V with the residual placed correctly: a financing trust holds the machines, the buybacks, and the resale desk; Robot Service Partners put in modest capital and earn on uptime against a published scorecard; spare-parts pools cap the downside.
$400MFramework Ventures tokenization fundToday's claimants are early and partial. Equipment-finance houses already write robotics paper, on a revenue basis, because there is no resale collateral to lend against, which is precisely the gap the records close. Framework Ventures closed a $400 million fund in June 2026 positioned explicitly around tokenization as a financing layer for AI and robotics. No robot-fleet SPV and no rated securitization of an operating fleet has closed anywhere.
A government route-or-hour contract would provide the strongest demand guarantee, but none has been signed. Foundation's Phantom humanoid has won about $24 million of Pentagon SBIR research contracts, including a Phase 3 designation that qualifies it as an approved vendor; vendor qualification buys no robots. DPA Title III carries about a billion dollars of direct financing authority through 2027 plus a five-billion-dollar Industrial Base Fund, and the MP Materials structure shows what the state will do when it decides a layer is strategic. The first government counterparty to guarantee route-or-hour demand (postal, base maintenance, municipal) gives operators the DSP treatment and starts the loss history everyone else will later need.
4. The collection businesses
As generic capture floods, pricing power moves to the rights-clean long tail: applied rules, expert practice, the trades. Somebody has to organize the recording, domain by domain, with consent and rights clean enough that a frontier lab can contract for the output directly.
395humanoids collecting across national facilitiesThe expert-knowledge end already pays. Mercor runs about $2 billion annualized paying over $2 million a day to a contractor network of doctors, lawyers, and scientists at $85-plus an hour; Surge is reported seeking $15 billion on over a billion of revenue. The physical-task end has one pace-setter and it is a state: China's coordinated program has 395 humanoids collecting across four operating national facilities with five more building, 7.19 million data points in a year, on top of roughly forty locally funded centers. The private Western version of that, organized physical-task capture under a gold-standard consent envelope, is unbuilt.
Starting one is a contracts problem before it is a technology problem: a buyer relationship and a clean consenting-capture entity. The defensible verticals are the ones no gig network can reach (licensed trades, clinical settings, regulated kitchens), where the consent envelope is the moat and the rights chain survives a lab's diligence.
5. The exams
$311,000median annual cost, accredited labIndependent tests kept out of training decide whether a model can do a job. Entry is cheap, but maintaining the position is not. Standards-committee seats cost attendance through national mirror committees, and the median initial accreditation assessment fee runs about $7,250. The median total annual cost of maintaining an ISO/IEC 17025 accredited lab is about $311,000, from a dated survey and directionally right. That recurring cost buys the top of the margin stack: the testing-inspection-certification industry runs ~$254 billion at 18–23% EBITDA, UL Solutions runs 23–28%, and the ratings agencies' exam franchises run operating margins around 60%.
$350of equipment, two employees, 1894Underwriters Laboratories is the founding precedent. In 1894, after the electricity displays at the Chicago World's Fair kept catching fire, fire insurers chartered William Henry Merrill's testing lab with two employees and $350 of equipment. The buyers of risk demanded a neutral referee; its mark still gates products into stores, and its commercial descendant is a $17 billion public company. It was buyer-chartered, started before volume arrived, and has owned its scope for a century without a comparable de-novo peer.
Today's seat map: NIST has proposed a humanoid baseline performance benchmark with physical test apparatuses to be distributed to manufacturers and regional facilities: the public-metrology stake. Robocurve, a two-person YC company, is the first private mover on independent continuous robot benchmarking. China's MIIT is drafting the definitions of "qualified" first, which everyone else will inherit or answer. Halos took machine safety; dataset quality, fleet performance, and the robot-hour have no owner. The accreditation clock is the catch that makes this position urgent: 18–24 months, and it starts only after a track record exists, so a lab that matters for a 2030 fleet starts its clock in 2026 or 2027.
6. The robot city
One building serving every layer at once: a showroom where robots work in front of visitors, a depot, a consented capture floor producing the naive-user data no gig network can, and the ground-truth lab the exams position needs.
4,000 m²Beijing Robot Mall, fifty-plus productsChina has already run the format through three generations: Beijing's 4,000-square-meter Robot Mall with fifty-plus products, a Shenzhen "6S" store, and a Wuhan "7S" store that combines experience, rental, purchase, operator training, and maintenance-technician certification under one roof. The Western entries so far are thin and retail-first: a three-day SoHo pop-up, a Manhattan "humanoid experience center" opening as demos-plus-retail. No neutral, public, consent-clean, multi-function building exists in any Western media capital, and no US city or state has yet offered a named incentive package to court one. That is a gap, given what jurisdictions pay for far less anchoring facilities and what Tesla and Figure are building privately (a five-million-square-foot Optimus expansion; a 12,000-unit-a-year line). Fit-out for the building itself runs twelve to twenty-four months, the shortest clock on this list. The first city that hosts it gets the industry's front door, and the operators, the capture floor, and the exams all get their venue in the same transaction.
7. Qualified supply outside China
Strain-wave lines, roller screws, and magnets outside China. The slowest and most expensive position, and the queue math from the parts arithmetic says qualified output gets bought the day it exists.
$1.4Bfederal package for Vulcan ElementsWashington has already written the template (capital in, a price floor, a procurement ban) and the magnet layer shows it working. Vulcan Elements went from a first facility in March 2025 to a $1.4 billion federal package by November; Noveon raised $215 million to expand the sole incumbent US sintered-magnet line; Neo Performance opened Europe's first oxide-to-magnet plant in Estonia. The US remains over 90% dependent on China for finished magnets, so the layer is still early, but it is funded, floored, and multiplying.
The layers where the humanoid arithmetic actually bites have received none of that treatment. In strain-wave reducers, the moves are incumbents adding capacity: Harmonic Drive expanded its Massachusetts operations, Nabtesco is doubling RV capacity, Spinea sits under Timken. No de-novo Western strain-wave line exists, and no DPA-structured crash program has started; the state money went to magnets. Roller screws are European and thin: GSA with Rollvis at over half the global market, Schaeffler with Ewellix running a made-in-USA line. For anyone holding parts exposure meanwhile, the deflation-survival rules from the arithmetic part apply: spare-parts pools against contracted offtake, equity in the chokepoint owners, magnet positions riding the state floor. A naked stockpile melts.
The first qualified Western reducer line inherits everything at once: the floor precedent, the procurement-ban tailwind, and a buyer queue that the qualification chapter's math guarantees. It is also a multi-year clock nobody has started.
8. The window
Reduce the whole list to a scheduling problem and it reads like this. Qualifying a fatigue-critical part runs years. An accreditation clock runs 18–24 months and starts only after a track record exists. A securitization category forms about three years after standardized performance data first exists, measured twice now, solar in 2013 and compute in 2024. A trust layer needs years of accumulated records before anyone rates against them. None of these clocks can be bought down much, and none has been started by anyone in the West.
Against them, the rare-earth truce expires November 10, 2026; the Pentagon magnet ban binds January 1, 2027; the EU Machinery Regulation applies January 20, 2027; China's standards system landed in February 2026 and is drafting definitions now; and the demand scenarios put fleet volume at 2027–2030. Five of the seven positions therefore have start-by dates of 2026 or earlier. Qualified supply outside China already needed to have started.
Three public triggers start years-long clocks: the first manufacturer to warranty a new supplier's reducer at volume; the first fleet-telemetry specification adopted across manufacturers; and the first rated securitization of an operating robot fleet. One question determines whether the stack forms: does the operator's utilization record become portable, or does each manufacturer keep telemetry captive and prevent a neutral schema from reaching the financing market?
Four failure conditions remain untriggered: utilization that cannot service debt at achievable lease rates; generations that become obsolete faster than a resale desk can re-lease them; an architecture shift from precision reducers to cheap gears and software; or platform incumbents financing their own fleets before an independent standard exists. The dependency runs one way: operators produce records, records unlock capital, and capital fields more fleets. The positions needed for a 2030 fleet are being committed now.