Feasibility audit · Contract manufacturing for robots

Robot
Foundry

Can a new company become the Foxconn or the TSMC of robotics? What already exists, the problems you would have to solve, and how hard it really is.

1 2 3
  1. 1Circular spline. Fixed ring with 202 internal teeth.
  2. 2Flexspline. Thin steel cup with 200 teeth. It flexes into an oval.
  3. 3Wave generator. The oval input. One turn shifts the cup two teeth, so a single stage gives 100:1.

Schematic strain-wave gear. Joints like this, with the hands, make up about three-quarters of a humanoid's cost, and US robot makers report shortages of them.

Document
RF-001
Revision
A · first issue
Date
5 Oct 2026
Basis
6 research threads, key claims checked at source
Scope
Humanoids, quadrupeds, mobile robots, arms, drones · US, Europe, China
Verdict
Possible, and hard. 9/10 as pitched. 7/10 by the route in this audit.

Summary

The short answer

A neutral company that builds other people's robots, the way Foxconn builds iPhones and TSMC builds Nvidia's chips, does not exist at scale anywhere yet. Pieces of it are appearing fast, mostly in China. A new entrant can win, but only by starting narrower than the vision.

Does it already exist?

Partly

In China, Apple suppliers already assemble other companies' robots. Lens Technology and Luxshare each built about 3,000 humanoids in 2025, and Foxconn opened a turnkey robot-manufacturing venture in Beijing on 21 Sep 2026. In the US and Europe there are only one-off deals: Jabil for Apptronik, Bosch for Humanoid. Nobody owns the TSMC layer.

How hard is it?

9 / 10

As pitched, this is harder than almost any hardware startup. Volumes are small, the best robot makers build their own robots, and robots have no standard parts to build around. The narrower entry described here is about 7 / 10.

What does it cost?

$25–50M

Estimated capital to reach a first certified US line making 500–2,000 robots a year. Reaching about 10,000 robots a year takes an estimated $150–400M and four to five years. A Foxconn-scale business is a ten-year project.

Where should you start?

US first

Since 28 Jul 2026, new foreign-made mobile robots cannot get FCC authorization in the US without a Department of War exemption. Become the US production partner for robot companies without factories, and win on test, calibration and certification. Standardized actuators and modules come next; that is where TSMC-style margins live.

Timing

Why 2026 is different

Three things changed in the last four months. Together they create demand for US robot manufacturing capacity that most robot companies cannot build alone.

  1. 28 Jul 2026 · FCC

    New foreign-produced mobile robots are blocked from the US market. The FCC added "advanced robotic devices" made abroad, such as humanoids, quadrupeds, warehouse AMRs and delivery robots, to its Covered List regardless of the maker's nationality. New models cannot receive equipment authorization unless the Department of War grants a conditional approval. Previously authorized models and federal purchases are unaffected. Per Sidley's reading of the determination, "foreign-produced" follows the Buy American test: made in the US, with US components above 65% of component cost through 2028 and 75% from 2029.

    FCC fact sheet, 28 Jul 2026 · Sidley, Aug 2026
  2. 3 Sep 2026 · Tariffs

    Drone tariffs of 25–100% took effect. Section 232 duties apply at 100% on drones over 25 kg and on critical components, and 25% on smaller drones. More drone parts follow at 25% from 9 Feb 2027. A parallel Section 232 review of robots and industrial machinery, opened in Sep 2025, has not yet produced a decision.

    KPMG, Aug 2026 · STR on the robotics 232 review
  3. Sep 2026 · Shortages

    US robot makers report they cannot source enough parts. Apptronik's CEO described major shortages of gears and put actuators at up to 60% of a humanoid's bill of materials. Tesla's CEO said on the July earnings call that for Optimus "there is no supply chain."

    Humanoids Daily, Sep 2026 · Tesla Q2 2026 call
  4. 2025–2026 · Capital

    Money is plentiful. Robotics startups raised a record $15B in 2025 and $18.8B by 22 Jun 2026. Investors also pay for factories framed as reindustrialization: Hadrian, an automated machining company, raised $1.37B at a $7.87B valuation in Aug 2026.

    Crunchbase, Jun 2026 · CNBC, Aug 2026

The FCC rule turns US manufacturing from a preference into a requirement for new mobile robots sold in America. Most robot companies have no factory that can meet it.

Who exists

Is there already a Foxconn for robots?

In China, yes, in an early form. Elsewhere, no. The table groups the players by what they actually do for other robot companies today, with my rating of how much each one threatens a new entrant.

PlayerWhat it does for other robot makersScale and statusThreat to you
Contract assemblers already at volume · China
Lens TechnologyChangsha · Apple supplier Whole-robot assembly plus head and joint modules. Runs a joint venture with AgiBot and took over AgiBot's humanoid line-up in 2025. More than 3,000 humanoids and 10,000 quadrupeds assembled in 2025. Named first-batch core supplier to a North American robot maker in Jan 2026. NBD 4
LuxshareDongguan · Apple supplier Makes core parts (harmonic gears, tendon-driven hands) and does final assembly, all for outside customers. Investor and industrial partner of EngineAI. About 3,000 humanoids in 2025. Told investors in May 2026 it will not invest heavily while the industry is immature, and that robots need standard components. Sina 4
Lingyi iTechShenzhen, Beijing · Apple supplier Runs the full chain on one flexible line: encoders, drivers and sensors, then hand and limb modules, testing and full-robot integration. More than 30 robot clients. Capacity plan of 10,000 robots a year in 2026 rising to 500,000 by 2030. 10jqka, Sep 2026 · TNW 4
Dongfang Precision × LejuGuangdong Leju designs the robot and Dongfang builds it. Dongfang owns 2.8% of Leju, an early version of the foundry-plus-equity deal. A 10,000-a-year line producing one robot every 30 minutes, with 77 inspection points (Mar 2026). Interesting Engineering 3
HuaqinShanghai · phone and PC maker Manufacturing partner for Dexmal's data-collection robot. Also builds its own robots. First mass-production delivery in Feb 2026. Sites in Vietnam, Mexico and India. China Daily 3
Foxconn
Foxconn / Hon HaiTaiwan, China, US, Vietnam Futekang, a Beijing venture launched 21 Sep 2026 with IAT, sells robot startups a turnkey path from R&D and prototypes to mass production. Seven partners signed on day one. Robot lines in Vietnam reach full production in Nov 2026. Led a ~$200M investment in Agility. Its own humanoids are still pilots: fewer than 10 worked on a Houston server line in Apr 2026. It also builds robots of its own, which weakens its neutrality. 36Kr · TechNode · Instrumental 5
Western contract manufacturers · one-off deals
JabilFlorida Apptronik's "worldwide manufacturing partner" for the Apollo humanoid since Feb 2025. Builds Symbotic's warehouse robots. Approaching 20,000 Symbotic bots by Sep 2026, its real robot volume. No public evidence of Apollo volume production yet. Apptronik · Jabil Q4 FY26 call 4
BoschGermany Contract manufacturer and design-for-manufacturing partner for Humanoid, the UK startup. Also an investor. Signed May 2026 after a proof of concept. Robotics Tomorrow 3
MagnaCanada, Austria Exclusive contract maker of Serve's delivery robots (2024). Its 2024 Sanctuary AI deal never reached production, and Sanctuary moved to software in Jun 2026. Up to 2,000 Serve robots planned. Benzinga 2
Flex, Celestica, SanminaUS, Canada Flex makes key components for Universal Robots. The others show little or no robotics business. Focused on AI data-center hardware. Flex, Apr 2026 2
Actuator and module suppliers · competing for the TSMC layer
SchaefflerGermany Joint actuators for Humanoid and Neura. Will supply more than half of Humanoid's joint actuators through 2031, a "seven-digit" volume. Humanoid, May 2026 3
Hyundai MobisKorea Actuators for Boston Dynamics' Atlas, its first robotics customer. Grippers, controllers and battery packs are next. Hyundai plans a US robot plant for 30,000 units a year by 2028. Boston Dynamics, Jan 2026 3
Tuopu, SanhuaChina · auto suppliers Rotary and linear actuators for Tesla Optimus. Tesla does final assembly itself. Reported first orders of about 5,000 sets, from anonymous sources. Tesla North 3
Captive factories · unlikely customers
Tesla, Figure, 1X, Agility, Hyundai–Boston DynamicsUS, Korea Build their own robots and most critical subsystems: actuators, hands, battery packs. Tesla's Fremont line is designed for 1M a year. Figure's BotQ went from one robot a day to one an hour in under 120 days and had delivered 350+ robots by Apr 2026. Agility's RoboFab is sized for 10,000 a year. Figure · Tesla Not a customer
Unitree, AgiBot, UBTech, XPeng, BYDChina Vertically integrated. Unitree makes about 86% of its G1's components in-house. Unitree listed in Shanghai on 19 Aug 2026 at about a $9B valuation and rose about 460% on day one. AgiBot passed 10,000 robots built in Mar 2026. CNBC Not a customer
Adjacent "new industrial" startups
Hadrian, Divergent, Machina Labs, Bright Machines, Vention, Re:Build, FictivUS, Canada Precision parts, 3D-printed structures, sheet metal, automation cells, and on-demand sourcing. None does whole-robot assembly, calibration and delivery. Possible suppliers or partners. Fictiv was bought by MISUMI for $350M in 2025. The Robot Report 2

Closest to your idea today

Foxconn's Futekang. It offers robot startups the full path from prototype to mass production. It sits in Beijing, so robots built there count as foreign-produced and cannot get FCC authorization for new US models without a Department of War exemption. Taiwan's government said in May 2025 that it would back a TSMC-model robot maker. No such company is visible yet.

What nobody is doing well

A neutral foundry that never sells its own robot. US-compliant production for the 100–10,000-units-a-year band, which is too small for Foxconn and too complex for parts marketplaces. Robot-specific calibration, burn-in and certification. A catalog of standard, pre-qualified actuators and modules. After-sale repair, refurbishment and field service.

The foundry test

Why TSMC and Foxconn worked, and whether robots pass the same test

Both models grew out of specific conditions. Robotics in 2026 meets none of them fully outside China. It looks like the chip industry before TSMC was founded in 1987, when every chipmaker also ran its own fab.

Today Maker A Maker B Maker C Design & AI Actuators Electronics, battery Assembly Test, calibration Design & AI Actuators Electronics, battery Assembly Test, calibration Design & AI Actuators Electronics, battery Assembly Test, calibration Customers
Every maker builds every layer, so every maker carries its own factory, actuator line and test stations at low utilization.
With a robot foundry Maker ADesign & AI Maker BDesign & AI Maker CDesign & AI DESIGNS + ORDERS ROBOT FOUNDRY Standard modules: actuators, battery packs, safety Assembly Test, calibration, burn-in Certification, software provisioning Customers get ready-to-work robots
Makers keep design and AI. The foundry spreads modules, assembly, test and certification across all of them. The highlighted layer is where TSMC-style margins come from.
PreconditionWhat chips and PCs hadRobots, Oct 2026Result
A standard interface between design and manufacturing Mead and Conway's design rules (1980), then process design kits, let any chip designer target any compliant fab. PCs had standard buses and modules. None. Every robot is a bespoke mechanical and electrical stack. Optimus has about 10,000 unique parts. There is no product safety standard for humanoids yet; ISO 25785-1 is still a draft. ✕ Fail
Many customers who cannot justify their own factory TSMC served 534 customers with 12,682 products in 2025. Many companies, with more than 150 humanoid makers in China alone. The leaders are building their own plants. ◐ Partial
Creditworthy anchor customers with visible demand Apple, Nvidia and AMD: large, profitable and predictable. Mostly venture-funded startups. 69% of humanoids shipped in H1 2026 went to research, education, exhibitions and government data centers. Agility's $300M order book is one customer. ✕ Fail
A process advantage the manufacturer owns TSMC's process nodes. Foxconn's in-house connectors, enclosures and scale. Final assembly is easy to copy: Tesla cleared an old car line for robots in 46 days. Real advantage is available in actuators, test, calibration and certification. ◐ Partial
Stable, high-volume designs One iPhone design a year at hundreds of millions of units. About 13,000 humanoids shipped in 2025. Designs change every 12–18 months. ✕ Fail
Neutrality: never compete with customers TSMC makes only its customers' chips, so customers share their roadmaps. Foxconn, Huaqin, BYD and Hyundai build their own robots. Picea, iRobot's manufacturer, now owns iRobot. The position is open. ◐ Open
A dense cluster of suppliers and talent Hsinchu for chips. Shenzhen for electronics. China made about 85% of 2025's humanoids and quadrupeds. The US may leave about 1.9M manufacturing jobs unfilled through 2033. ✓ China✕ US
0 of 7 preconditions are fully met in the US; three are partly met. The first row matters most. Whoever writes the standard interface for robots, the equivalent of a chip process design kit, becomes the TSMC of robotics. TSMC · VLSI Project · Tech Briefs · Deloitte · Focus Taiwan

Market

Fast growth from a tiny base

Humanoid shipments grew more than fivefold in a year, and forecasts for 2030 run to the high hundreds of thousands. By units, the whole category is still about 2% of the industrial-robot market.

Humanoids shipped, 2025

~13,000

AgiBot 5,168 · Unitree 4,200 · UBTech about 1,000. Chinese makers dominate.

Omdia via TechNode, Jan 2026

Humanoids shipped, H1 2026

~25,000

Up 432% year on year. 69% went to research, education, exhibitions and government data centers.

IDC via Humanoids Daily, Sep 2026

Industrial robots installed, 2025

600,000+

Up 11%. China installed 354,000 (59%). The US is second at about 38,500.

IFR World Robotics, Sep 2026

Humanoid forecast, 2030

750k–890k

Goldman Sachs expects 75,000 in 2026 and 6.5M units worth $138B by 2035. IDC expects 750,000+ in 2030.

Goldman via 24/7 Wall St, Sep 2026

Robotics venture funding, 2026

$18.8B

Raised by 22 Jun 2026, already above all of 2025 ($15B, itself a record).

Crunchbase, Jun 2026

Home cleaning robots, 2025

32.7M

The one robot category already at Foxconn scale. It is built mostly by Chinese brands and ODMs.

IDC, Mar 2026

Robots made or sold in 2025, by kind

Units

All humanoids together came to about 2% of industrial-robot installs in 2025. IFR counts only about 7,000 full-size humanoids sold for commercial work, because it excludes research units and robots under 140 cm. Plan capacity that can also serve quadrupeds, mobile manipulators, warehouse robots and drones.

Supply chain

Where the money and the choke points are

A humanoid's cost sits in its joints. The parts that matter most are also the ones that are scarce in the US or controlled by China.

Where a humanoid's bill of materials goes

Share of cost

CMBI's estimate for Tesla's Optimus: about RMB 318,000 (≈$44k) per robot today, falling about 70% at scale. The parts that move, hands plus the two kinds of joint actuator, make up about three-quarters of the cost. Other analysts put actuators alone at 44–60%. CMBI, Aug 2026

  • Actuators and precision gears Critical

    US makers report gear shortages. Two firms, GSA-Rollvis and Schaeffler's Ewellix, supply over 70% of planetary roller screws, and the thread-grinding machines that make them are back-ordered. China's Leaderdrive leads harmonic reducers at home with 27.5% share, at prices 40–60% below Japanese rivals.

  • Rare-earth magnets Critical

    China makes about 94% of the world's sintered rare-earth magnets. Its April 2025 export licensing is still in force and delayed Tesla's Optimus; wider controls are suspended only until 10 Nov 2026. Non-China supply is starting: Neo's plant in Estonia began commercial production in Sep 2026, and MP Materials and Vulcan Elements are scaling with Pentagon money.

  • Cost of a non-China build Serious

    Morgan Stanley estimates an Optimus Gen 2 at about $46,000 with Chinese suppliers and about $131,000 without them; actuators alone go from $22,000 to $58,000. Unitree's G1 sells from $16,000.

  • Hands Serious

    The least durable subsystem. The forearm was Figure 02's top hardware failure point at BMW. Treat hands as wear parts, with spares and refurbishment planned in.

  • Compute Moderate

    Over 80% of humanoids run on NVIDIA's platform; the Jetson AGX Thor developer kit costs $3,499. A small share of cost with little room to add value. Secure your allocation.

One trap to avoid. Goldman Sachs found nine Chinese suppliers building capacity for 100,000 to 1 million robot sets with no confirmed large orders, and Leaderdrive's gross margin has fallen from 52% in 2021 to 32% in H1 2026. Buy commodity parts; do not build capacity for them. Put your capital into actuator assembly, test and calibration. Humanoids Daily on Goldman · 36Kr on Leaderdrive · 36Kr on roller screws · IEA · Morgan Stanley via SCMP · Neo · Figure at BMW · ABI via humanoid.guide

Economics

Assembly pays 3–8%. Owning the process pays 50%.

Foxconn earns thin margins on enormous, steady volume from a few creditworthy customers. TSMC keeps half of every dollar as operating profit because customers design to its process and cannot easily leave. A robot foundry that only assembles will be priced like Foxconn.

Operating margin, latest fiscal year

Percent of revenue

Celestica, Flex and Magna figures are adjusted; Jabil is core margin. Margin in a robot foundry has to come from what customers cannot copy: standard modules, test and calibration know-how, certification and service.

What happened to manufacturers who bet on startups

Precedents

Magna Steyr and Fisker

Magna built the Fisker Ocean in Austria. Fisker filed for bankruptcy in Jun 2024. Magna booked $316M of charges plus about $75M of supplier obligations, and cut about 500 jobs.

One startup customer can erase years of margin.

Motor1

Foxconn and Lordstown

Foxconn bought the Lordstown plant for $230M in 2022 to build EVs for startups. Lordstown Motors and Fisker both failed. Foxconn sold the plant for $375M in 2025 to a SoftBank affiliate making AI-server equipment.

Generic factory assets saved it; startup volume never came.

DCD

Picea and iRobot

iRobot relied on one manufacturer, Picea. By late 2025 it owed Picea about $160M. Picea bought iRobot's debt, and after the Dec 2025 bankruptcy it owned the company outright.

A manufacturer that is also the main creditor ends up owning the brand. Customers know this.

The Register

NIO and JAC

NIO paid JAC about RMB 3B in fees from 2018 to 2022 to build its cars, then bought the plants and its own licence in Dec 2023.

Successful customers take production in-house. Plan for it in pricing and contracts.

CnEVPost

Fitbit and Flex

Fitbit paid for and owned all tooling, on a one-year contract with 90- to 180-day notice. When demand fell in 2016, Fitbit carried the write-downs and owed Flex for unused capacity.

This is the contract template to copy: customer-owned tooling and component liability.

Fitbit 10-K

TSMC's start

Founded in 1987 with 48% of its capital from the Taiwanese government and 27.5% from Philips, which also contributed process technology. In 2025 it earned a 50.8% operating margin on $122B of revenue.

The foundry model needed patient state and strategic capital, plus a standard interface.

TSMC history · Q4 2025 results

Problems to solve

The 14 problems you have to solve

Ranked by how likely each one is to kill the company. The first four are existential. Each comes with what solving it looks like in practice.

  1. 01

    Demand is real, but tiny, and most of it is not work yet

    Critical

    Why it is hard

    About 13,000 humanoids shipped in 2025 and about 25,000 in the first half of 2026, and 69% of those went to research, education and exhibitions. Agility, the most-deployed US humanoid, booked $1.8M of revenue in 2025 and plans about 800 units in 2027. A factory needs steady volume; robotics has forecasts.

    What solving it looks like

    Flexible lines that can run humanoids, quadrupeds, mobile manipulators, warehouse robots and drones on shared stations. Pricing built on engineering fees plus per-unit fees, so idle capacity does not sink you.

    IDC · The Robot Report

  2. 02

    The biggest robot makers will never be your customers

    Critical

    Why it is hard

    Tesla, Figure, 1X, Agility, Hyundai–Boston Dynamics, Unitree, AgiBot, UBTech, XPeng and BYD all build in-house. Tesla and Figure say they do it because no supply base exists, and once they have their own plants they will not come back.

    What solving it looks like

    Sell to the second tier: well-funded startups without factories, AI-model companies with no hardware (Physical Intelligence, Skild AI, Google DeepMind) that need fleets of reference robots, and leaders that want second-source or overflow capacity.

    Figure on BotQ · Tesla Q4 2025 update

  3. 03

    Your customers can go bankrupt holding your money

    Critical

    Why it is hard

    Most customers are venture-funded, and China's planning agency has already warned of a humanoid bubble. At 3–8% margins, one customer default wipes out years of profit, as Magna learned with Fisker and Foxconn with Lordstown.

    What solving it looks like

    Customer-owned tooling, deposits and engineering fees up front, take-or-pay volumes, component liability clauses on the Fitbit–Flex model, and a hard cap on exposure to any single customer.

    Bloomberg on the NDRC warning

  4. 04

    Robots have no standard interface to build around

    Critical

    Why it is hard

    TSMC works because chip designers design to its rules, so one process serves hundreds of customers. Every robot today is custom: Optimus has about 10,000 unique parts, designs turn over every 12–18 months, and Luxshare says the industry still lacks standard components.

    What solving it looks like

    Write the interface yourself: standard actuator families, battery packs, a safety controller, a wiring and communication architecture, design-for-manufacture rules and test specifications. Make designing onto your platform cheaper and faster than designing around it.

    Yahoo on Optimus parts · Luxshare investor Q&A

  5. 05

    The incumbents have already started

    Serious

    Why it is hard

    Foxconn launched a turnkey robot venture in Sep 2026 and starts robot lines in Vietnam in Nov 2026. Jabil, Bosch and Schaeffler have signed robot deals, and Lingyi plans capacity for 500,000 robots a year by 2030.

    What solving it looks like

    Work in the gaps incumbents ignore: US-compliant production, 100–10,000 units a year, many product types, fast new-product introduction, and robot-specific test. The window is roughly two to three years.

    36Kr on Futekang · TechNode on Vietnam

  6. 06

    Actuators and gears are the choke point

    Serious

    Why it is hard

    Hands and joint actuators are about three-quarters of a humanoid's bill of materials, and US makers report gear shortages. Figure had to build more than 9,000 actuators itself. Two firms supply over 70% of planetary roller screws, and the grinding machines that make them are back-ordered.

    What solving it looks like

    Assemble, test and characterize actuators in-house from day one. Qualify two allied sources for gears, screws, bearings and motors, and reserve their capacity years ahead.

    Apptronik CEO · Figure ramp report

  7. 07

    China can cut off your magnets

    Serious

    Why it is hard

    China makes about 94% of the sintered rare-earth magnets inside robot motors. Its 2025 licensing delayed Tesla's Optimus. A wider set of controls is suspended only until 10 Nov 2026, and the US defense bill in conference would phase in China-free magnets for military work.

    What solving it looks like

    Offtake agreements with non-China magnet makers such as MP Materials, Vulcan Elements, e-VAC and USA Rare Earth; qualify rare-earth-free motors such as Niron's; hold a strategic buffer.

    Mining Technology · The Register on Optimus · Columbia CGEP

  8. 08

    China builds the same robot for a fraction of the price

    Serious

    Why it is hard

    Morgan Stanley estimates the same humanoid costs about 2.8 times as much to build without Chinese suppliers. Unitree's G1 starts at $16,000 and Booster's K1 at $4,999, while Agility's Digit v4 costs about $125,000 to build.

    What solving it looks like

    Do not compete on unit cost. Compete on what China cannot offer US buyers: FCC and defense compliance, IP security, speed, proximity and quality data. Close the cost gap over time with design-for-manufacture and automation.

    Morgan Stanley via SCMP · Agility SPAC filing

  9. 09

    Margins are thin and capital is heavy

    Serious

    Why it is hard

    Assembly earns 3–8% operating margin. Tooling costs an estimated $2–15M per robot generation, and factories, test equipment and long-lead parts must be paid for before revenue arrives.

    What solving it looks like

    Customers fund tooling. Use public capital: Pentagon Office of Strategic Capital loans, and the 2025 tax law's 100% write-off for new factory buildings started before the end of 2028. Earn margin in modules, test and services.

    IRS · Formlabs on tooling

  10. 10

    Testing and calibrating robots is its own discipline

    Serious

    Why it is hard

    A phone is tested on a fixture in seconds. A humanoid needs joint-by-joint calibration, full-body checks and burn-in. Figure runs more than 50 in-process inspections and 80 functional tests per robot, and still reports end-of-line first-pass yield just above 80%.

    What solving it looks like

    Build your own test, calibration and burn-in stations and the data system behind them. This is your process technology, the part customers cannot easily copy.

    Figure, Apr 2026

  11. 11

    Safety certification is unsettled

    Moderate

    Why it is hard

    No product safety standard for humanoids exists yet. Arms follow ISO 10218 (2025), service robots UL 3300, and the EU Machinery Regulation applies from 20 Jan 2027, with third-party checks for safety functions that use machine learning. Budget an estimated $0.5–2M and 12–24 months per new robot.

    What solving it looks like

    Pre-certified safety controllers and battery packs, an in-house pre-compliance lab, and a repeatable certification pipeline you can sell as a service.

    ISO 10218 · BAuA on the EU regulation · UL 3300

  12. 12

    Customers have to trust you with their designs

    Moderate

    Why it is hard

    Foxconn, Huaqin, BYD and Hyundai all build their own robots, and Picea ended up owning iRobot. Robot makers will not share roadmaps with a likely future competitor.

    What solving it looks like

    A public pledge, as TSMC made, never to sell your own robot. Separate teams per customer, and clear IP and tooling ownership in every contract.

  13. 13

    Talent and ramp speed

    Moderate

    Why it is hard

    The US may leave about 1.9M manufacturing jobs unfilled through 2033. Even Tesla missed its 2025 Optimus goal of about 5,000 units and built a few hundred. Engineers who have ramped a robot line are rare.

    What solving it looks like

    Hire from EV, aerospace and automotive programs, automate test from the start, and locate in an existing manufacturing cluster.

    TipRanks on Optimus

  14. 14

    "Ready to work" means owning uptime

    Moderate

    Why it is hard

    Buyers pay for working robots. In the US, robots-as-a-service is now the main business model for service robots, and Formic, a robot leasing firm, advertises 99.3% uptime. Spares, repair and field service are where many robot programs quietly fail.

    What solving it looks like

    Offer deployment, field service, spare parts and refurbishment as a paid layer, or partner with a field-service company.

    IFR, Sep 2026 · Formic

Difficulty

How hard is it? That depends on which company you build.

"Robot foundry" covers five different businesses. My ratings weigh capital, competition, customer risk and time to revenue.

Robot Foxconn

Assemble anyone's robots at high volume. Needs $0.5–1B+, anchor customers that do not exist yet, and a fight with Foxconn, Lens and Luxshare at 3% margins.

9.5

Robot TSMC

Standard actuators and modules sold to many robot makers. Deep engineering and capital, and you must win against Schaeffler, Hyundai Mobis and Chinese suppliers while makers keep actuators in-house.

9

US robot foundry for startupsRecommended

Prototype-to-volume production for 100–10,000 robots a year per customer, built to pass the FCC rule. The demand is new and the gap is visible, but customers are fragile.

7

Defense drones and ground robots

Pentagon-compliant production and components. Demand is funded by programs and tariffs, but buying is lumpy and competition is crowded.

6.5

Robot readiness services

Test, calibration, certification, deployment and field service, without owning assembly. Low capital and fast revenue, but a smaller prize on its own.

5

What it takes, stage by stage

Estimates
StageCapitalTeamSpaceOutputRevenue
Prove demandMonths 0–6$2–5M6–1210–20k sq ft pilot cellEngineering builds for 3 design partnersEngineering fees
First certified lineMonths 6–24$25–50M60–15060–100k sq ft500–2,000 robots a year$10–40M
The robot design kitYears 2–5$150–400M400–1,000200–400k sq ft10,000–30,000 robots a year, plus modules sold to others$0.3–1B
ScaleYears 5–10$1B+2,000+Several plants100,000+ robots a year$2B+

Benchmarks behind the estimates: Agility's RoboFab is about 70,000 sq ft for 10,000 robots a year. Figure's first BotQ line is sized for 12,000. Foundation builds 300–500 a year at a small San Francisco site. ISO 9001 takes about 6–8 months for a 50-person plant; automotive-grade IATF 16949 takes 12–18 months. CNBC on RoboFab · Figure · JumpCloud

Ways in

Five ways in, and the one to take

Start where the demand is forced and the incumbents are weak, then move down the stack toward the parts every robot shares.

OptionWhat you buildStart capitalMoatMain rivalsVerdict
A · Robot Foxconn High-volume final assembly for anyone $500M+ Scale only Foxconn, Lens, Luxshare, Lingyi, Jabil Do not start here. It can grow out of B and C later.
B · Robot TSMC Standard actuators, battery packs and safety controllers: the robot design kit $100–300M High, once designed in Schaeffler, Hyundai Mobis, Tuopu, Sanhua, Harmonic Drive The end goal. Build it from what repeats across your first customers.
C · US startup foundryStart here Prototype-to-volume manufacturing, test and certification for robot companies selling in the US $25–50M Compliance, process know-how, customer lock-in Jabil, Bosch, Magna, in one-off deals Best entry. The FCC rule creates the demand and incumbents are not organized for it.
D · Defense drones and UGVs Pentagon-compliant drones, ground robots and components $10–30M Compliance and clearances Neros, Unusual Machines, Skydio, Re:Build Strong parallel cash engine. Drone Dominance is a $1.1B program and drone tariffs are in force.
E · Readiness services Test, calibration, certification, deployment, field service $3–10M Know-how and data Formic, integrators, field-service startups Bundle with C from day one. It is the fastest revenue and the real differentiator.

Capital figures are estimates. Drone sources: Breaking Defense · KPMG.

Recommended path

From startup to robot foundry, in four phases

Each phase has a gate. Do not raise or build for the next phase until you pass it.

Phase 0 · months 0–6

Prove demand

  • Interview 30 robot makers and 10 AI-model companies.
  • Model a compliant bill of materials for two reference robots against the 65% US-content test.
  • Sign supply agreements for magnets, gears and motors.
  • Run engineering builds in a small pilot cell.
$2–5Mseed

Gate: three customers pay engineering fees and commit to 2027–28 volumes.

Phase 1 · months 6–24

First certified line

  • 60–100k sq ft US plant with actuator assembly, final assembly and end-of-line test.
  • ISO 9001, then a safety and pre-compliance lab.
  • Three to five customers, 500–2,000 robots a year.
  • Proprietary test and calibration software from day one.
$25–50MSeries A

Gate: first-pass yield above 90% and two customers in repeat production.

Phase 2 · years 2–5

The robot design kit

  • Standardize what repeats across customers: actuator families, battery packs, safety controller, harness.
  • Sell modules to robot makers who assemble elsewhere.
  • 10,000–30,000 robots a year.
  • Finance with equity, government loans and customer prepayments.
$150–400Mgrowth

Gate: five or more customers designing new robots on your modules.

Phase 3 · years 5–10

Scale

  • Second plant, US or allied.
  • Joint ventures for magnets and gears.
  • 100,000+ robots a year.
  • Become the default way to build a robot outside China.
$1B+scale

Gate: volume from repeat customers covers fixed costs.

Your first 90 days

0 of 7 done

Where to build

2026–2030
LocationRoleWhy
United StatesFinal assembly, actuators, battery packs, circuit boards, harnesses, testRequired for new FCC-covered mobile robots and for defense work. Highest cost, best customer access, strongest incentives.
Japan, Taiwan, Korea, EUPrecision screws, reducers, bearings, motors, sensors, chipsAllied and high quality, but counted as foreign under the FCC test. Keep them inside the 35% foreign allowance (25% from 2029).
MexicoEnclosures, harnesses, fixed industrial armsLow cost and tariff relief, but Mexican content counts as foreign for FCC-covered robots.
Gulf statesCapital partner and regional assemblyLarge funds and local demand: Saudi Arabia's QSS plans 10,000 humanoids with Humanoid over five years. Chinese ties can create US compliance risk.
ChinaAvoid for US-bound productsLowest cost and the deepest cluster, but effectively closed for new US mobile robots, drones and defense.

Sidley on the FCC test · Arab News

Watch list

Dates that can change the plan

Several decisions are still open. Track these, and decide in advance what would make you stop.

  • 10 Nov 2026

    China's suspension of its Oct 2025 rare-earth rules expires.If it lapses, magnet supply tightens for every robot maker.

  • Nov 2026

    Foxconn's Vietnam robot lines reach full production.The first large non-China robot capacity from an incumbent.

  • Pending

    US Section 232 decision on robots and industrial machinery, a robotics executive order, and the FY27 defense bill.Tariffs could raise both your imported parts and your competitors' costs.

  • 20 Jan 2027

    The EU Machinery Regulation applies.New rules for autonomy, cybersecurity and machine-learning safety functions.

  • 9 Feb 2027

    A 25% tariff on more drone parts takes effect.

  • 1 Jan 2028

    Deadline to apply for a Department of War exemption from the FCC robot rule.Foreign makers that miss it must build in the US or stay out.

  • 2 Aug 2028

    EU AI Act rules for AI built into machinery apply.

  • 2029

    The FCC content threshold rises from 65% to 75% US.

Stop or pivot if

  • Fewer than three robot companies will pay engineering fees or deposits within six months.
  • The FCC rule is narrowed to China only, or Department of War exemptions become routine.
  • Robotics funding in H1 2027 falls below half of H1 2026. Pivot to defense drones or readiness services.
  • You cannot reach 65% US content at a price customers accept.

Sources: Mining Technology · Inside Government Contracts · Addleshaw Goddard

Sources

Sources

Primary sources were checked for the FCC rule, the drone tariffs, IFR, IDC, Omdia, Goldman Sachs, Crunchbase, Lingyi and the rare-earth deadline. Figures marked as estimates are my analysis. Some company figures are single-source and are attributed in the text.

Policy and trade

Market and funding

Manufacturers and robot makers

Components and supply chain

Economics, history and standards