THE JOINT THAT HAS TO SURVIVE

A traceable second source for precision motion | 12 September 2026

A Fieldbook for founders and investors. A first-person opportunity narrative, checked against technical, company, regulatory and market evidence.

I am interested in strain-wave reducers for the same reason I am interested in bridges: the drawing is easy to admire, but the business is decided by what happens after the millionth cycle.

01 · The possibility

THE SMALL PART INSIDE THE BIG MACHINE

Robots, semiconductor tools, optical instruments and spacecraft all need to move accurately while carrying force. The component that makes that possible is often invisible: a compact reducer between a motor and a joint.

A strain-wave reducer can provide a large reduction in one compact stage, with high torque density and a design intent of very low lost motion. That is valuable wherever a motor must become a precise, strong, small movement. Harmonic Drive describes its technology for industrial robots, semiconductor equipment, medical equipment and aerospace. Harmonic Drive technology

The founder opportunity is not “copy a famous gear.” It is to become a trusted second source for one defined axis: the supplier who can show the material lot, the process history, the measured lost motion, the endurance spectrum and the failure analysis.

That is a smaller claim than owning the humanoid supply chain. It is also one a startup might actually prove.

02 · The mechanism

THREE PARTS, ONE CONTROLLED DEFORMATION

The mechanism is easiest to understand as three pieces. An elliptical wave generator sits inside a thin, flexible cup called the flexspline. The flexspline engages an outer circular spline at two regions around the long axis of the ellipse. The two gears have a small tooth-count difference, so one turn of the wave generator produces a much slower relative movement.

The reducer wins by deliberately deforming metal. That deformation creates the same risks that create the compactness: alternating stress in the cup and diaphragm, tooth-root stress, bearing load, sensitivity to alignment and a catastrophic overload mode called ratcheting. Harmonic Drive says its S-tooth design can engage up to 30% of the teeth simultaneously; that is an incumbent product claim, not a free recipe for a new entrant. Harmonic Drive technology

The original Musser patent was filed in 1955 and issued in 1959; the public record marks that foundational patent expired. That does not make the field unpatented. Later tooth profiles, materials, manufacturing processes and product arrangements still require a freedom-to-operate claim chart. US 2,906,143

I would measure lost motion or backlash under a stated torque, temperature and fixture. I would not write “zero backlash” as though it were a measured property of my product.

03 · The corrected market

THE HEADLINE MARKET DOES NOT CLEAR THE FIRST TEST

The supplied story says the global market is about $4.8bn in 2026 and more than $7.8bn by 2030. I cannot verify those figures from a primary source with a clear definition. They may include actuators, other reducer mechanisms, services and channel revenue. I treat them as an ungrounded scenario, not a total addressable market.

The public company anchors are more useful, and more limited. HDS reports FY2024 consolidated sales of ¥55.645bn, but it sells speed reducers and mechatronics and does not isolate strain-wave revenue. HDS FY2024 profile

Nabtesco reports ¥79.3bn for its recast FY2025 Component Solutions segment and attributes ¥11.7bn of year-on-year growth to precision reduction gears. Its filing also records the hydraulic-equipment business as discontinued from Q3. This is segment scale, not SWG-only revenue. Nabtesco’s famous RV product is planocentric/cycloidal; a separate German entity advertises strain-wave products. Nabtesco FY2025 results Nabtesco RV products Nabtesco strain-wave gears

IFR’s public page shows the scale and geography of industrial robotics, but the exact 2024 installation count was not independently readable in this pass. It is directional context, not a reducer-demand input. IFR World Robotics 2025

04 · Where the first customer lives

DO NOT START WITH A HUMANOID BOM

The first customer is more likely to be a robot, motion-control or equipment company with one painful axis than a humanoid manufacturer promising millions of units. I would look for a stable specification, a documented dual-source requirement or an incumbent lead time longer than 16 weeks, 1,000–20,000 annual units, and no immediate flight-critical liability.

Humanoids are useful as a possibility, not as a forecast. Figure says its BotQ line targets up to 12,000 humanoids per year and 100,000 over four years. It also says critical actuators, batteries, sensors, structures and electronics were designed in-house. Its 35-DoF control claim is not a reducer count: a degree of freedom can use a belt, screw, cycloidal stage, direct drive or a shared actuator. Figure 03 Figure Helix

The jobs-to-be-done are concrete: remove a single-source risk; hit a specified torque, stiffness, efficiency and lost-motion envelope; shorten a customer’s lead time; provide serialised acceptance data; and make a design review easier. The buyer is paying for risk removed, not for the word “harmonic.”

05 · The material and fatigue moat

THE CUP IS THE COMPANY

The hard part is not turning a gear-shaped object. It is keeping a thin flexspline alive while it bends, transmits torque, sees lubrication and tolerates the small errors left by manufacturing and assembly.

Harmonic Drive’s catalog states infinite flexspline fatigue only at rated conditions and gives wave-generator bearing L10 lives of 7,000 hours for one family and 10,000 hours for another. Those are conditional catalog ratings; they are not a universal 10,000-hour dynamic-payload demonstration for my design. The catalog also warns about ratcheting. CSF Mini catalog

Published work identifies diaphragm and tooth-root fatigue as difficult problems. A space-tribology study reports shorter life in vacuum than in air, linking wear to the wave generator and flexspline. That makes lubricant, temperature, contamination and environment part of the product, not footnotes. Strain-wave fatigue study Space tribology study

My process hypothesis is deliberately unglamorous: certified alloy and blank; controlled forming or rough machining; stabilisation and heat treatment; finish geometry; bearing sourcing; washing; assembly preload and grease control; calibrated torque-angle testing; metrology; endurance; teardown. I need hardness, microstructure, tooth profile, runout, surface finish, serial genealogy and change control. If a supplier will not share raw data and retain destructive samples, it is not a partner for a traceability business.

06 · The buyer’s qualification gate

“AEROSPACE-GRADE” IS NOT A MATERIAL

There is no universal aerospace reducer certificate. IAQG 9100 is a quality-management standard used across aviation, space and defense supply chains. NASA distinguishes qualification—showing a design survives its specified environment—from acceptance, a smaller check applied to each flight unit. NASA-STD-7001C is a payload vibroacoustic standard, not a universal gearbox qualification rule. IAQG 9100 NASA Product Verification NASA-STD-7001

The customer must define the drawing, duty cycle, load spectrum, temperature, lubricant, contamination, allowable lost motion, torsional stiffness, efficiency, noise, life confidence and acceptance tests. A flight program may require years, special processes, environmental testing, insurance, export controls and an approved-source process. A non-flight robotics axis can be a much better first proof.

The China story also needs precision. DFARS restrictions apply to specified USML/600-series items and Chinese military-company status; they are not a blanket legal ban on every Chinese reducer. I would get counsel to classify the exact part and obtain the customer’s written provenance requirements. DFARS Subpart 225.7

07 · The outsourced pilot

SELL THE EVIDENCE BEFORE BUYING THE FACTORY

I would begin with a paid NRE and a named customer’s incumbent in the test fixture. Pick one torque/diameter family. Freeze the interfaces and spectrum. Audit two Taiwan, Korea or Malaysia suppliers for equipment, ownership, heat-treatment records, material origin, subcontractors, calibration, genealogy and change notification. “Asian” is not a provenance category.

Run two identical pilot lots and one tolerance-stressed lot. Correlate supplier and founder metrology. Measure lost motion under load, torsional stiffness, ratio error, efficiency, temperature, vibration and noise. Run instrumented endurance against the customer spectrum, then open the failed and surviving parts. Do not extrapolate a 100-hour anecdote to 10,000 hours without a fatigue model and a statistical basis.

The wedge is a traceable, tested, configurable second source for a narrow non-flight robotics axis or space-ground-test actuator. It is not “cheaper Japanese gear,” and it is not a claim that Western buyers are universally prohibited from China. The first deliverable is an evidence package a customer can sign off.

08 · What the economics actually say

A SMALL PILOT CAN BE CHEAP; A QUALIFIED LINE IS NOT

These are planning scenarios, not observed market prices. They show why I would stage capital.

Case	ASP	Manufacturing + outsourced processes	Test/QA + freight	Warranty/scrap reserve	Contribution/unit	Annual volume	Annual contribution before fixed cost
Low	$350	$245	$35	$28	$42	4,000	$168k
Base	$500	$275	$40	$25	$160	80,000	$12.8m
High	$650	$260	$45	$20	$325	600,000	$195m

The low case cannot support a serious validation team. The base case works only after qualification and working-capital financing. The high case needs customer BOM evidence that I do not have. Bare gear ASP is not actuator ASP; motors, encoders, controls, integration and warranty may capture more value.

The supplied $300k–$500k capex claim is not credible for a complete aerospace-grade in-house line. A fabless learning lot is estimated at $60k–$180k. A captive validation cell is $250k–$700k while keeping specialized heat treatment outsourced. An integrated aerospace-grade cell is $1.5m–$4m or more. Public machine and furnace listings are directional, not quotes. 5-axis listing Vacuum furnace listing

09 · The competitor map

THE GAP IS QUALIFICATION, NOT ABSENCE

HDS is a major pure-play strain-wave incumbent with global engineering and application support. Nabtesco is a serious alternative-mechanism supplier through RV and is expanding a separate strain-wave range in Germany. Leaderdrive offers Chinese cup, hat and ultra-flat products, including a claimed third-harmonic variant. Nidec-Shimpo and other Asian suppliers are additional audit leads. HDS installation considerations Leaderdrive products

This map changes the pitch. I cannot say “HDS and Nabtesco are the aerospace-grade SWG oligopoly.” I can say incumbents have process data, special-product experience, qualification history and customer relationships. Nabtesco’s company-estimated roughly 60% share is for medium/large industrial-robot precision reduction gears, principally its RV family—not SWG-only market share. Nabtesco robot reducers

A startup’s defensible position would be a measured application package: faster development, dual-source provenance, customer-witnessed testing and a configuration the incumbent will not prioritise. That is a service-and-evidence moat before it becomes a manufacturing moat.

10 · The 12–24 month attack

EARN THE RIGHT TO SCALE

Months 0–3: interview 20 motion-control, robotics, space-ground and defense-prime engineers. Secure three written specifications and one paid NRE. Freeze one family and spectrum. Build later-patent claim charts, classify the part for export control, and audit two suppliers.

Months 3–9: produce two pilot lots plus one stressed lot. Correlate metrology. Measure lost motion, stiffness, efficiency, thermal behaviour, NVH and ratio error. Run endurance and teardown. Report yield, scrap, Cpk/Ppk, landed cost and warranty reserve.

Months 9–18: run a customer-witnessed incumbent comparison. Deliver serialised acceptance data, nonconformance and corrective-action records. Convert one non-flight design-in. Add equipment only when a demonstrated bottleneck justifies it.

Months 18–24: scale only with a 12-month forecast, pilot yield above 90%, demonstrated life at the customer spectrum, a second qualified source for every critical process and more than 35% gross margin after scrap, test and warranty. Otherwise remain an engineering/test business or stop.

11 · The bear case

THE INCUMBENT MAY BE GOOD ENOUGH

The strongest bear case is not that the mechanism is impossible. It is that the “gap” is a qualification and integration gap, not a supply vacuum. HDS may have acceptable lead times and decades of process data. Nabtesco offers a high-rigidity alternative mechanism and a growing SWG line. Chinese and other Asian suppliers already compete on cost. A robot OEM may want a complete actuator, not my bare gear.

Humanoids may use integrated actuators, belts, screws, cycloidal reducers or direct drive. Aerospace qualification may take years and require special processes, traceability, insurance and customer authority I cannot finance. Outsourcing can expose the process secrets I claim as a moat; buying the process in-house can consume millions. One flexspline crack or undocumented material substitution can create field liability.

So I set explicit falsifiers: stop if no five qualified buyers will provide a specification or paid sample order; stop after two failed pilot iterations; stop if the customer-spectrum life fails, critical-dimension Cpk stays below 1.33, pilot yield stays below 70% during learning, or contribution falls below 35% at a signed-forecast volume; stop if no customer design-in exists by month 18. These are not pessimism. They are the company’s safety system.

12 · The founder decision

BUILD THE TEST CELL, NOT THE MYTH

I would pursue this conditionally. I would fund the $60k–$180k learning lot only after a named buyer provides an interface, load spectrum and paid NRE. I would build a captive validation cell only after two suppliers and the customer’s incumbent make the test question concrete. I would not buy a complete aerospace-grade factory on the strength of a $4.8bn/$7.8bn market report, a 30–40-reducer humanoid assumption, a 100-hour failure anecdote, a universal 10,000-hour claim or a $300k–$500k capex story.

The company I want to build sells proof: this exact reducer, from this material lot, through this process, measured on this calibrated fixture, survived this customer spectrum, with this acceptance record. If a customer pays for that proof and signs the design-in, manufacturing scale becomes a question worth answering. If not, the honest outcome is a small test business—or no company.

13 · The founder’s ladder

LEARN IN PUBLIC, QUALIFY IN PRIVATE

The founder does not need to cosplay as a thirty-year gear veteran. Credibility comes from a narrow question, an external reviewer, calibrated measurement and a clear statement of what each experiment cannot prove. I would hire a contract gear designer, an independent test engineer and a trained metallography laboratory for specialist work while I own the requirements, supplier comparison, data trail and customer loop.

There are four different articles in the development ladder:

Article	Purpose	Boundary
Polymer kinematic demonstrator	A hand-cranked or low-energy printed model makes the tooth-count difference, engagement and ratio obvious	It says nothing about metal stress, torque, wear or life
Engineering mule	A commercial reducer or non-flight metal prototype proves the fixture, controls, interface and teardown workflow	It does not validate the candidate material or process
Pilot article	Serialised units from a controlled lot test repeatability, CTQs and failure modes under a named spectrum	It is not flight acceptance or a 10,000-hour claim
Qualification article	A frozen design is tested in the customer-approved environment with agreed margins and evidence	Qualification belongs to this configuration and programme, not to a marketing adjective

The first 30 days are education and customer discovery: build the polymer demonstrator, read public NASA/ESA test programmes, interview the people who own joints, reliability, suppliers, qualification, returns and procurement, and write a one-page specification packet. Public RFPs are vocabulary and hypothesis material; they are not permission to infer demand.

14 · The rig must be boring

A SHORT TEST IS A SCREEN, NOT A LIFE CLAIM

I would rent a calibrated dynamometer for the first mule. The test record should contain input speed, output torque, angle, temperature at the housing and bearing, lubricant, vibration/noise where relevant, raw time series, stopping criteria and a failure definition. ASTM E467 makes dynamic verification specific to the actual machine configuration and specimen; a transducer calibration certificate alone is not enough. ASTM E467

Only after that fixture is understood would I commission a founder rig: rigid reaction frame, positive retention, interlocked full enclosure, independent emergency stop, overspeed and overtemperature trips, load shedding, physical limits, remote video, data heartbeat and a tested restart inhibit. Rotating parts and stored energy need guarding and lockout/tagout. An unattended overnight run is not a clever shortcut; it requires a written hazard analysis and a demonstrated safe state. OSHA machine guarding OSHA hazardous-energy control

Nor would I follow improvised heat-treatment or Nital recipes. A thin flexspline needs controlled process records, distortion control, hardness/microstructure evidence and witness coupons. Nitric acid is corrosive/oxidising; ethanol is flammable; OSHA has recorded both serious nitric-acid burns and an explosion involving acid and alcohol. Send polished sections to a trained metallography lab with a chemical-hygiene plan instead. OSHA nitric acid OSHA laboratory chemical hygiene OSHA incident

For variable loading, use the customer’s measured history and cycle-count it under ASTM E1049. A Basquin/S–N fit can support interpolation inside a tested regime; ASTM E739 warns against extrapolation outside that interval and against sparse low-percentile claims. Miner’s linear damage sum is a model to validate, not a magic conversion from a 100-hour over-torque run to 10,000 hours. NASA’s reliability guidance likewise separates gross-defect screening from statistically credible reliability evidence. ASTM E739 ASTM E1049 NASA reliability test methods

Every result gets an uncertainty budget and a teardown. Report repeatability, drift, calibration traceability and expanded uncertainty; NIST treats uncertainty as a property of the specific measurement configuration. Record crack location, tooth wear, fretting, bearing damage, grease condition, hardness/microstructure and dimensional drift. A survivor is evidence of survival in one fixture, not evidence that a competing failure mode does not exist. NIST TN 1297

15 · Ask for the packet, then ask for the cheque

THE FIRST CUSTOMER BUYS RISK REDUCTION

The first design partner should be a robotics or motion-control OEM with one painful, non-flight axis. I would ask for an interface drawing, continuous/peak/overload torque, speed and duty cycle, lost-motion definition and test torque, torsional stiffness, efficiency, temperature, lubricant and contamination constraints, target life/reliability/confidence, acceptance tests for every unit, qualification/environment tests for the design, provenance clauses, forecast, incumbent price and lead time. Public NASA verification guidance and a SAM.gov custom-gearbox solicitation both show why “a gearbox” is not a sufficient specification. NASA product verification SAM.gov custom gearbox solicitation

Interview by responsibility, not by fame: joint/motion engineering; reliability and test; manufacturing and supplier quality; programme qualification; field service and returns; procurement and finance. Ask for the smallest anonymised artefact they can share under NDA—a duty-cycle trace, acceptance template, supplier scorecard or returned-part teardown. Do not scrape or publish personal names.

The offer is a four-to-six-week paid requirements-and-evidence sprint. The partner gets a requirements baseline, FMEA, supplier process plan, metrology correlation, test matrix, costed design and a go/no-go review. The founder gets a fixed NRE fee, a representative spectrum or agreed proxy, an incumbent unit or benchmark and a named acceptance decision-maker. Data rights, witness rights, change notification, liability and no-exclusivity terms go in writing. A free sample with no load spectrum is a marketing prop.

ESA’s public EuroSMG record is a helpful precedent: it names a user specification, interfaces, an engineering model and a qualification model as distinct steps. That is the language to borrow—not the conclusion that any new reducer is qualified. ESA EuroSMG

16 · Own the evidence before the factory

COST-DOWN FOLLOWS THE BOTTLENECK

Stage 1 outsources blanking, heat treatment, tooth finishing and specialist metrology to audited suppliers while the founder owns drawings, CTQs, acceptance tests, genealogy and teardown. Stage 2 brings measurement, cleaning, assembly, grease control, final test and nonconformance data in-house. Stage 3 buys only the process whose demonstrated yield, lead time or IP exposure justifies it—often fixtures, finishing or metrology before a furnace. Keep two approved sources for every critical process and requalify after a material, recipe, tool or subcontractor change.

The cost model must include first-pass yield, scrap/rework, outsourced charges, inspection and test minutes, freight, financing, warranty reserve and customer audits. A lower nominal piece price is not cost-down if it increases escapes, inspection or qualification risk. China may be a supplier geography; it is not a universal competitiveness conclusion or legal moat. Provenance means the actual manufacturer, beneficial ownership, material origin, process location, subcontractors and change controls.

Here is the execution clock:

Day	Deliverable	Stop/go gate
30	Six role-based interviews, one-page packet, one named non-flight axis, two supplier calls, lab quote, hazard register and polymer demonstrator	Go only if a buyer shares a spectrum or defensible proxy and agrees to paid NRE
60	Signed NRE, incumbent benchmark, frozen interfaces, FMEA, test procedure, lab booking, supplier audit and cost/yield model	Go only if acceptance criteria are named and measurable with an uncertainty budget
90	Mule data, supplier-lot plan, guarded-fixture review, metrology correlation plan, pilot PO and witness dates	Go only if the main measurement risks close and the written forecast can clear 35% gross margin after scrap, test and warranty

Stop after two failed controlled pilot iterations; repeatable crack, wear, ratcheting or drift before the requirement; missing genealogy/change control; critical-dimension Cpk below 1.33 after process lock; learning yield below 70% or pre-scale yield below 90%; no certifying authority and acceptance matrix for aerospace; or no customer design-in by month 18. A short over-torque test can stop the project. It cannot create a 10,000-hour claim.