The RMA-repair pain is real but narrow: who else has it, and who only looks like they do
Landscape input for the founders. Executes the pressure-tested scan in the RMA target plan, grounded in the archetype from the removability-gate memo. Per RDI methodology and the assumption-ledger rule, this surfaces evidence, ranges, and disqualifiers — it does not conclude. Every non-trivial figure carries a source label; every load-bearing financial was checked against a primary SEC/IFRS filing, not a blog restatement. The compliance/export wedge is treated as killed; where a defense/government angle surfaced, it is flagged for a human, not reframed.
Outline changes from the approved plan
The plan’s two research tiers were executed as five parallel legs (electronic spine as one; the far-field analog scan split into aviation/defense, space + negative controls, medical/EV, and heavy-equipment/rail/oil-&-gas) so each industry got primary-filing depth rather than uniform shallowness. No sections were dropped; the negative controls and horizontality test are retained and, if anything, carry more weight than the plan anticipated. One method caveat the plan pre-registered — that warranty reserves are the wrong tell for industries that book repair economics in service-segment margins or core-exchange accounting — turned out to be decisive for the far-field and is treated as a first-class finding below.
1. BLUF
The NVIDIA-style RMA pain — trapped capital in a repair-vs-replace loop at high ASP, on a process that “worked when we were a chip company” and now doesn’t — is structurally common but acutely, unsolvedly painful in very few places. Fourteen industries were run through a three-layer screen (structural/removability → warranty-accrual-rate signal → emerging-vs-already-solved). The removability gate and the maturity gate together disqualify most of the field:
- Structural fit is everywhere. Rotable / line-replaceable-unit economics show up in datacenter hardware, telecom, robotics, drones, semicap, batteries, aviation, defense, rail, heavy equipment, medical imaging, EV powertrain, and oil & gas. The gate does real work only at the edges — it cleanly rejects the negative controls (satellites in orbit, installed wind/gas turbines, elevators) and splits system-vs-component in imaging, semicap, robotics, and BESS.
- The financial pain signal mostly disconfirms. Where a warranty rollforward exists in the far-field, the accrual rate is flat-to-declining, the inverse of NVIDIA’s climb: GE Aerospace, RTX (a single powder-metal artifact, amortizing down), Honeywell, Caterpillar, Deere, GE HealthCare, Cisco, Ciena, AeroVironment. Big warranty numbers are not the same as NVIDIA-shaped pain.
- Genuinely emerging, structurally-fit, unsolved pain concentrates hard. Only NVIDIA clears every layer with public evidence and two internal interviews. AMD is the one public “probable,” one cycle behind. Beyond that, the strongest structural fits are private and unmeasurable (humanoid robotics, enterprise drones) or ambiguous (Fluence in grid batteries; Wabtec in rail; semicap subsystem suppliers).
- Horizontality is unconfirmed, not disproven — and the difference is the whole game. Counting genuinely independent buying centers that clear all three layers with visible evidence today, the answer is ~1 confirmed (NVIDIA) + ~1 probable (AMD) + an unmeasurable emerging frontier (humanoids, drones, EV battery packs, semicap modules). But two measurement caveats keep this from being a verdict that the market is narrow: (a) the primary financial tell — a rising warranty-accrual rate — is structurally blind to the large share of the field that books repair economics in service-segment margins, core-exchange/LTSA accounting, or not at all (private firms), so many disqualifications are “not visible in warranty data,” not “no pain”; and (b) “already solved by incumbents” is doing heavy lifting at Layer 3, yet NVIDIA also had incumbents (Syncron, Servigistics, ReverseLogix) and the pain persisted because none delivered a unified solution — so a mature-industry disqualification is only as good as an operator’s confirmation that the incumbent tooling is actually adequate. Net: the confirmable market collapses toward NVIDIA + AMD; whether the actual market is that narrow is exactly what the operator interviews in §9 exist to test.
The rest of this brief walks the logic so the list can be trusted or challenged, then ends in the questions that would tighten it.
2. The archetype (the structural signature distilled from NVIDIA)
The reusable pain, stated as a wedge, is the “rotable-pool repair economy” from the removability-gate memo: B2B units expensive enough to repair and modular enough to pull, failing often enough that a standing balance sits idle in a central repair loop — trapped capital an identifiable owner would pay to unlock. Four gates, each of which excludes real candidates:
- Repairable, not disposable — repair cost sits well below replacement cost, so repairing beats scrapping.
- Removable, not installed — the unit detaches and ships to a depot/pool, rather than being serviced in place by a technician. This is the ceiling a value-only screen misses.
- Recurring, not rare — units fail and return often enough to keep a standing balance in the flow.
- Trapped capital of real magnitude, with an owner —
unit value × units-in-flow × dwellis large and sits on one identifiable party’s balance sheet.
NVIDIA’s distinguishing feature is not the structure (aviation and Caterpillar share it) but the maturity: the process is manual and unsolved because it scaled faster than the tooling — “this worked when we were a chip company” [Interview: Lonny Orona, 2026-05-12]. Its public tell is a warranty reserve that has climbed $82M → $306M → $1.29B → $2.81B (FY23–FY26), an accrual rate rising from ~0.46% to ~1.15% of revenue — verified this session directly against SEC XBRL (FY25 $1,290M; FY26 $2,807M ending balance, CIK 0001045810) [Public: NVIDIA FY26 10-K, accession 0001045810-26-000021]. A widely-cited $8.2B NVIDIA warranty figure is a WarrantyWeek data error (~3× the filed number) and is the reason every financial below was checked against primary filings [Public: WarrantyWeek, 2026-04-09, contradicted by primary].
3. The screening method (and its limits)
Applied in order, to every industry and firm:
- Layer 1 — structural fit (the qualifier): high unit value; rotable/LRU form factor (the removability gate: physically removed and shipped to a pool, not bolted-in and field-serviced); a real return stream with warranty/service obligation; contract-manufacturer dependence; growth outrunning process maturity.
- Layer 2 — pain-signal overlay (the prioritizer): product-warranty reserve growing YoY and outpacing revenue (accrual rate rising, not reserves merely tracking sales); plus rapid growth; plus qualitative tells (recalls, depot build-outs, reverse-logistics hiring, refurb/secondary-market activity). Limit, and it is a big one: many far-field industries book repair economics in service-segment margins, LTSAs, or core-exchange accounting, not a warranty line — so warranty absence there is not signal absence. Private fast-growth firms have no filings at all and are quarantined as speculative.
- Layer 3 — maturity check (the real differentiator): is the repair-loop pain emerging/unsolved (NVIDIA) or already-solved (mature reverse-logistics IT + entrenched incumbents)? Only emerging pain clears. Limit: “incumbents exist” is not the same as “pain solved.” NVIDIA had incumbents too (Syncron, Servigistics, ReverseLogix) and the pain was real because none stitched the flow into a unified system. A Layer-3 disqualification on maturity grounds is therefore provisional — it holds only if an operator confirms the incumbent tooling actually closes the loop, which for the far-field industries below we have not yet tested. Where a mature-industry disqualification would flip on that test, it is flagged.
Pre-registered disqualifiers (candidates shown as disqualified, not silently dropped): fails the removability gate; low unit value; warranty growth that is only revenue growth (flat/falling accrual rate); pain already solved by incumbents; a reserve spike tracing to a single consumer recall rather than a structural reman stream.
The three layers are AND-gated: a firm must pass structure, show a live signal, and have unsolved pain. Most of the field dies at Layer 3, not Layer 1.
4. Industry-by-industry walk
Deep-dive spine (electronic near-neighbors — most NVIDIA-like)
4.1 Datacenter compute / networking / storage — SPLIT: emerging only at the AI-accelerator layer
- Structural fit: passes industry-wide. GPUs (~$25–40K), switch line cards, all-flash arrays, configured servers are all pulled and RMA’d; heavy ODM/EMS build (Foxconn, Quanta, Wistron, Celestica, Flex, Jabil, Sanmina) creates the multi-party reconciliation that made the flow manual
[Public: EMS/ODM structure; Synthesis]. - Pain signal — quiet everywhere except the accelerator: whole-system integrator reserves are flat-to-declining despite the AI-server boom — Dell $467M→$450M (FY23→FY26), HPE $318M→$284M, Supermicro ~$17M on >$20B AI-server revenue (~0.2–0.3% rate)
[Public: OEM 10-Ks]. The load-bearing reason: GPU defect cost indemnifies back up to NVIDIA via supplier agreements, so structural fit at the integrator layer does not become warranty pain there[Synthesis; internal reverse-logistics brief]. - Emerging-vs-solved: networking (Cisco, Juniper) and storage (NetApp) are mature/solved — 20+ years of advance-replacement, stable books. AI accelerators (NVIDIA; AMD trailing) are the emerging exception. Incumbent reverse-logistics vendors already serve the category (Reconext, Ingram Micro Lifecycle; software Syncron, Servigistics/PTC, Baxter, ReverseLogix) — the gap the anchor describes is a unified platform, not the absence of point tools.
- Named firms: NVIDIA — confirmed-signal (emerging; rising accrual; two interviews). AMD — probable (reserve $85M→$188M→$308M, rate 0.56%→1.03%, one cycle behind — but the 10-K does not segment warranty to data-center, so AI attribution is inferential)
[Public: AMD FY25 10-K, accession 0000002488-26-000018]. Cisco — disqualified/mature ($333M→$399M FY23–FY25, rate0.6–0.7%, revenue flat)$25–27M flat; a stray “$146M” XBRL read was a mistag — the same class of error as the WarrantyWeek $8B figure, caught by reading the rollforward), Arista (fastest grower, +28.6% to ~$9.0B FY25, but no warranty rollforward disclosed — the one “watch, don’t count” item if AI back-end switching throws a failure tail), Pure/Dell/HPE/Supermicro — all disqualified[Public: Cisco FY25 10-K, accession 0000858877-25-000111]. Juniper ($29M→$31M, now inside HPE), NetApp ([Public: respective FY24–FY26 10-Ks].
4.2 Telecom / RAN equipment — DISQUALIFIED: signal runs the wrong way, in a capex down-cycle
- Structural fit: passes on form (an RRU is unbolted and depot-repaired) but weak on unit value — a single RRU/baseband unit is mid-4-to-5-figure, ~an order of magnitude below the GPU-server class; pain would scale through site count, not per-unit trapped capital
[Synthesis]. - Pain signal — inverted: Ciena (cleanest US-GAAP data) warranty flat-to-down ($45.5M→$55.5M FY22–FY25, provisions charged declining)
[Public: Ciena FY24/FY25 10-Ks]; Nokia modest; Ericsson’s warranty line not isolable from restructuring-dominated provisions. And global RAN equipment revenue fell$45B→$35B (2022→2024), flat 2025 — killing “growth outrunning process”[Public: Omdia/Dell'Oro]. - Emerging-vs-solved: strongly already-solved — decades of depot repair/spares pooling with a dense incumbent ecosystem (Tempest, DTC Telecom, TXO). Two disruptors remove themselves from the return stream: Mavenir is exiting RAN hardware and Rakuten Symphony pivoted to software licensing — the very players you’d expect to have immature reverse logistics are abandoning the hardware
[Public: Light Reading, 2025].
4.3 Robotics / humanoids / industrial automation — the strongest emerging structural candidate, but unmeasurable
- Structural fit: system-vs-component split is decisive. Installed robot arms/gantries and Symbotic structures fail the gate at the system level (field-serviced on-site); the LRUs inside pass cleanly (servo drives, motors, controllers, VFDs, PLC modules — third-party depots advertise “3–5 day turnaround, 2-year in-service warranty”). Humanoids/AMRs are more likely whole-unit-swappable — the strongest NVIDIA-shaped tell, with NVIDIA-shaped CM dependence (Apptronik + Jabil) and growth outrunning maturity (Unitree 5,500+ units shipped by mid-2026; Tesla Optimus / Figure / Apptronik shipping to pilots simultaneously)
[Public: RivCut/Sacra, 2025–2026]. - Pain signal: the one clean public filer, Symbotic, is ambiguous — reserve÷revenue rising 1.75%→1.94% (NVIDIA-directional) but provision÷revenue dead flat at 1.03%, so the reserve build is usage-lag, not accelerating accrual
[Public: Symbotic FY25 10-K]. ABB is large/flat/mature; Rockwell and Fanuc rollforwards were not primary-verified (open). - Emerging-vs-solved: established automation (ABB, Fanuc, Rockwell) is already-solved (deep GES-style component-repair depot ecosystem + Syncron/PTC/IFS software). Humanoids/AMRs are potentially greenfield — no depot SLA standard, no rotable-pool convention, no reman market yet, because the installed base is 1–3 years old and just hitting volume
[Synthesis]. - Named firms — speculative tier (private, quarantined): Figure AI, Apptronik (Jabil ODM; GXO/Mercedes pilots), Agility Robotics (RaaS bundles a service obligation), Unitree (highest volume → most likely a real emerging return stream), Tesla Optimus (captive/internal-first, may have no external RMA yet). All
[Speculation]— structural fit only, no financial confirmation.
4.4 Drones & UAS — BIFURCATED, and the biggest value+growth is disqualified by design
- Structural fit: passes cleanly at the enterprise/defense-ISR tier ($10K–$M, returnable), fails on value at the consumer tier (sub-$5K), and — distinctively — fails by design for attritable munitions/CCAs (Switchblade, Anduril Altius/Fury are meant to be consumed; the repair loop does not exist)
[Public: AeroVironment/Anduril product pages]. - Pain signal: the one clean filer disconfirms — AeroVironment warranty balance÷revenue is flat-to-declining (~77→51→44 bps FY24–FY26); the FY26 balance jump is smaller than the BlueHalo-merger revenue jump — “warranty growth is only revenue growth”
[Public: AVAV 10-K XBRL, 2026]. Red Cat has the best growth shape (+161%, Army SRR win) but warranty is unmeasurable and small. - Emerging-vs-solved: mature at the defense-depot end (Tobyhanna); at the enterprise end an incumbent may already own it — Robotic Skies (250+ service centers, 50 countries) + OEM captive pools (Skydio Care Enterprise, DJI Care Enterprise) push enterprise drones toward “watch,” not “wedge.” Skydio is the cleanest private structural match but discloses nothing
[Speculation].
4.5 Semiconductor capital equipment (module/subsystem level) — structurally on-thesis, financially unverified
- Structural fit: the tool fails the gate (a $150M+ scanner is field-serviced in the fab), but the module/subsystem layer passes — RF generators, mass-flow controllers, vacuum/turbo pumps, chucks, chambers, EUV optics are removable LRUs cycling through refurb/exchange pools, at a heavily-outsourced supplier layer, with AI/HBM/advanced-packaging capex driving growth
[Synthesis]. - Pain signal: the least-complete leg — flagged honestly. The subsystem-supplier warranty rollforwards (MKS Instruments, Advanced Energy, Ichor, Ultra Clean, Entegris) and the AMAT/Lam/KLA tool-level notes were not primary-verified this session; per the $8B precedent, no figures are asserted from memory.
- Emerging-vs-solved: likely substantially solved at the OEM level — Applied Global Services, Lam CSBG/Reliant, KLA services, plus used-tool brokers (SurplusGlobal) are the incumbent aftermarket. Any emerging pain would sit at the subsystem supplier’s own RMA process, which is the open question.
4.6 Energy storage / grid batteries — PROBABLE, on the single most ambiguous reserve in the scan
- Structural fit: the site/container fails (fixed installed asset), the battery module/rack and PCS/inverter pass (removable, swappable; mid-life module swaps are a growing practice). Heavy cell-supplier dependence (CATL, LGES, Samsung SDI); Fluence even books “recoverable warranty costs from suppliers” — direct pass-through evidence; and growth is real (Fluence $12.7B/6.8GW backlog)
[Public: Fluence FY24/FY25 10-Ks, CIK 1868941]. - Pain signal — the crux: two distinct things share the GAAP warranty line and separating them is the whole game — (1) a long-dated capacity/degradation-guarantee actuarial reserve discharged by augmentation, not a depot cycle, and (2) an active module-swap reman stream. Fluence’s reserve rose ~50% (to $40.2M gross / $27.5M net FY24) while revenue was flat-to-down → rate rising, but the degradation-vs-reman split is unresolvable from the filing
[Public: Fluence FY24 10-K, Note 15]. Enphase’s big reserve is 82% long-tail actuarial and benefited in FY24 on a low-value microinverter (fails Layer-1 value). Generac is a single-recall artifact (SnapRS). Tesla Energy is blended with autos. - Emerging-vs-solved: the battery-reman ecosystem exists but points the wrong way (EV-pack-into-stationary repurposing, not grid-BESS module reman); the grid-native module-reman depot layer looks thin/emerging — possible greenfield. Note: the one mature financial product here (Munich Re/TWAICE Li-ion performance-warranty insurance) attaches to the degradation tail, not the trapped-capital repair loop — a wrong-analog warning.
- Named firm: Fluence — probable (module/PCS LRU flow, explosive backlog, rising reserve on flat revenue), gated entirely on the degradation-vs-reman question.
Far-field analog scan (removability-gate fits — classified emerging vs mature)
4.7 Aviation & aerospace — DISQUALIFIED: the textbook decades-mature/served case
Passes Layer 1 cleanly (it is the origin of the rotable/LRU concept; engines $12–45M, LTSAs, independent MROs). But repair economics live in services-segment margin and LTSA/ASC-606 accounting, not warranty — GE services are 75% of commercial-engine revenue — and where warranty is visible it is flat: GE Aerospace reserve ~$595M (+1%), accruals −12%; Honeywell −7%; RTX’s headline $2.9B is a single powder-metal AD artifact amortizing down ($2.7B→$1.7B→$0.9B) — the inverse of a rising tail, and the pre-registered “single-artifact” disqualifier [Public: GE 10-K FY25 accession 0000405450-26-000008; RTX 10-K FY24 accession 0000101829-25-000005; WarrantyWeek ww20260611]. Layer 3 is emphatically solved: ILS (since 1979), AMOS/IFS/Ramco/TRAX/Rusada, Aeroxchange, AerSale/AAR/StandardAero — a ~$8B mature MRO-software market. (Boeing’s reserve +31%/accruals +127% is real but a single-driver 737 MAX/787 quality story, not a rotable-reman signal — out of scope.)
4.8 Defense electronics & platforms — emerging pain exists, but in the wrong (government) buying center
Passes Layer 1 (LRUs/WRAs, PBL contracts, Derco “rotable pool & exchange”). The rotable mechanics are mature/served (IFS A&D, L3Harris depots, organic DoD depots). But the F-35 evidence is genuinely NVIDIA-shaped and unsolved: 141-day repair turnaround vs 60–90 target, 10,000+ unit depot backlog, and “DoD does not have records for all the parts it purchased, where they are, or what they cost” — with DoD declining a PBL contract citing data-quality issues [Public: GAO-24-106703, GAO-26-108113]. The surprise: the closest analog to NVIDIA’s actual pain in the entire far-field is the government’s depot enterprise — which the killed-compliance boundary says to flag for a human, not pursue commercially. Named firms (RTX, Lockheed, Northrop, L3Harris, BAE) either disqualify on flat warranty or sit behind a government buyer.
4.9 Space / satellites — DISQUALIFIED at flight-hardware, with one surprising pass and one financialization tell
The removability gate does its sharpest work here: a failed satellite is on-orbit-spared, deorbited, or graveyard-orbited — never RMA’d (Iridium runs on-orbit spares; LEO burns up; GEO boosts to disposal). Even at extreme ASP, form makes removal impossible — the gas-turbine failure mode at altitude. The surprise that passes: SpaceX Falcon-9 booster / Merlin-engine refurbishment is a genuine jet-engine-style rotable-refurb loop (recover-inspect-refurbish-refly) — but it is captive and n≈1, so it passes removability and fails horizontality (a pool of one) [Public: reuse reporting 2024–26; economics blog-sourced, LOW confidence]. Consumer Starlink terminals pass on form, fail on value. Orthogonal finding: satellite build contracts carry orbital-performance-incentive / warranty-payback clauses — the industry’s financial substitute for a return loop, relevant to the parametric-insurance wedge, not this one [Public: Maxar 10-K FY2020].
4.10 Heavy equipment / off-highway — DISQUALIFIED: fit is perfect, the tell is present and inverted
Caterpillar, Komatsu, and Deere pass the gate cleanly via explicit core-exchange programs. Contrary to the plan’s premise that they bury everything in core accounting, Cat and Deere both book full warranty rollforwards — and the rate is flat-to-declining (CAT 1.43%→~1.13%, ending liability fell $1,894M→$1,626M; Deere ~2.2% flat), the inverse of NVIDIA [Public: CAT FY22/FY24/FY25 10-Ks; Deere FY24 10-K accession 0001558370-24-016169]. Cat Reman is the picture-perfect fit that is exactly the mature/served trap (rising physical core returns 127M→147M lbs run through a decades-old, Syncron-topped, deposit-accounted system). Komatsu’s JP filing was not verified — the one open corner.
4.11 Rail — DISQUALIFIED overall, but Wabtec is the leg’s one genuine rising-tell
All three pass Layer 1 (unit-exchange of locomotive components; Alstom services 35,500+ vehicles including non-Alstom fleets). Wabtec is the surprise going the NVIDIA direction: warranty expense÷sales 0.94%→0.98%→1.13% (FY22–FY24), outpacing revenue [Public: Wabtec FY24 10-K accession 0001628280-25-005100] — but an order of magnitude milder than NVIDIA, and Wabtec self-serves its reverse-logistics IT (Wabtec Digital/ExpressYard, ex-GE Transportation, 30+ years). Alstom/Siemens Mobility book no warranty rollforward (IFRS; repair economics in Services revenue), and Siemens Mobility’s aftermarket is a top profit driver — “solved and monetized,” the opposite of distress.
4.12 Medical imaging & devices — DISQUALIFIED: passes the gate, but actively de-provisioning
The system (MRI magnet/gantry) fails the gate (field-serviced); the component-exchange layer passes (RF coils, X-ray tubes, boards cycle through ISO-certified exchange pools, next-day). But the anchor’s tell points the wrong way: GE HealthCare’s warranty reserve and absolute provisions both fell ($193M→$192M→$168M; provisions $238M→$216M→$202M) on flat-to-up revenue — a cleaner negative than mere flatness, primary-verified [Public: GE HealthCare FY24 10-K, accession 0001932393-25-000005]. Layer 3 is emphatically solved: OEM refurb (GE GoldSeal, Siemens ecoline, Philips Diamond Select) + a deep ISO ecosystem (MXR, DirectMed, Block Imaging, Agiliti, TRIMEDX) in a ~$21B-by-2030 maintenance market.
4.13 EV powertrain & charging — EMERGING industry, but the fit and the financials sit on different layers
The most genuinely emerging far-field industry — and the most internally contradictory. Battery packs and inverters pass the gate (removable, high-value) and second-life/pack reverse logistics is explicitly immature (“logistics a major cost,” SoH-data gaps, EU battery-passport from 2027). Charger cabinets fail the gate (field-serviced); swappable power-modules pass. But the clean rising-accrual numbers sit at the whole-vehicle level, which fails the gate — Lucid 24% / Rivian 4.9% accrual vs mature OEMs ~2% [Public: WarrantyWeek EV, 2024-09-19, secondary — flagged for primary verification]. The two in-scope charging names both fail the specific tell: ChargePoint’s warranty spiked once and reversed ($16.7M FY24 → $6.0M → $5.5M — a disqualifier, not a ramp) [Public: CHPT FY26 10-K accession 0001777393-26-000013]; EVgo’s pain is own-asset uptime/impairment (the ReNew program), structurally different from a customer-return repair loop. Net: EV’s emerging pain (packs) and EV’s best financials (whole vehicles) are attached to different layers of the stack — a real contradiction to sit with.
4.14 Oil & gas downhole / rental tools — DISQUALIFIED: no tell, and the rotable half is contracting
SLB, Halliburton, Baker Hughes pass on the downhole/OFSE side (tools pulled, inspected, redeployed; M&E is 76–90% of gross PP&E). But none books a product-warranty rollforward (SLB books $14.5B of product revenue and still discloses no warranty line), the repair loops are mature and in-house (SLB Delfi/Lumen, HAL DecisionSpace; pure-play DTI runs proprietary COMPASS fleet-tracking over a “70-year” commercial norm), and the rotable half is the contracting half (SLB Well Construction −11%, HAL −3 yrs, BKR OFSE −8%). NVIDIA’s “growth outrunning process” is inverted [Public: SLB/HAL/BKR FY25 10-Ks].
Negative controls (must fail the gate — and do, each for a different reason)
- Installed wind turbines — nacelle components serviced up-tower/crane-down (removal-cost barrier); a minority gearbox-exchange sub-flow exists, which proves the gate measures removal cost, not dollar value. Wind’s real signal is parametric downtime, a different wedge. (Vestas, GE Vernova, Siemens Gamesa.)
- Installed gas turbines — bolted into the plant, serviced via on-site outages (GE Vernova “Live Outage”); the canonical gate failure. Real capital pain exists (order backlogs) but it is supply/capacity, not repair-loop. (GE Vernova, Siemens Energy, Mitsubishi Power.)
- Elevators / escalators — serviced in place under recurring maintenance contracts (Otis maintains >2M units; ~75% of EBIT from servicing the installed base); a route-density service business, no rotable pool. (Otis, KONE, Schindler, TK Elevator.)
5. Horizontality test — count the independent buying centers
Counting genuinely independent organizations that clear all three layers plus the disqualifiers, not industries or logos:
| Tier | Count | Who |
|---|---|---|
| Confirmed (all layers, public evidence + interviews) | 1 | NVIDIA |
| Probable (structural fit + partial/ambiguous signal, public) | ~2–3 | AMD; Fluence (gated on degradation-vs-reman); Wabtec (mild, self-served IT) |
| Speculative frontier (structural fit, no measurable signal) | unmeasurable | humanoid robotics (Figure, Apptronik, Agility, Unitree, Optimus); enterprise drones (Skydio — possibly already served); semicap subsystem suppliers (MKSI/AEIS/ICHR/UCTT/ENTG); EV battery-pack/inverter makers |
The honest reading: the confirmable archetype does not fan out into a broad horizontal market today — it collapses toward NVIDIA + AMD + a thin, unmeasurable emerging frontier. But “confirmable” is load-bearing, and the collapse is driven as much by the limits of the screen as by the market. Two structural blind spots inflate the disqualified column: the accrual-rate tell cannot see service-margin, core-exchange, or private-firm economics (so much of the far-field and the whole private frontier are “invisible,” not “painless”); and the Layer-3 maturity disqualifications are provisional — “incumbents exist” only closes the case if those incumbents deliver a unified solution, which NVIDIA’s own incumbents did not, and which we have not tested for aviation, medical, heavy-equipment, or rail. So the finding cuts three ways: (a) a “horizontal RMA platform” thesis is not yet supported by visible evidence; (b) the strongest structural fits are precisely the pre-revenue private firms whose numbers we cannot see — where NVIDIA itself sat before we got inside, so absence of signal in a 1–3-year-old installed base is weak evidence of absence; and (c) several mature-industry disqualifications would re-open if an operator revealed the incumbent tooling is point-solutions rather than a closed loop. The number to carry forward is not “the market is 1–2 buyers” but “only 1–2 buyers are confirmable without more conversations” — which is a research instruction, not a verdict.
6. Evidence-tiered target map (landscape, not a ranking)
Confirmed-signal — NVIDIA.
Probable — AMD (rising accrual, DC-attribution inferential); Fluence (module/PCS reman if the reserve is reman not degradation); Wabtec (rising rate, self-served).
Speculative (private/fast-growth, structural fit, no visible numbers) — Figure AI, Apptronik, Agility Robotics, Unitree, Tesla Optimus (humanoids); Skydio and enterprise DJI (drones); MKS Instruments, Advanced Energy, Ichor, Ultra Clean, Entegris (semicap subsystem — public but warranty unverified); BorgWarner / EV inverter & pack makers.
Watch (public but signal immaterial/ambiguous) — Arista (AI back-end switching failure tail?), Symbotic (reserve rate up, provision flat), Red Cat, Rivian/Lucid (whole-vehicle rising accrual, fails gate — verify primary filings).
Disqualified (shown, with reason) — Cisco/Juniper/NetApp/Pure/Dell/HPE/Supermicro (mature/quiet; GPU cost indemnifies upstream); Ciena/Ericsson/Nokia/Mavenir/Rakuten (down-cycle, mature, or exiting hardware); GE Aerospace/RTX/Honeywell/Safran (flat or single-artifact); Lockheed/Northrop/L3Harris/BAE (flat, or government buyer); satellites-in-orbit / SpaceX-captive (gate fail / n=1); Caterpillar/Komatsu/Deere (inverted tell, mature); Alstom/Siemens Mobility (no tell, monetized aftermarket); GE HealthCare/Siemens Healthineers/Philips (declining, solved); ChargePoint/EVgo/Enphase/Generac/Stem/Tesla Energy (spike-reversed / wrong-shape / value-fail / recall-artifact / blended); SLB/HAL/BKR (no tell, contracting); wind/gas turbines, elevators (negative controls).
7. Assumption ledger
| # | Assumption | Value | Source | Conf | Impact if wrong |
|---|---|---|---|---|---|
| 1 ⭐ | GPU defect cost indemnifies upstream to NVIDIA (why the whole integrator layer is quiet) | qualitative | [Synthesis; internal reverse-logistics brief] | M-H | Highest — load-bearing for why datacenter pain concentrates at the accelerator, not the OEMs |
| 2 ⭐ | Horizontality collapses to ~1 confirmed + ~2 probable + unmeasurable frontier | see §5 | [Synthesis across all legs] | M | Highest — this is the headline; if the private frontier is real, the count could grow materially |
| 3 ⭐ | Fluence reserve is reman pain vs degradation-guarantee actuarial | 50% YoY, $40.2M FY24 | [Public: Fluence FY24 10-K] | L | High — the entire battery vertical hinges on this split |
| 4 | NVIDIA benchmark accrual trajectory (the yardstick) | $82M→$2.81B FY23–26; ~0.46%→1.15% | [Public: SEC XBRL, FY25/FY26 confirmed] | H | High — the comparison collapses if the yardstick is wrong (survived the $8B scare) |
| 5 | Far-field warranty rates flat/declining, not rising | CAT 1.4%→1.1%; GE Aero +1%; GE HC declining; RTX artifact | [Public: primary 10-Ks] | H | High — the core disconfirming pattern for the far-field |
| 6 | Humanoids/drones structurally fit but are unmeasurable (private) | — | [Speculation] | M | High — determines whether the frontier is real or wishful |
| 7 | Semicap subsystem-supplier warranty | UNVERIFIED | OPEN | L/gap | High — least-complete industry; re-run needed |
| 8 | EV startup accrual rates (Lucid 24%, Rivian 4.9%) | secondary, whole-vehicle | [Public: WarrantyWeek 2024] | M | High — drives the “EV emerging” read; not primary-verified, and the vehicle fails the gate |
⭐ = the three most load-bearing to validate next: (1) the upstream-indemnity mechanism, (2) the horizontality count, (3) the Fluence reman-vs-degradation split.
8. Surprises & contradictions (the RDI payload)
- Warranty magnitude is anti-correlated with NVIDIA-shaped pain. Cisco (~$399M) carries a bigger reserve than HPE or nearly Dell, yet is the textbook solved case; RTX’s $2.9B is a declining single-artifact. A “biggest reserve = biggest pain” screen ranks exactly backwards. Rate-of-change and maturity are the signal, not size.
- Structural fit is strongest exactly where the financial signal is invisible. The most NVIDIA-shaped candidates (humanoids, enterprise drones) are private with zero disclosure; the visible-number firms mostly disconfirm. Evidence and confidence are inversely correlated across the frontier.
- The closest analog to NVIDIA’s actual pain is the US government’s F-35 depot enterprise — bad parts data, 141-day turnarounds, a PBL contract abandoned over data quality — which the killed-compliance boundary tells us to flag, not pursue.
- Two “disruptor” archetypes engineer the return stream out. Mavenir/Rakuten exit hardware; attritable munitions (Anduril Fury, Switchblade) are consumed by design. High value + high growth ≠ a repair loop.
- A space sub-layer unexpectedly passes the gate — SpaceX booster refurb — but is a pool of one, so removability and horizontality point in opposite directions.
- The one mature financial product in batteries solves the wrong bucket (Munich Re/TWAICE insures the degradation tail, not trapped repair capital) — a wrong-analog warning for anyone reaching for the insurance precedent.
- A live XBRL mistag surfaced in real time (NetApp “$146M” vs the ~$27M rollforward) — the same class of error as the WarrantyWeek $8B figure, caught only by reading the footnote. The primary-filing discipline earned its keep.
9. Gaps, missing perspectives, and questions to test next
The dominant gap: we have zero interviews outside the NVIDIA arc. Every judgment on 13 of 14 industries is public-filing + synthesis, with no operator who lives the problem. The RDI method says the target list is a set of conversations to have, not a conclusion.
Questions, with who could tighten each:
- Is the humanoid/AMR frontier real yet, or 12–24 months out? Unitree’s volume (5,500+) is the first place to look. Needs an operator conversation — our nearest live robotics relationship is Eoin (Mundane Robotics), though his pain was supply-chain intelligence, not repair.
- Does the archetype actually clear more than 1–2 buying centers? The horizontality count is the load-bearing strategic question — validate by getting inside one speculative-tier firm the way we got inside NVIDIA.
- Fluence: how much of the reserve is active module reman vs degradation-guarantee actuarial? A Fluence/Wärtsilä/Powin services controller could answer; it decides the whole battery vertical.
- Semicap subsystem suppliers — re-run the primary financials (MKSI/AEIS/ICHR/UCTT/ENTG) and ask an AMAT AGS or Lam CSBG services lead whether emerging RMA pain sits below the OEM exchange programs.
- EV: is EVgo’s “keep my own assets up” a different (uptime/parametric) wedge than the customer-return repair loop? And do Rivian/Lucid’s rising accruals decompose to powertrain/battery (rotable) vs body (not)? Verify Rivian/Lucid/Tesla FY24 10-Ks directly.
- Is the F-35 depot data problem a defense/government wedge the founders want to touch at all — adjacent to the killed-compliance boundary? A founder call, not an agent call. Nearest contact: Chris Moran (Lockheed Ventures), via H.R. McMaster.
- Arista — is AI back-end switching (Etherlink/800G) throwing a field-failure tail below disclosure materiality? The one datacenter “watch.” No networking-ops interview in the vault.
This brief surfaces the sized evidence and the load-bearing assumptions; it does not draw the conclusion. Sources: 2026-07-11-repair-flow-analog-removability-gate; reverse-logistics-warranty-tam-2026-05-29; rma-reverse-logistics-warranty-software-landscape-2026-07-08; 2026-06-23-nvidia-repair-flow-economic-model; scrap-pile-rma-target-plan; NVIDIA arc interviews (Lonny Orona 2026-05-12; Alex Zhu 2026-05-27). Primary filings cited inline (NVIDIA FY26; AMD FY25; Cisco FY25; Fluence FY24; Symbotic FY25; AeroVironment; GE Aerospace FY25; RTX FY24; GE HealthCare FY24; Caterpillar FY22/24/25; Deere FY24; Wabtec FY24; ChargePoint FY26; SLB/HAL/BKR FY25; GAO F-35 reports; Maxar FY2020). External market/trade sources labeled inline.