Seventeen companies in the humanoid-robot supply chain, examined with one question: is a Micron/Nvidia-type earnings explosion coming that the market hasn't priced? The robot buildout is real — but at today's prices, almost none of these suppliers are the way to own it.
Research: July 13–14, 2026 · Prices as of July 13 · Automated research pipeline (sourced filings/teardowns/capacity data → adversarial review → deterministic math), summarized by hand. Corrected July 14 after an independent verification pass. Not investment advice.
A supplier only produces a Micron-style earnings explosion when four things line up: a demand step you can compute in units; supply that can't respond — not just the company being sold out, but the whole qualified supplier list, including the cheapest ones; a pricing structure that lets shortage turn into higher prices; and earnings forecasts that are wrong by a multiple. That last point matters more than the P/E: Nvidia traded at 25–30× forward earnings in late 2022 and still rose ten-fold — not because it was cheap, but because its earnings came in several times above what anyone modelled. So the question is never "is this expensive," it is "what earnings does the price require, and can the physics deliver them?" Two failure patterns dominate this basket: China built component capacity ahead of the robots (so shortages resolve by orders shifting to a cheaper supplier, not by prices rising), and the market already paid for the robot future.
We called this "the cleanest avoid" because its factory runs well below capacity. That was backwards — idle capacity you can fill without spending is the precondition for an earnings explosion, not evidence against one. So we re-ran it with the bull case built first. The bull won on the physics and lost on the arithmetic, and the arithmetic is one sentence: fill every machine it owns, sell those units at the price humanoid robots actually pay, and you land almost exactly on what analysts already forecast.
Everything the bull said was true. Harmonic Drive really does have four to five times more factory than it uses — it makes roughly half a million reducers a year on machines built for 2.6 million — and its capital spending runs below its depreciation, meaning it can fill that capacity without spending a yen. It earns about ¥65 of profit on every extra ¥100 of revenue. And uniquely among all seventeen companies in this study, it actually raised prices during tightness (a stated ¥500M contribution to last year's margin). Every Chinese competitor's file reads "no price rise found." On the physics, it is the only company here shaped like Nvidia.
And then the arithmetic ends it. Its entire nameplate — 2.64 million reducers — sold at the price the humanoid market actually clears (about ¥26,600, where Chinese suppliers set it) comes to roughly ¥70bn of revenue. Add its motion-control business and you get about ¥86bn. Analysts already forecast ¥83.9bn. So: sell out completely, every machine running, zero new investment — and you have booked 3% more revenue than consensus, on a worse profit mix. To beat the forecast, it needs a higher price; but a higher price is exactly what loses it the volume to the Chinese suppliers who are half its cost. Volume needs price; price kills volume. That is the whole company in one line.
And the demand isn't reaching it anyway. Tesla's screws and reducers go to Chinese suppliers; Figure builds in-house and buys Chinese (teardowns find zero Harmonic Drive part numbers); Apptronik designed harmonic reducers out, with its engineer saying on record: "you don't get the price break you think you'll get with harmonic drives." Its finance chief explicitly told analysts not to extrapolate its humanoid order book. Meanwhile the widely-cited "8.7 million vs 5 million reducer shortage" requires more than 500,000 humanoids in 2027 — roughly ten times the mainstream forecast, and it double-counts Harmonic Drive's own output on the supply side.
The number that settles it: even if the bull is completely right and 2028 earnings come in at double what analysts expect, the stock still returns −33% at a normal exit multiple. It is not that the earnings can't grow. It is that the price already pays for the good outcome twice over — and the earnings are bounded by a cost curve, not by a factory.
The one company with a genuine physical shortage — running above its rated capacity and still unable to fill humanoid orders — and still the most extreme valuation in the basket. To justify its price it would have to become roughly 19× its current size. This is the most instructive name here: sold out, and it still doesn't work.
The shortage is real: monthly output has run at 131–150% of official nameplate, and its next big factory is delayed to 2028 (the precision grinders it needs are export-controlled and back-ordered). That is exactly the setup we hunt for.
Why it still fails: being sold out means nothing if you cannot raise prices. Despite running flat out, Leaderdrive's average selling price fell 15% last year, because competitor Laifual profitably sells at less than half its price and has spare capacity. Now run the bull at maximum strength — full post-2028 capacity of 1.64M reducers plus a complete shift into higher-value integrated modules at ¥6,101 each. That entire revenue envelope is about ¥3.06B. To be a 20× stock it needs about ¥11B. It falls short by 3.6× with every lever pulled at once.
Correction: an earlier version used a 1.4M-unit capacity figure. That was our misreading of the project's ¥1.402 billion capital budget as a unit count. Real nameplate is 640,000. The error made the case look better than it is.
~250× trailing earnings; a consensus its own quarter already falsifies (revenue fell 19% while estimates assume +29%); zero finished roller-screw capacity today; negative gross margins on the robot parts it does sell; a formal exchange reprimand for misleading humanoid disclosures; and absence from the official Optimus-supplier list its bulls cite.
We first framed this as a pricing problem ("needs 11× the price"). Review sharpened it: it is a financing-and-machinery problem that does not resolve before 2029. Installed finished-screw capacity today is literally zero; the first line arrives February 2027. The celebrated 980,000-unit expansion needs roughly 336 specialised inner-thread grinders it has not ordered — some ¥3–6B of additional capital against ¥809M of short-term debt, ¥34M of free cash flow, and a self-identified ¥1.41B funding hole. Its realistic 2027 output ceiling is 6,000–12,000 sets.
What it actually sells today is ¥200 semi-finished blanks, tier-2, at a negative 9.3% gross margin — not the ¥500 finished screws in the plan. In fairness to the bull: management's own study projects ¥14.7B of revenue at full ramp, which would clear the bar. The question is not whether the plan is attractive; it is whether this balance sheet can finance and build it before the window closes.
Its roller-screw capacity is not small. It is zero. The 150,000-unit line the story rests on makes ball screws — a different product — and even that runs at 2.3% utilization. Actual roller-screw output: about 250 sample sets, and ¥223,000 of unaudited sample revenue, against a ¥10.7B market value.
We originally said Best "needs a 4× price increase." Review showed that was far too kind: it credited a roller-screw thesis with a ball-screw line's nameplate, and priced it at an industry aspiration (¥1,500) rather than anything Best actually charges.
The honest picture: zero installed roller-screw lines; ~250 effective sample sets a year; a 400,000-unit expansion that exists on paper against an unaudited plan for 58 imported grinders, only about a quarter of which appears funded. Grant the entire 2028 build-out at the aspirational price and it still covers under a quarter of the valuation gap; at the price the market actually clears (¥499), under a tenth. Meanwhile the tech press names other suppliers for Optimus screws, Unitree's IPO supplier list omits it, and Figure's robot is all-rotary — it uses no screws at all.
The inverse of the setup we hunt: unit demand is growing, but Ouster's factory is only ~25–33% utilized, its average selling price has fallen from $6.3k toward $5.1k, and Chinese rivals sell capable lidar for $200–1,000. Growth without pricing power.
Ouster makes solid 3D lidar with real customers — warehouse robots, traffic intersections, mining trucks, defence drones. The problem is market structure: Hesai and RoboSense ship millions of units a year from China at list prices as low as $200–1,000, while Ouster's blended price is above $5,000 and drifting down every quarter. Its contract manufacturer's line could build 100k+ units a year; Ouster shipped 25k.
It needs roughly $286M of annual revenue to stop burning cash; consensus for this year is about $220M. So the math requires much faster volume (usually bought with price cuts in this structure) or prices holding (which three years of data contradict). Robots getting cheap lidar is great for robot makers — it is the sensor supplier being commoditized.
Filed roller-screw output last year: zero qualified units. It has denied the Tesla rumour twice through investor relations. Roughly 45% of its market value (~¥64B) is a humanoid option premium the evidence values at ¥1–2B. But the residual is a genuine hydraulics business — so this is "the robot story is worthless," not "the stock is worthless."
On the screws the kill is emphatic and got stronger on review. Official status is "sample delivery plus initial mass production." Filed qualified output: zero. Design capacity is 750 units a year; realistic current output under 300 sets; roughly 10–20 of a planned 25 precision grinders installed, and the ramp cannot inflect before mid-2027.
Where our first version was unfair: we asked what would make Hengli a 20× stock, then blamed roller screws for the answer. But the market pays 35–45× for Hengli because it is a hydraulics compounder (its pump-valve segment grew 21% at a 48.8% gross margin), not because of robots. The honest, narrower statement: the ~¥64B robot premium is unsupported and should not be paid. Whether the remaining ~40× hydraulics cyclical is worth owning is a different question — one this study did not answer.
Still the only name here we would keep on a list — but not for the reason we first gave. We said its capacity made the numbers work at today's screw prices. That rested on a 1,000,000-unit expansion the company cancelled in July 2025. Real capacity is 100,000 units, with commercial output still unproven.
The correction. On real capacity, Shuanglin would need roughly ¥3,096 per screw to justify its price — over six times its own ¥499 list price. By the standard we applied to every other name, that is a fail. Our published "works at today's price" was an artifact of a dead number.
Why it stays on the watchlist anyway. The reason is simpler: it barely needs the screws. You pay about 22× earnings for a real automotive business whose core profit is compounding (+37% last year), after a 55% share-price fall. The gap to a "cheap" valuation is only about ¥310M of extra profit — six percent of revenue — and it has three ways to earn that without heroic screw pricing: a built 120,000-unit joint-module line selling at ¥1,499 (3× the bare screw), an in-house subsidiary now selling the precision grinders that are the industry's real bottleneck, and a new robot ball-screw line landing this autumn.
But be clear what it is not. Even at full expanded capacity and today's prices, screws add a fraction of its revenue base — its earnings cannot be wrong by a multiple. It is a reasonably priced auto supplier with a cheap option attached, not a Micron. Triggers: audited half-year results (~August) showing real screw revenue and proof the new line produces, or a named design win.
Robots need to feel force; VPG's strain gages are the precision standard with a contractually protected niche; the company has only 13.3M shares, so small wins move earnings violently; and there is a dated go/no-go event this year. But it already trades near 37× 2028 consensus — you are paying up for the option, so size it like one.
A strain gage is a foil element that measures force by how much a metal part flexes — it is how a robot hand knows how hard it is gripping. VPG has shipped preproduction force sensors (~$600k so far, ~$5M expected this year) to unnamed humanoid maker(s), at roughly $400–500 of content per robot. The production decision is expected in the second half of this fiscal year — a rare, dated, binary catalyst.
The upside is violent because the share count is tiny: at ~650k humanoids in 2028 and ~20% of the sensing content, that is roughly +$1.74 of EPS on a $3.07 consensus — a +57% surprise. And unlike the reducer and screw markets, this niche has real protection: VPG's former parent is contractually barred from competing, and cheap alternatives are not precise enough for the calibration-locked sockets it serves.
Three reasons it is an option, not a conviction buy: the stock already re-rated on its AI test-equipment boom; VPG historically monetizes through volume and mix, never price increases; and Minebea Mitsumi mass-produces two million cheaper force sensors a month — if robot makers decide "good enough" beats "perfect," the option expires worthless. Playbook: small starter, add only when the production decision lands, exit on a slip or a competitor win.
How this was made: each company got a ~130–170-node automated research sweep over filings, teardowns, capacity and pricing data (every claim source-tagged), adversarial review by independent models, and a deterministic supply/demand model computing implied earnings against consensus. On July 14 a clean-context verifier attacked the results and found seventeen input errors — most of which, at that point, favoured the bull. Both directions of error are now documented above.
What this is not: investment advice, a complete analysis, or an audit. Numbers marked as estimates come from research corpora, not company guidance. Prices move; these verdicts are conditional on July 13, 2026 prices. Avoid calls are value statements, not momentum calls — names at 100×+ re-rate on narrative and can look wrong for a quarter. Nothing here is a short recommendation. Do your own work.