Humanoid supply chain

The closest thing to a monopoly on the list. Strain wave gears are the compact, zero-backlash gearboxes inside precision joints, and there is no substitute at volume. The ~85% share claim is consistent with how the market is generally described, and the humanoid order book is real but still small: roughly ¥2.5bn of humanoid-related orders guided for FY3/26, which management has suggested could double or triple in FY3/27.

The catch is that today's profit does not come from humanoids. FY3/26 landed at ¥59.6bn revenue (+7.0%) with operating profit down 94.4% year on year. The recovery is now visible, though: April–June revenue of ¥16.68bn (+23.6%) and operating profit of ¥1.84bn against ¥122m a year earlier, on consolidated orders of ¥24.13bn (+55.7%) — roughly ¥3bn of which was a one-off North American surgical-robot order. On 7 August the company lifted FY3/27 guidance to ¥74.5bn revenue and ¥8.5bn operating profit, from ¥68bn and ¥6.2bn. The stock is still priced on that forecast rather than on trailing numbers.

WatchQuarterly humanoid order disclosure. This is the single number that justifies the multiple. Eight sessions of it now, and the shape is a staircase with one broken step: +5.5% on 4 September, +4.2% on 7 September, −4.5% on 8 September as the yen firmed, +2.1% on 9 September, +2.7% on 10 September to ¥5,980 — the highest close since 17 August — then −3.2% on 11 September with the Nikkei down 1.93%, +0.7% on 14 September against an index down 0.81%, and +0.3% to ¥5,850 on 15 September on a flat tape. Net over the eight sessions it is 7.7% higher, and the last two of them came while Tokyo sold its AI complex and then sat still — neither of which is a gearbox event. The month is still 10.1% lower and the drawdown 36.1%. Nothing in that sequence came from the company: no order disclosure has been published since the 7 August guidance upgrade, which is already priced.

The heavy-duty counterpart: cycloidal RV reducers for medium-to-large joints, with roughly 60% global share. Of the six, this is the one where the cited financials check out cleanly — FY2025 delivered ¥307.9bn revenue (+9.8%) and ¥20.7bn operating profit (+60.3%), with FY2026 guided to ¥327bn and ¥27.7bn.

Crucially, Nabtesco is committing capital ahead of the demand: it is doubling RV-reducer capacity into 2026 and bought Slovak cycloidal maker Spinea in 2023. That is a company acting on the thesis, not just narrating it.

WatchUtilisation of the doubled capacity. Idle new capacity turns a margin story into a fixed-cost problem quickly. It has been the most volatile name on the tier, and has now gone quiet: −5.7% on 2 September, then +1.6%, +1.4% and +3.6%, then −6.1% on 8 September and +4.5% on 9 September — and since then −0.7%, −1.3%, −0.3% and +0.2% to ¥4,437 on 15 September — four sessions inside 1.5% after a fortnight of violent ones, and net over the whole ten-session run it is 1.0% higher than where it started. It sits 27.4% below its 52-week high, the tightest on the tier apart from Nidec, and 9.4% lower over a month.

The most concrete commercial traction on the list. Schaeffler signed a strategic partnership with Swiss Hexagon Robotics in April 2026 for strain wave and planetary gear actuators, and a separate technology and supply partnership with UK-based Humanoid. It is also a customer: it plans to deploy at least a thousand Hexagon robots across its own plants within seven years.

The in-house manufacturing point is fair — motors, power electronics and encoders are built internally. The offset is that Schaeffler remains predominantly an automotive supplier, so a humanoid win is diluted by a large, cyclical, structurally pressured base business. That mixed identity shows in the tape: down 16.9% YTD despite being up 28.5% over twelve months, and 42.0% below its high. The fortnight of going nowhere ended on Friday. Four sessions had sat inside nine cents — €7.34 on 7 September, €7.43 on 8 September, €7.40 on 9 September and €7.34 on 10 September — and then 11 September broke the range, −2.0% to €7.19, the lowest close since 2 September. Monday went further: −3.3% to €6.95, the lowest close since 24 March — nearly six months undone in two sessions, and the largest single fall on the tier that day apart from Ambarella’s. It never joined the mid-August drawdown, so for a fortnight it had nothing to give back; it is giving it back quickly now, and the month reads 6.0% lower where a week earlier this was the one tier-one name essentially flat on that measure. Note what it is not: a chip name. Schaeffler fell on the same session as the semiconductor selloff but for its own reasons, and at 0.60 to the robotics basket it is one of the two loosest fits on the tier, alongside Ambarella at 0.54 — the two names here that are really something else. Both of those closes were reconstructed from the closing auction print when the German daily feed swallowed the session. The 11 September bar has since filled at exactly €7.19, confirming the method; the 14 September bar has not yet appeared — see the method note.

One correction worth flagging, because it changes the read. Measured on the thin SFFLY ADR, Schaeffler's correlation to the robotics basket looks like 0.16 — essentially unrelated. Measured on the liquid Xetra line it is 0.60. The ADR was not tracking the company; it was tracking its own illiquidity. Schaeffler belongs firmly inside the robotics bloc — and at 0.46 it is also the tier-one name that moves most with REMX, which fits a supplier whose bill of materials is heavy in magnets and bearing steel.

WatchWhether actuators are ever broken out as a reported segment. Until then the robotics revenue is invisible inside an auto-parts P&L.

The world's largest electric motor maker, building integrated humanoid motion units that fold motor, gearbox and controller together — FLEXWAVE strain-wave reducers, KINEX cycloidal gearboxes and planetary systems shown at Automate 2026. The scale argument is real: roughly ¥2.6 trillion in annual revenue (about $17bn, not $2.6tn as the post's garbled phrasing implies).

This is the one name carrying a governance problem rather than just an execution problem, and on 4 September 2026 one half of it was finally published. Nidec disclosed the Investigation Committee's report on the quality strand, received on 2 September and released once confidentiality review was complete, together with its recurrence-prevention measures. The committee, established in May 2026, found a series of quality-related failures: 4M change-rule violations, falsification and fabrication of testing and inspection results, work outside approved inspection and manufacturing conditions, the use and shipment of substandard and unauthorised items, and breaches of record preparation and management rules, arising from an “unhealthy state” of excessive cost-reduction pressure, inefficient risk monitoring and inadequate quality-assurance resourcing. The decisive line for holders is financial: Nidec's evaluation of the individual matters found no material impact on its consolidated financial statements. Remediation gives quality managers authority to suspend shipments, revises reporting lines, adds permanent complaint boxes, puts every executive and employee through quality-compliance training and starts continuous quality audits; disciplinary action is promised but nobody is named. The stock rose 5.7% on the day and a further 7.7% on the Monday after it, to ¥2,812, before losing 3.2% to ¥2,722 and a further 4.8% to ¥2,591 — giving the whole post-report rally back by Wednesday — then taking most of it straight back with +5.6% to ¥2,737 on 10 September, the largest single gain anywhere on this page, and easing 1.7% to ¥2,691 on 11 September and a further 1.3% to ¥2,655 on 14 September. Five of the seven sessions since the report moved more than 3%, in both directions — +5.7%, +7.7%, −3.2%, −4.8%, +5.6%, −1.7%, −1.3% — and the two small falls that end the run are the quietest pair since the report landed. Nidec has disclosed nothing at all since 4 September. This is a stock being repriced on interpretation rather than on information. Everything else is still open — though it is worth being precise about what "open" means, because this page previously was not. The separate third-party committee on the accounting irregularities has finished: appointed in September 2025, it delivered a mid-term report on 27 February 2026 and its final report on 17 April 2026. What outlasts the inquiry is its consequence. Nidec is still correcting accounts back to the year ended March 2022, has not closed FY3/26, and has repeatedly postponed results — including the April–June quarter, delayed again on 5 Aug 2026. A tariff strand runs alongside: additional customs duties from declaration errors were estimated in March 2026 at US$69.7m, about ¥11.14bn, including interest. The quality strand covers changes to materials, processes and designs made without customer approval, mishandled inspection data and mislabelled production locations, and on 5 August an individual shareholder formally requested that the company pursue liability against current and former directors. A second, larger request followed: on 20 August multiple shareholders jointly demanded the same action under Article 847 of the Companies Act, disclosed the next day. Nidec has an Executive Responsibility Investigation Committee, running since 13 March 2026, to decide whether directors, auditors and executive officers bear legal liability. The US ADR has barely traded since February, which is why the Tokyo line is the only usable quote.

Watch30 September 2026, and it is now fifteen days out. The quality strand is closed; the accounting strand is not, and it has a hard date on it. Nidec obtained approval on 30 June to extend the filing deadline for its annual securities report for the year ended March 2026 from 30 June to 30 September — so the restatement the company has been working through since September 2025 is either filed inside fifteen days or the extension itself becomes the story. The April–June quarter behind it is still missing too. Until both are current, position sizing here remains a governance decision rather than a robotics one. On the closed strand, the verified count is worth carrying: the 4 September report found 844 cases of quality misconduct over about a decade — mostly changes to materials and processes without customer approval, alongside falsified tests and country-of-origin mislabelling — and Nidec states no individual matter has a material impact on its financial statements. Press reporting that attaches a ¥250bn write-down to that report is conflating it with the accounting strand; the claim audit now carries this. Note too that Nidec still shows as the tightest tier-one name against its 52-week high, now at 9.6%, and positive over twelve months at +1.6% — neither of which describes a stock that has recovered. That second figure is being drained by its own base rather than by the price: Nidec rose 4% on 12 September 2025, and that session rolling into the comparison cut the twelve-month return from +7.5% to +2.1% on a 1.3% price move, with two further quiet falls since taking it to +1.6%. Tomorrow the base becomes 16 September 2025’s ¥2,671 — so at an unchanged price the twelve-month number goes negative, to about −1.1%, with nothing about the company having changed. It is the cleanest illustration on this page of a return figure that moves because the calendar did.

One of very few firms that can already supply precision motion at industrial scale, now pointed at humanoids: it acquired 100% of Tokyo Robotics, whose whole-body torque sensing and impedance control fill a genuine capability gap, and ships MOTOMAN Next with NVIDIA GPUs for autonomous control.

But the profit claim in the source post is backwards. For the fiscal year ended February 2026, revenue edged up while operating profit fell more than 5%, with pre-tax and net profit down harder. The growth is in the forecast — FY2/27 guidance of ¥580bn revenue (+7%) and ¥47bn net profit (+33%). Up 41.8% over twelve months but 45.7% below its high — still the widest gap on the list — it is the clearest example here of a stock priced on guidance rather than results. Its year-to-date return turned negative on 1 September and has kept going, to −9.7%: a 4.4% fall on 8 September wiped out three modest up sessions, a 1.6% loss on 10 September took it below its April low, a further 3.0% on 11 September took it to ¥4,301, then two yen lower on 14 September and three yen lower again to ¥4,296 on 15 September. Both of those are fresh closing lows, the weakest since 3 April, and the manner of them is the point: not a capitulation but a slow bleed of two and three yen a session while the rest of the tier drifts higher. It is 20.9% lower in a month, the steepest on the tier, and the only tier-one name making new lows rather than working off the August peak. The guidance it is priced on does not get tested until the first-half numbers.

WatchFirst-half progress against that +33% net profit guide. It is a forecast-driven multiple with a recent record of missing.

The only US-listed name and the only one with a clean, current, verifiable set of numbers. CV7 launched at CES on 5 January 2026 — an 8K edge AI vision SoC for multi-sensor perception across robotics, drones, industrial automation and automotive. Fiscal 2026 revenue was $390.7m, up 37.2%, with fiscal 2027 growth guided at 10–15%. Q2 FY2027, reported on 3 September, came in at $108.1m (+13.2%) with non-GAAP EPS of $0.18 — both a shade ahead of consensus — and Q3 guided to $115–124m.

The robotics pipeline is disclosed rather than implied: 15+ robotics design wins with lifetime revenue above $100m, and 30+ customers in the pipeline. The honest caveat is that robotics is still the smaller part of a business anchored in security cameras and automotive — and the deceleration from +37% to a guided 10–15% is the number bulls have to explain.

Nor did the follow-up. Fermi Wang appeared at Citi's Global TMT Conference on 9 September and was blunter there than on the call: robotics has “a couple design wins,” but the volume is not going to change the financial performance materially, and humanoids are, on his account, harder from a technology standpoint than a level-5 autonomous car. He also named the constraint that is not in the guidance — memory. Customers, he said, could not get allocation commitments from their memory suppliers even for November, which he called the biggest uncertainty and expected to run into 2027. For a fabless designer that is a customer-side supply problem, not a demand one, but it lands on the same revenue line.

The Q2 call moved the narrative without moving the disclosure. Fermi Wang said edge AI revenue “reached record levels” on balanced sequential growth in Auto and IoT and very strong growth from the 5nm CV75 and CV72 SoCs, and the company raised its five-year serviceable-market forecast for edge AI and physical AI, citing higher-value products and new distribution through Macnica and CapGemini. Margins went the other way: non-GAAP gross margin of 59.3% against 60.5% a year earlier, with the Q3 guide at 59–60%. Still a GAAP loss — $6.7m, or $0.15 a share — against $8.2m of non-GAAP net income, on $272.3m of cash.

WatchRobotics revenue disclosed as its own line — and Q2 did not do it. “Physical AI” now appears in the serviceable-market framing, but there is still no robotics number to hold anyone to, so this name remains a narrative read rather than a volume read. Then the tape did something the disclosure did not explain. Ambarella rose 8.4% to $68.80 on 9 September, the largest single move on this page in weeks and its highest close since 31 August. The sequence is worth recording, because the obvious attribution is wrong: the stock opened flat at $63.07 against a $63.48 previous close, ran roughly 10% in the first thirty minutes to $70.28 by 10:00, then faded to $66.43 by 13:00 — before the Citi appearance began at 13:55 — and closed at $68.80 on heavy final-hour volume. The move was made in the opening half-hour, hours ahead of the only scheduled event of the day, and the stock actually fell through the appearance itself. It was not a sector move either: the semiconductor ETFs added 0.7% and 0.1%, and NXP, still the reported suitor, closed lower. No company announcement accompanied it — and a sixth of it came back the next session, −1.4% to $67.87 on 10 September, against a semiconductor index down 2.7%. On 11 September it did nothing at all, a single cent lower at $67.86, while the semiconductor index rose 1.8% — the one name on this page that joined neither the rare-earth selling nor the chip rebound. Then on 14 September the whole unexplained 9 September gain came back at once, −5.7% to $63.98, its worst session since 16 July and within fifty cents of the 8 September close it started from. This time the attribution is easy and it is not about robots: public calls from several AI-lab executives for a slower pace of frontier development took the semiconductor complex down with them — SOXX −5.6%, Nvidia −3.4%, Broadcom −4.8%, AMD −4.4% — and Ambarella tracked the sector almost exactly. That is the useful fact about this position. On the way up, its 9 September move was idiosyncratic and unexplained; on the way down it was pure sector beta. It leaves Ambarella 22.0% lower over a month and 33.8% off its high, and it is the only tier-one name whose drawdown is being set by the AI trade rather than by industrial demand. Ambarella is still the name closest to breaking robotics out, which would make it the cleanest read on real humanoid volume anywhere on this list — the next chance is Q3, in December.

The only vertically integrated Western play, and the one genuine bridge between "rare earths" and "robot actuators." Q2 2026 revenue of $108.5m (+89%) came with NdPr production of 840 tonnes (+41%) and sales of 1,006 tonnes — more than double the prior year — pushing adjusted EBITDA positive and narrowing the diluted loss to $0.11 from $0.19.

The government backstop is unusually deep: the Department of Defense took $400m of preferred stock in July 2025 to become the largest shareholder, with a ten-year $110/kg NdPr price floor, and has since signed a ten-year offtake for 100% of the output of the planned 10X facility in Northlake, Texas — a $1.25bn-plus campus commissioning from 2028 that is meant to lift MP's total US magnet capacity to roughly 10,000 tonnes a year, not to produce 10,000 tonnes on its own. Beijing noticed: China's Ministry of Commerce added MP to its export-control blacklist in June 2026, and the stock fell nearly 30% through July.

The 13 Aug earnings call surfaced a new line: a multi-year, nine-figure agreement signed in July to supply separated gadolinium oxide to an undisclosed US aerospace and defense manufacturer, recovered as a co-product of the existing Mountain Pass ore stream. Terms are undisclosed so it cannot be sized precisely, but it is the clearest evidence yet that MP is building toward a multi-element refinery rather than a single-product (NdPr) business.

WatchMagnetics segment revenue. At $16.5m against $95.6m from materials it is still the small half of the business — but it is the half that matters for anything robotic, and the only number here that scales with Western magnet independence. The tape went the other way meanwhile: a 5.5% fall to $51.32 on 10 September inside a broad metals liquidation, then a further 1.6% to $50.51 on 11 September against a rising market and a rising gold complex, and then it stopped — +0.3% to $50.68 on 14 September, one of only two tier-two names to hold a session that took REMX to a six-week low. That leaves MP flat to a rounding error year to date and 49.5% below its high — a company whose magnetics revenue nearly doubled over the year has given the whole year back in the price. One column to read carefully this run: MP shows 13.7% lower over a month, against 6.7% a run earlier, and the stock rose in between. The 30-day lookback base rolled off 12 August onto 14 August, a session on which MP closed 8.6% higher, and that base change is most of the move. The one-month figure describes the calendar here more than it describes the tape.

The source post has this one materially wrong. It is not waiting on 2028 — USAR commissioned its Phase 1a commercial magnet line at Stillwater, Oklahoma and began filling customer orders for sintered NdFeB magnets in Q2 2026, pulling its timeline forward by about two years. Phase 1a is guided to a 600 mtpa run-rate by the end of Q4 2026 and 1,200 mtpa combined with Phase 1b in Q1 2027, against roughly 5,000 tonnes at full build.

Q2 2026 revenue was $5.8m against a $10.3m net loss (down from $142.5m a year earlier, when the company had no commercial revenue at all). The balance sheet is the actual headline: $1.53bn of cash at quarter-end, up from $359.9m at year-end 2025. USAR is deploying it into corporate structure as much as capacity — it closed the acquisition of Texas Mineral Resources Corp on 7 Aug, giving it outright ownership of Round Top, and finalized a roughly 13.6% stake in French rare-earth processor Carester. It agreed to buy Brazil's Serra Verde, the only scaled ex-China producer of all four magnet rare earths, for about $2.8bn; shareholders approved the share issuance on 28 August — 108,248,297 votes for against 1,403,269 opposed, with 16,879,393 abstentions — and the combination completed on 3 September 2026, announced the following morning and filed the same day on Form 8-K under Item 2.01. This page carried it as pending for a week longer than it was; that was wrong, and the correction is the more important half of this card.

What actually closed is bigger than the price tag suggested. USAR paid $300m in cash plus 126,849,307 new shares to Serra Verde's shareholders, lenders, warrant holders, employees and consultants, and took on Serra Verde's financing with it: a facility of up to $565m from the US International Development Finance Corporation, sized to fund the Brazilian optimisation and expansion programme through to positive cash flow. The Goiás operation has been producing since January 2024 and is finishing a commissioning and optimisation programme whose first stage targets a run-rate of about 4,000 tpa of total rare earth oxide by the end of 2026. Sir Mick Davis joins the board; Thras Moraitis becomes President immediately and succeeds Barbara Humpton as chief executive on 1 October. USAR now owns an operating heavy rare-earth mine, a magnet line and a separation stake — upstream, midstream and downstream in one company, which is what MP has and no one else outside China does.

The financing behind that deal landed on 24 August 2026 and is the most consequential development on this page since the last refresh. The special-purpose vehicle that will buy 100% of Serra Verde's Phase 1 output completed an upsized $1.55bn capitalisation: a $750m Department of War investment, raised from the $500m originally contemplated, a bank senior secured borrowing-base revolver of up to $500m, and a US government forward purchase contract for not less than $300m of rare-earth product over five years. The offtake runs fifteen years on take-or-pay terms with guaranteed price floors on the magnetic rare earths — including, on the company's account, the first such floors anywhere for dysprosium and terbium. Read against the $110/kg NdPr floor MP already has, Washington is now underwriting the price of the heavy elements too, which is the part of the chain China actually controls.

And on 9 September 2026 it broke ground on a second magnet plant, which this page should have carried a run earlier. The site is Blacksburg, South Carolina — 124 acres in Bailey Industrial Park, Cherokee County, about 800,000 square feet, roughly $1.2bn of investment and some 490 manufacturing jobs, commissioning in 2028. The capacity is the part that matters: 6,400 tpa of sintered NdFeB magnets plus 5,000 tpa of strip-cast metal and alloy. On its own that is larger than Stillwater's full build of roughly 5,000 tonnes, and it places a second and bigger magnet line behind the company's own metal. Measured against MP's 10X campus, which is meant to lift MP's total US magnet capacity to about 10,000 tpa from 2028, Blacksburg is the largest single increment to announced ex-China magnet tonnage anywhere on this page. The company says it screened nearly 275 sites before choosing it. None of this moved the stock, which fell 8.8% across the two sessions that followed.

The Round Top deposit in West Texas is the part that is genuinely years out, and the distinction matters: the magnets shipping today are made from purchased feedstock, not from USAR's own mine. Round Top did produce a first dysprosium oxide sample at 99.1% purity in January. Like MP, USAR was added to China's export-control list in June 2026. The stock sits 64.3% below its 52-week high, still the widest drawdown of the robotics-adjacent names, and closing the deal did nothing to lift it. It fell 5% on the day the SPV funding was announced, 6.5% on the day of the vote, 7.4% on 31 August when the result was disclosed and 3.1% on 1 September; it then drifted through six sessions either side of completion, fell 3.4% to $17.06 on 9 September, a further 6.0% to $16.04 on 10 September — the largest fall on the page that day, though that one arrived with the whole metals complex rather than alone — and 3.0% to $15.56 on 11 September, which did not: the complex rebounded that session and USAR made a fresh low regardless, its weakest close since 31 July. The one-way tape then broke, mildly — +1.0% to $15.71 on 14 September, one of only two tier-two names to hold a session that took REMX to a six-week low. Six weeks of decisive corporate progress have still been met with a tape that has given back 21.4% over a month. The market is pricing the dilution, which is arithmetic — 126.8m new shares against 244.7m outstanding — ahead of the assets, which require execution.

WatchTwo run-rates now, both due by the end of this year: Stillwater's 600 mtpa of sintered magnets and Serra Verde's 4,000 tpa of rare earth oxide. Every financing gate has been cleared — the vote, the $1.55bn SPV, the DFC facility — so from here the only thing left to miss is production, and the first management test is a leadership handover three weeks after the merger closed. The tape has read each corporate step as dilution arriving rather than risk lifting; the fourth-quarter numbers are where that gets settled either way.

Operationally the strongest of the five, and now with a full year on the board. FY26 results, reported 26 August 2026, were records across the line: revenue of A$977.9m, up 76%, net profit after tax of A$222.4m against A$8.0m the year before, REO production of 13,089 tonnes (+25%), NdPr production of 7,260 tonnes (+11%) and a record average realised price of A$80.7/kg. More strategically, it became the first commercial producer outside China of separated dysprosium and terbium — first terbium oxide in June 2026, on a heavy rare-earth circuit of about 1,500 tonnes a year.

The market did not take it as a clean beat. The ADR fell 5.9% on results day before recovering 5.0% the next session, then gave back another 8% into 1 September with the rest of the mining complex, and rose in four sessions to $11.23 on 8 September. That run ended abruptly: −2.9% to $10.91 on 9 September, then −4.3% to $10.44 on 10 September in the broad metals selloff, −2.2% to $10.21 on 11 September after it reversed, and −3.2% to $9.88 on 14 September — its lowest close since 3 August, and back under $10 for the first time since then, with nothing from the company behind any of the four. Over a month the ADR is 15.6% lower against REMX's 12.2% — still the steeper of the two, which is roughly what a 0.64 correlation and a higher beta produce together. The reason is on the cost side: capital expenditure on the heavy rare-earth circuit has escalated 63%, to A$294m from A$180m, which the company attributes to sourcing equipment outside China. That is the price of the scarcity premium, stated in cash.

That heavy-element capability is the real scarcity. Terbium oxide outside China has traded at $4,500–$4,700/kg against under $1,000 domestically, and dysprosium and terbium are what keep magnets stable at motor operating temperatures. The post's claim that Lynas is "now building a US processing facility" is out of date: on 16 March 2026 the US government redirected the US$96m earmarked for the Seadrift, Texas heavy rare-earth plant into a four-year purchase agreement for Lynas product instead, with a US$110/kg NdPr floor. The Texas plant is in limbo, not under construction.

WatchWhether Seadrift is revived or quietly shelved, and the dysprosium/terbium ramp against that 63% capex escalation. The heavy circuit is the part China cannot easily replace and the part no one else outside China has — but USA Rare Earth's new US-backed Dy/Tb price floors are the first sign that someone else intends to be paid for it too. A permanent CEO is also still outstanding: Pol Le Roux has run the company on an interim basis since 1 July.

A well-financed critical-minerals project that has almost nothing to do with robotics. Perpetua secured a $2.9bn EXIM loan, board-approved in May 2026 and expected to become available in the second half of the year, on top of a $255m private placement in October 2025 backed by Agnico Eagle and JPMorgan. Construction at Stibnite in Idaho is underway.

Two caveats the post skips. First, this is predominantly a gold mine: reserves of roughly 4.8 million ounces and about 450,000 ounces of gold a year over the first four years, with antimony supplying an estimated 35% of US demand alongside it. Second, production does not begin until 2029. Antimony goes into flame retardants, ammunition and alloys — not robot joints. It earns its place on a critical-minerals list, not on a robotics one.

The tape adds a third caveat, and ten sessions have now sharpened it in both directions. Perpetua shows −10.7% over the past month: the early-August run — $18.41 on 31 July to $24.72 by 7 August — had already rolled into the lookback base in full, so the month starts from the top of that run. What the month shows is still a gold trade rather than a critical-minerals one, and a badly lagging one: bullion is 0.7% lower over those four weeks while the gold miners are 4.6% higher, and Perpetua, at 10.7% lower, sits well below both. On 31 August gold rose and Perpetua fell 4.7% anyway, alongside the gold miners; on 1 September gold fell 1.9%, the miners 3.9%, and Perpetua 5.8%. Then it ran in reverse: on 2 September the miners rose 3.1% and Perpetua 4.1%. On 3 September the relationship held in direction but not in size — gold 1.9%, the miners 4.0%, and Perpetua only 2.4%, for once lagging the equities it usually amplifies. Then on 4 September it decoupled altogether: gold fell 1.4% and the miners 2.2%, and Perpetua did nothing at all, +0.2% to $25.09. On 8 September the usual relationship resumed quietly: gold eased 0.8% and Perpetua 1.3% to $24.76. On 9 September it went missing again: bullion added 0.9% and the gold miners 1.1%, and Perpetua closed unchanged at $24.76, to the cent. Then on 10 September it came back with force — bullion −1.7%, the gold miners −3.5% and Perpetua −4.4% to $23.67. On 11 September it went quiet once more: bullion rose 0.6% and the miners 1.1%, and Perpetua eased 0.2% to $23.63, its lowest close since 1 September and the only tier-two name that barely moved on a day the rest of the tier fell. Then on 14 September it amplified the complex again and harder than either leg of it: bullion −1.3%, the gold miners −3.1%, and Perpetua −4.6% to $22.55, its lowest close since 4 August. Three times in ten sessions it has ignored a moving metal and a moving miners index, and twice it has multiplied one. That session carries a caution worth recording, because this page names GOLD as the instrument to check before attributing a Perpetua move. GOLD is Barrick, and Barrick rose 11.2% that day against a falling metal and a falling miners index. A single-name divergence that wide is company-specific, not a gold signal — which is why the comparison above is drawn against the miners index rather than against Barrick alone. Across all ten sessions Perpetua tracked the mining equities, not the metal. The exposure is to gold equities, not to gold. Antimony went nowhere throughout: $51.80/kg, down 5.85% year to date and about 43% from a year earlier. The commodity that earns this name its place on a critical-minerals list is flat to falling either way.

WatchFinancial close on the EXIM facility, still guided to the second half of 2026 as of the 17 August quarterly and not yet done. Approved is not drawn — and with first metal four years out, financing terms matter more than commodity prices here. Cash at 30 June was $574.2m.

The weakest name on either list, and it deserves a plain warning rather than a bullet point. Novonix makes synthetic graphite for battery anodes in Chattanooga, Tennessee — a battery input, not a magnet input, and not part of any robot's bill of materials in the way the tier-one components are.

The financials are severe, and on 31 August 2026 the auditor said so in the accounts. The half-year report for the six months to 30 June shows a loss of $24.6m against $20.1m a year earlier, net operating cash outflows of $23.1m, and cash down to $59.5m from $79.9m at 31 December. Net current liabilities stand at $2.5m where six months earlier there were net current assets of $5.6m. The independent auditor's review report carries a material uncertainty that may cast significant doubt on the group's ability to continue as a going concern. The same report has the company backing away from a land purchase beside its Chattanooga plant and stretching the Riverside ramp-up, which is what conserving cash looks like in a capital project. The genuine positive is still the US$103m of 48C tax credits certified in April 2026 for the Riverside project — but those vest only when the first 11,000 tpa of capacity is in service before 7 April 2028, and a slower ramp is the one thing that deadline cannot absorb. Mass production for lead customer Panasonic is not expected before 2027.

The going-concern note names the trigger, which is what makes this more than a slow burn. $35m of face value in unsecured convertible debentures issued to Yorkville carries an A$0.12 floor on the ASX share price: a fall below it may constitute an amortisation event, under which the company could be required to redeem part or all of the outstanding balance in cash. On 31 August, the day that note was published, the ordinary shares closed at A$0.11, down 12.0% on 12.3m shares against a recent average nearer 3.5m, and the Nasdaq line fell 12.3% the same session. They have not recovered: A$0.11 on 1 September, A$0.105 on 2 and 3 September, A$0.11 on 4 September, A$0.105 on 7 September, then −9.5% to A$0.095 on 8 September — the first close below ten cents — A$0.098 on 9 September, a 7.1% bounce to A$0.105 on 10 September on 10.8m shares against a recent average nearer 3.5m, and A$0.105 again on 11 September — ten consecutive closes below the floor, the lowest of them 20.8% under it. Then on 14 September the shares rose 14.3% to A$0.12 on 7.0m shares, their first close back at the threshold since 28 August, with no ASX announcement attached to it. At the floor is not above it: the clause turns on a fall below A$0.12, so this close leaves the question open rather than answering it, and it leaves no margin at all. The directors rest the going-concern basis on a history of raising capital, including $14.5m from a June 2026 placement; the share purchase plan that followed it raised A$0.96m.

The US listing is a separate problem, already solved once. On 12 August the company agreed with BNY to change its ADS ratio from one ADS per four ordinary shares to one per forty — economically a one-for-ten reverse split of the ADSs, effective 27 August 2026, done to lift the traded price back over Nasdaq's minimum bid threshold. The underlying ASX ordinary shares are untouched; only the US wrapper changed. On the new ratio the US line fell to $2.86 on 3 September, rebounded 5.2% to $3.01 on 4 September, fell 8.3% to $2.76 after Labor Day, eased 0.7% to $2.74 on 9 September — a two-year closing low — then rose 6.2% to $2.91 on 10 September, the only US line on this page to rise that session and a day behind the Australian bounce it was following, added a further 2.4% to $2.98 on 11 September, and then jumped 8.4% to $3.23 on 14 September — its highest close since 28 August and the largest single gain anywhere on this page that session. Four consecutive rises, on a day the rest of tier two fell, which says more about how far it had already fallen than about anything improving; no company announcement accompanied any of them. The two lines are now disagreeing across a one-session offset: the ASX ordinary rose 14.3% to A$0.12 on 14 September and gave back 8.3% to A$0.11 on 15 September, so the Australian bounce the US line was following has already half-reversed at home. It leaves the US line 91.6% below its high and 74.4% down over a year. A ratio change buys compliance, not solvency: in the eleven sessions since it took effect the US line has gone from $3.51 to $3.23, still clear of Nasdaq’s $1 threshold but well inside the range where it matters.

WatchWhether the A$0.12 floor is breached on a basis that triggers the Yorkville amortisation clause. A cash redemption against a $59.5m balance falling by roughly $23m a half is the live question. The shares closed below the floor on ten consecutive sessions and then, on 14 September, exactly at it — a 14.3% single-session recovery with nothing disclosed behind it, on the largest volume of any up day in the run. Dilution or a raise was already the base case; the auditor has now written the reason into the accounts. It is the only name on this list where solvency is the question rather than growth.

The most substantial magnet business you can actually buy. TDK is the largest non-Chinese neodymium magnet manufacturer by revenue and the most globally diversified, producing sintered NdFeB in Japan and China alongside ferrite and metal-powder products. The company's origins are literally magnetic — it was founded in 1935 to commercialise ferrite.

Two honest caveats. Magnets are a modest slice of the group: TDK is dominated by its energy business, including ATL, one of the world's largest lithium battery makers, plus passive components and HDD heads. And its 0.54 correlation to the robotics basket — the highest of any name outside tier one, above every robotics ETF tested — largely reflects its status as a large Japanese electronics exporter moving with the same market, not humanoid demand reaching its P&L. You would be buying a battery-and-components group with a major magnet business attached.

WatchAny segment disclosure sizing magnets against the battery business. Without it, the magnet exposure is real but unmeasurable — the same problem Aichi Steel has, at much larger scale. The 0.54 correlation keeps earning its keep: TDK fell 6.9% on 8 September — the worst session of any name on this page — in a yen-driven selloff that had nothing to do with magnets, recovered 1.4% and a further 1.4% to ¥2,850.5 by 10 September, gave back 1.5% on 11 September as oil and bond yields took the whole Tokyo market down, added 0.3% on 14 September while the index fell 0.81% on its AI complex, and rose a further 0.6% to ¥2,833.5 on 15 September. It moves with Murata and Taiyo Yuden rather than with the gearbox makers, which is the point: this is a Japanese electronics exporter that happens to make magnets. Note what it did not do on 14 September: the AI-pace selloff that took the US semiconductor complex down more than 5% that session left TDK higher on both sides of it. A passive-components maker is not an AI-silicon name, and the tape agrees even when the sector label does not. Down 11.7% over a month, with a twelve-month return of +43.1% that has swung between +33% and +51% inside a fortnight, purely on which day the lookback base lands.

The cleanest direct way to own a magnet maker on this list. A Toyota Group company engineering permanent magnets tuned for robotic actuators, with an explicit design goal of reducing dependence on constrained rare earths — the same strategic bet Proterial and Niron are making, but in a listed vehicle.

It is also the steadiest chart in this report: up 9.9% over a year and 10.4% below its 52-week high, against 28–46% for five of the six tier-one names. That is partly a virtue and partly a warning: magnets are a modest slice of a diversified steelmaker, so the humanoid narrative barely moves it in either direction, up or down. It sat out the 4 September rally with a 0.8% fall, sat out Monday's 2.1% index rally with a 0.2% fall, and lost 1.8% on 8 September on a day the component names lost 3–7%. What it had not sat out was the drift: four consecutive falls, ending −0.8% to ¥3,115 on 10 September, on a day Harmonic Drive rose 2.7% and Nidec 5.6%. Then it did the rarest thing on the page — on 11 September it closed unchanged at ¥3,115, to the yen, while the Nikkei fell 1.93% and every other Tokyo name here lost ground. On 14 September it added 0.3% to ¥3,125, its first rise since 3 September, on a day the index fell again, and on 15 September it added 0.8% to ¥3,150 — its best session since 3 September and the largest gain of any Tokyo name on this page that day. Steadiness cuts both ways, and across the three sessions since 10 September it is the whole story: the Nikkei 2.7% lower in total, and a magnet maker 1.1% higher. It is now the tightest name to its 52-week high anywhere on this page bar Nidec.

WatchAny disclosure sizing the magnet business separately from steel. Without it you are buying a steelmaker with a magnet option attached, and the option is not priced because it is not visible.

A real magnet business you cannot buy directly. Arnold makes permanent magnets, precision thin metals and magnetic assemblies to the tolerances robotics OEMs need — and in March 2026 signed a distribution agreement with USA Rare Earth, linking it straight to tier two. But it has been a subsidiary of Compass Diversified since a $130.5m acquisition in 2012, and CODI is the only way in.

That is a poor trade for magnet exposure, and the reason is not subtle. CODI is a diversified holding company whose stock fell 59% in May 2025 when a probe found pervasive accounting fraud at its Lugano Diamonds unit. It restated fiscal 2022–2024, deconsolidated Lugano after a Chapter 11 filing in November 2025, triggered credit-facility defaults requiring forbearance, and suspended its distributions. The stock is up 135.6% year to date because it is recovering from that hole, not because of magnets — and it fell 2.7% on 9 September and a further 2.3% to $10.96 on 10 September, its lowest close since 6 August, in a session that had nothing to do with either magnets or diamonds. It has since taken most of that back, +0.5% on 11 September and +2.7% to $11.31 on 14 September, the largest tier-three gain that session and one of the few green closes anywhere on this page. Nothing about Arnold changed in either direction across those four sessions, which is the recurring problem with owning a magnet business this way.

The Q2 2026 call, held 17 Aug, is the freshest read on Arnold specifically. Subsidiary adjusted EBITDA came in at $91.5m (+12.6% year over year), helped by the 1 May sale of the Sterno food-service business, whose proceeds (over $280m) went straight to paying down the senior secured term loan — total debt fell to $1.59bn from $1.89bn at year-end 2025. COO Zach Sawtelle, who succeeds retiring CEO Elias Sabo at the end of 2026, called Arnold's segment "a standout, up nearly 50%," with a strong backlog tied to non-China rare-earth magnet sourcing and progress at a Thailand facility. Management said it is still pursuing further divestitures to cut debt, but did not name Arnold as a candidate — if anything, a segment growing 50% is the one they would want to keep. CODI shares still fell 4.8% on the day.

WatchWhether CODI divests Arnold. Management's 17 Aug commentary points the other way for now — Arnold was singled out as a growth standout, not flagged among the divestiture candidates — but a holding company still repairing its balance sheet post-Lugano remains a seller in waiting.

Not investable, but the most strategically interesting name in this section — because it is trying to make the bottleneck irrelevant rather than relieve it. Niron's iron nitride permanent magnets, out of University of Minnesota research, use no rare earths at all. If they work at scale, the entire NdPr supply argument underpinning tier two weakens.

It just got a serious endorsement: on 7 August 2026 the Department of War's Office of Strategic Capital issued a conditional commitment for a direct loan of up to $150m over 20 years for a 287,000 sq ft plant in Sartell, Minnesota — intended as the world's first full-scale iron nitride production site, targeted at robots, defence systems and electric motors. The same release names a second $150m loan from the Shakopee Mdewakanton Sioux Community for the same site, so the plant carries up to $300m of committed debt against roughly $50m of venture capital raised to date.

Construction reached a visible milestone on 10 September 2026: Niron raised first steel at Sartell and, in the same release, disclosed an equity investment from Honda, made through Honda Xcelerator Ventures on undisclosed terms. That is the part worth marking. Every dollar behind this company so far has come from venture funds or from government — the Department of War loan, the tribal loan. An automaker putting its own balance sheet into rare-earth-free magnets is a customer-side endorsement, and customers are the constituency whose behaviour would actually reprice tier two.

The number that matters for tier two is the capacity. Sartell is scoped at up to 1,500 tonnes a year of rare-earth-free permanent magnets, operational in 2027 — small against MP's planned ~10,000 tonnes, but Niron says a follow-on 10,000 tonne US plant breaks ground in 2028, with site selection already under way. If the second plant lands on schedule, the rare-earth-free capacity arriving at the end of this decade is the same order of magnitude as the Western NdFeB capacity being subsidised into existence today.

WatchSartell's commissioning, still pointed at 2027 now that the structure is going up, whether the 10,000-tonne follow-on site is actually named, and whether any other original-equipment maker follows Honda in. Track it as a threat to the tier-two thesis rather than as a position — the first credible rare-earth-free magnet at volume reprices every name in that tier downward.

Proterial — formerly Hitachi Metals, a genuine heavyweight in advanced magnetic materials and rare-earth-free development — has not been buyable since October 2022, when a Bain Capital-led consortium completed a roughly $7.5bn take-private and the shares were delisted. It was renamed in January 2023. Any list presenting it as a way to invest in magnets is out of date by nearly four years.

Noveon Magnetics is a private Texas manufacturer producing sintered NdFeB through its EcoFlux recycling process — a closed-loop route that sidesteps mining entirely. It has raised a $75m Series B and is operational, which makes it a real competitor to MP and USAR, just not a tradeable one.

On the Chinese side, the incumbents are what make the ~90% midstream share real. Two names circulating in research summaries, Ningbo Zhanhao and Mainrich Magnets, could not be verified as listed entities and should not be treated as tickers. The clearest listed Chinese pure-play in rare-earth permanent magnets is Ningbo Yunsheng (600366.SS) — A-share access and Chinese policy risk apply, which is much of the reason the West is building alternatives at all.

WatchConsolidation. A Western magnet maker being acquired, floated or spun out is the most likely route by which this layer becomes investable at all — and every name here is a candidate.