Where AGI's bid for hands and legs runs into the wall of specialty manufacturing. Picks-and-shovels > humanoid OEMs.
Once recursive self-improvement in cognition starts to saturate, the binding constraint moves to the physical world. Every AGI lab, every government, and every Tier-1 manufacturer will want embodied agents — humanoid for general labor, specialized for fabs and labs — as fast as possible. The brains are getting cheap. The bodies are not.
The interesting question isn't "who builds the best humanoid?" — it's "what does every humanoid OEM need that physically cannot be expanded fast enough?" The answer is a short list: precision strain wave (harmonic) and RV reducers, rare-earth permanent magnets, high-resolution force/tactile sensors, frameless torque motors, and high-end machine vision. That's where the rent goes. Humanoid OEMs themselves are racing to commoditize each other; the component suppliers behind them are not.
| Source | 2026 units | 2027 units | Notes |
|---|---|---|---|
| Tesla Optimus (Musk public target) | Gen-3 prod. launch summer | 1M/yr target (aspirational) | Texas Gigafactory dedicated line under construction; 10M/yr ceiling claimed long-term. |
| Unitree (China, IPO-track) | ~20,000 (target); 75k capacity | est. 40–80k | Quadrupeds 115k capacity. Humanoid revenue now >51% of mix. 60% combined GM. |
| AgiBot (China) | 10,000+ Expedition A3 shipped by Mar | est. 20–30k | Scaled 1k→5k→10k in three months. Auto + electronics + logistics customers. |
| Figure AI (BMW + Brookfield) | low thousands | 10k+ planned | $39B post-money Series C (Sep 2025). BotQ factory ramp. Helix model. |
| Boston Dynamics Atlas (Hyundai) | First commercial units at Metaplant GA | Broader enterprise rollout | $140–150k unit price; 56 DoF; Gemini Robotics onboard. |
| China industry aggregate (TrendForce) | +94% YoY output growth | ~150k+ aggregate | Unitree + AgiBot ~80% of Chinese shipments. |
Reality check: in 2026, total real humanoid shipments worldwide are most likely in the 40–80k range, not Musk's million. But every one of those robots needs ~30 motors and ~14–28 precision reducers. The component math is what matters, not the headline.
| Component | Per humanoid | 2026 unit demand (60k bots, midpoint) | Capacity status |
|---|---|---|---|
| Strain wave (harmonic) reducers | 10–20 | 0.6–1.2M units | Tight Harmonic Drive Systems dominant; Leaderdrive (China) ramping. |
| RV / planetary reducers (heavy joints) | 4–8 | 240–480k | Reorganizing Nabtesco losing share (51.8%→33.8% China) to Fine Motion. |
| Frameless torque / BLDC motors | 28–40 | 1.7–2.4M | Adequate but premium tier tight Maxon, Kollmorgen, Nidec, Allied Motion. |
| NdFeB permanent magnets | ~2–4 kg | 120–240 t | Severely constrained NdPr oxide doubled in 7 months to $120k/kg. Mine lead time ~18 yrs. |
| 6-axis force/torque sensors | 2–6 | 120–360k | Tight ATI, Schunk, OnRobot — small specialty players. |
| Tactile / fingertip sensors | 10–20 | 600k–1.2M | Tight, immature GelSight, Contactile, BeBop — mostly private/early. |
| Depth cameras / lidar | 1–3 | 60–180k | Adequate Hesai pivoting to robotics; Intel RealSense; Orbbec. |
| Edge AI inference SoC | 1–2 | 60–120k | Adequate NVDA Jetson Thor, Qualcomm, Ambarella. See compute track. |
A precision strain wave reducer is one of the most demanding mass-manufactured mechanical parts on earth: sub-arcminute backlash, gear-grade alloy steel, hand-finished tolerances on a flexspline that flexes elastically every revolution. Harmonic Drive Systems (HDS, Tokyo: 6324) has spent five decades building that know-how. Capacity is not GPU-fab-bad, but you can't 10x it in 24 months by spending money — you need engineers who have seen 10,000 flexsplines fail and know why.
The same is true at the high end of RV reducers (Nabtesco) and torque sensors (ATI). These are quiet, century-old craft businesses that suddenly find themselves in the path of trillion-dollar capital.
That is the textbook setup for pricing power: a commodity-cost input becomes a scarce input, and the supplier extracts a much larger share of the value. Harmonic Drive's revenue is JPY 59.6B today; even if humanoid grows reducer ASPs only 40% and volume 3x, the math gets very interesting. The catch — see below — is that the stock has already moved 148% in a year and trades at ~471x trailing earnings.
Rare-earth magnets are the more uncontroversial bottleneck. NdPr oxide doubled in seven months to >$120k/kg. China still controls ~90% of refining. Each humanoid wants 2–4 kg of NdFeB. At Musk's 10B-by-2040 claim, you'd need >100x current global production. Mine lead times average 17.9 years. This is the cleanest physical-impossibility gap in the whole stack.
| Ticker | Market cap | Multiple | Why a bottleneck | Risks | Priced-in |
|---|---|---|---|---|---|
| 6324.T Harmonic Drive Systems | ¥757B (~$5.0B) | P/E ~471x | Dominant strain wave reducer for nearly every humanoid joint. Five decades of flexspline know-how. | Multiple is bonkers. China challengers (Leaderdrive, Fine Motion) catching up on patent-expired designs. Cyclicality if humanoid hype pauses. | 9 / 10 |
| 6268.T Nabtesco | ~¥500B (est) | P/E ~20x (est) | RV reducers for heavy joints, robotic precision actuators in industrial arms. | China share collapsed 51.8% → 33.8% (2021→2024). Fine Motion + Leaderdrive eating mid-tier. Domestic Japan exposure is solid but not growing. | 3 / 10 |
| 6594.T Nidec | ¥3.24T (~$21B) | P/E 27x, Fwd 19x | Precision motors at scale; the only company with the capex to flood humanoid motor demand. | Active accounting scandal (1,000+ quality misconduct cases); third-party investigation; cancelled dividend. Auto EV motor business under pricing pressure. | 4 / 10 |
| 6861.T Keyence | ~¥17T (est) | P/E ~35x | Best-run industrial business on earth (~55% op margin). Machine vision is the eyes for every reconfigured factory. | Already a quality-compounder darling; not undiscovered. Direct humanoid exposure is small relative to size. | 6 / 10 |
| 6954.T Fanuc | ¥7.5T (~$48B) | P/E 45x, Fwd 35–38x | Global #1 industrial robot; 21% op margins; ¥50B buyback announced. | ~45% of robot sales = China, which is rapidly domesticating (Estun, SIASUN). Industrial cobots, not humanoid, is the swim lane. | 5 / 10 |
| CGNX Cognex | ~$5B (est) | P/E ~35x (est) | Western alternative to Keyence in machine vision; barcode + factory automation. | Cyclically tied to automotive/electronics capex; lost some share to Keyence over decade. | 5 / 10 |
| HSAI Hesai Group | $3.35B | ~7x sales | #1 long-range ADAS lidar; "spatial intelligence" pivot to robotics (Kosmo platform, actuator modules); doubling shipments 2026. | US listing geopolitical (China-domiciled); auto lidar competition (Robosense, Innoviz) brutal. | 5 / 10 |
| MP Materials | ~$10B (est at $60) | High; pre-profit on magnet biz | Only at-scale rare earth + magnet producer in Western Hemisphere; DoD $400M deal + price floor; Texas magnet plant scaling 10x. | Pricing volatility tied to China policy; magnet plant ramp 2028; Apple offtake helps but doesn't de-risk fully. | 5 / 10 |
| 005380.KS Hyundai Motor | ~$45B (est) | P/E ~6x | Owns Boston Dynamics; first deploying Atlas at Metaplant GA; effectively a free humanoid call option on a cheap auto stock. | Auto cyclical; BD optionality not yet material; KOSPI/Korea governance discount. | 2 / 10 |
| TECN.SW Tecan | ~$3B (est) | P/E ~25x (est) | Liquid handling for cloud labs / bio automation — overlaps biology track. Lab robotics gets bid alongside humanoid. | Post-COVID biotech capex hangover; competition from Hamilton (private) and Beckman. | 3 / 10 |
| RSW.L Renishaw | ~£3B (est) | P/E ~25x (est) | High-precision encoders, the eyes that close the loop on every robot joint. Quiet, durable IP moat. | Lumpy capex demand; UK listing illiquidity; family-controlled. | 3 / 10 |
| TER Teradyne (Universal Robots) | ~$18B (est) | P/E ~30x (est) | Owns Universal Robots (cobot leader) + MiR (mobile). Test-equipment core still cyclical but Robotics segment getting AI lift. | Semi test cycle drives the stock more than robotics today; UR margin pressure. | 4 / 10 |
"Priced-in" score is subjective, 1 = market hasn't woken up, 10 = priced for narrative perfection. Estimates flagged with italic text where exact market cap or multiple not verified from a primary source in this session.
| Company | Latest valuation | What they do | Why it matters |
|---|---|---|---|
| Figure AI | $39B post-money (Sep 2025 Series C, >$1B raised) | Humanoid + Helix VLA model | Investors include NVDA, Brookfield, Intel Capital, Qualcomm, Salesforce. BMW deployment. Most likely 2026/27 IPO candidate. |
| Physical Intelligence (π) | ~$11B (March 2026 round in progress) | General-purpose robot foundation models (pi-0, pi-zero) | Doubled valuation in 4 months. Founders Fund, Lightspeed, Thrive, Lux. The "OpenAI of robots" framing. |
| Unitree | STAR IPO filed | China humanoid + quadruped, 75k/115k capacity | Most likely "buy when listed" candidate. 60% gross margin, real shipments. Geopolitical risk for US investors. |
| AgiBot | Private; IPO talked | Expedition A3 humanoid, China | Scaled 1k→10k in three months — fastest production ramp in the space. |
| Apptronik | ~$1.6B (est) | Apollo humanoid, Mercedes/GXO pilots | US, NASA-DNA, less narrative noise than Figure. Watch for Series C-D. |
| 1X | ~$1B+ (est) | Neo home humanoid, OpenAI-backed | Consumer angle is differentiated; risky. |
| Agility Robotics | ~$1B+ (est) | Digit, logistics-focused | Real Amazon pilots; less hype but more revenue. |
| Maxon Motor (Swiss private) | Family-owned | Precision DC motors, the Swiss watchmaking of motion | Not investable directly; track as a tell — if Maxon raises prices or sells out, demand is real. |
| Hamilton Company (private US) | Family-owned | Liquid handling, lab automation rival to Tecan | Bio robotics adjacency. Not investable directly. |
| Heidenhain (German private) | Foundation-owned | Precision encoders, the eyes of CNC and robotics | Co-monopoly with Renishaw. Not investable directly — buy Renishaw as proxy. |
| Play | Vehicle | Comment |
|---|---|---|
| Rare-earth magnets (Western) | MP, Lynas (ASX:LYC), Energy Fuels (UUUU), USA Rare Earth | DoD floor pricing on MP changes the entire investment case. Lynas is biggest ex-China primary producer. |
| Specialty bearings | NSK (6471.T), SKF (SKFB.ST), Schaeffler (SHA.DE), JTEKT (6473.T) | High-precision crossed-roller bearings inside every reducer. Less constrained than reducers themselves, but volume tailwind. |
| Gear-grade alloy steel | Daido Steel (5471.T), Sanyo Special Steel (5481.T) | Niche Japanese specialty steel for flexspline blanks. Lumpy but a real tell on reducer demand. |
| Rare-earth-free magnet R&D | Proterial (5486.T), Aichi Steel (5482.T) | Long-shot moonshot — if it works, it kills the magnet thesis. Watch news flow. |
| Lithium (humanoid batteries) | ALB, SQM, Pilbara (PLS.AX), CATL (300750) | Energy density needs are 1–2 kWh per humanoid. Marginal demand vs auto, but adds to bid. |
The contrarian pick. Yes, they lost China share (51.8% → 33.8%). But (a) the multiple already reflects this, (b) Japan domestic + Western humanoid OEMs are de-risking China and rebuilding non-China supply chains, and (c) RV reducers for the heavy hip and shoulder joints are harder to substitute than strain wave. You're buying the leading non-China actuator IP at roughly mid-cycle industrial multiples while Harmonic Drive trades at ~471x. If humanoid volumes materialize at even half of consensus, this rerates. Risk: if Fine Motion + Leaderdrive achieve cost-quality parity, the moat shrinks faster than rerating happens.
The cleanest physical chokepoint in the entire AGI-embodied stack. NdPr oxide doubled in seven months. China controls 90% of refining. DoD has put a price floor under MP and is funding the Texas magnet plant. Apple's offtake adds a non-defense anchor. You don't need humanoid to scale for this thesis to work — you only need the West to remain unwilling to depend on China for the magnets in every motor, generator, and missile. Humanoid is upside option. Risk: NdPr price reversal if China relaxes export controls or substitution accelerates.
A 6x P/E auto stock with Boston Dynamics baked in for free. Atlas commercial production at the Hyundai Metaplant GA starts this year at $140–150k/unit, with Gemini Robotics onboard. The market is valuing Hyundai as an auto OEM with a perceived Korea governance discount; the Atlas optionality is a rounding error in the price. If even one humanoid OEM gets to material commercial volumes by 2028, BD is a credible top-3 candidate and Hyundai owns it. Risk: auto cyclical drawdown; BD never materializes commercially.
Not a pure humanoid play, but the closest thing to a "buy quality and wait" name in the track. ~55% operating margins, dominant in machine vision, every reshored factory needs more of their gear. Already a quality compounder so not undiscovered, but you're buying a structural beneficiary of physical-AI capex without having to be right about which OEM wins. Risk: the multiple compresses if Japan equities derate; humanoid is small share of revenue today.
Pivot to robotics is real (Kosmo platform, actuator modules for spatial intelligence). Doubling shipments in 2026. Robotaxi optionality on top. Cheap-ish at $3.35B mkt cap with 30% revenue growth. Risk: China ADR overhang is real; any geopolitical event can mark this down 30% overnight. Size accordingly.
Report date: 2026-05-26. All live prices and market caps as of close 2026-05-26 unless flagged otherwise. Estimates labeled italic.
Author's note: This is a snapshot at one point in time. Several figures (Nabtesco mkt cap, Renishaw mkt cap, CGNX multiples, Tecan multiples) are marked as estimates because they were not pulled from primary sources in this session. Verify before sizing positions.