Rare-Earth Magnets in Humanoid Joints: The Neodymium Inside Every Actuator – and Who Controls It
Rare-earth magnets in humanoid joints: the neodymium inside every actuator
Every rotary joint in a humanoid spins on a ring of neodymium-iron-boron magnets. Take them out and the robot gets heavier or weaker – there is no third option on the market yet. That makes a few kilograms of sintered magnet, almost all of it made in one country, the most politically exposed part of the actuator.
Sources: Foundation for American Innovation, 2026; Rare Earth Exchanges, 2026; Fastmarkets, 2026. Per-robot figures are analyst estimates.
Where the magnets sit in a joint
Almost every humanoid joint motor is a permanent-magnet synchronous machine. A wound stator sits on the outside; a rotor carrying a ring of alternating north and south magnet poles turns on the inside. Current in the copper builds a rotating field, the magnets chase it, and the reducer or screw behind the rotor turns that spin into joint torque. It’s the same principle in the frameless torque motor of a hip, the motor behind a roller screw in a knee and the tiny motors in a dexterous hand. A single humanoid carries 30 or more of them.
The magnet of choice is sintered NdFeB, roughly 30% rare earths (mainly neodymium and praseodymium), 69% iron and 1% boron. It stores around 400 kJ/m³ of magnetic energy, more than any other commercially viable material, and that energy density is what lets a palm-sized motor produce the torque a biped needs without the limb turning into a club.
Figure 1 · Animated · Cross-section of a joint motor, with the NdFeB magnet ring on the rotor
Illustration: humanoid.guide. Generic surface-magnet rotor with 10 poles and 12 stator slots; schematic, not to scale. Animation is CSS-only and stops when reduced motion is requested.
Why not cheaper ferrite? Because mass compounds in a walking machine. A ferrite motor has to be about 30% heavier to deliver the same torque, and in a leg that extra weight loads the joint below it, which then needs a bigger motor of its own. Engineers can trade magnet grade for copper, current and cooling, but they can’t trade it for nothing.
Heat is why dysprosium matters
Neodymium magnets have a weak spot. Their coercivity – the resistance to being demagnetised – falls as they warm up, and a joint motor is a warm place. The magnets sit centimetres from windings that dump heat during every torque peak, and the same peak current throws an opposing field straight at them. Push a magnet past its limit and it loses strength permanently. The joint gets weaker for good.
The fix is chemistry. Magnet makers substitute some neodymium with the heavy rare earths dysprosium and, to a lesser extent, terbium, which raise the anisotropy field and with it the coercivity. The more heat a grade must survive, the more heavy rare earth it carries. That’s the hidden bill: analyst estimates put a humanoid at 0.57–1.3 kg of neodymium-praseodymium and 35–79 g of dysprosium. Terbium is roughly 2.5 times as effective as dysprosium per gram, and priced accordingly – more than four times as much.
Table 1 · NdFeB temperature classes
| Grade suffix | Nominal max operating temperature | Heavy rare earth (Dy/Tb) content |
|---|---|---|
| N (no suffix) | 80 °C | Little or none |
| M | 100 °C | Low |
| H | 120 °C | Moderate |
| SH | 150 °C | Higher |
| UH | 180 °C | High |
| EH | 200 °C | Very high |
| AH (TH) | 220–230 °C | Highest |
Source: Radial Magnet, NdFeB temperature classes (nominal ratings assume a permeance coefficient of about 1 or higher and no significant opposing field). Relative Dy/Tb content is indicative; real grades vary by maker and process.
Those temperatures are nominal, not promises. A magnet’s real limit depends on the magnetic circuit it sits in and how hard the drive pushes the current, which is why suppliers such as Mainrich tell actuator designers to size grades from hot demagnetisation curves rather than the suffix. The practical lever is grain-boundary diffusion: instead of mixing dysprosium through the whole alloy, it’s diffused into the grain boundaries where demagnetisation starts, reaching the same coercivity with less heavy rare earth. Savings are real, but every new route needs qualifying part by part.
The chokepoint map
China doesn’t just mine rare earths. It owns nearly every step between the ore and the finished magnet, and its share rises the closer you get to the part that goes into a joint.
Table 2 · China’s share at each stage of the magnet chain
| Stage | China’s share | Detail |
|---|---|---|
| Rare earth mining | 69% | 270 kt of 390 kt global output, 2025 |
| Separation | ~90% | Share of global separation capacity |
| Refining of magnet rare earths | 91% | NdPr, Dy and Tb refining |
| Sintered magnet production | 94% | Sintered permanent magnets, 2024 |
| NdFeB capacity, 2030 projection | >82% | ~280 kt/yr in China vs ~50–55 kt/yr rest of world |
Sources: GlobalData via Yahoo Finance (mining, 2025); Crosscurrents (separation, 2030 capacity); Rare Earth Exchanges (refining, sintered magnets, 2024).
Beijing has used that position twice in eighteen months. On 4 April 2025 it put licence requirements on seven medium and heavy rare earths – samarium, gadolinium, terbium, dysprosium, lutetium, scandium and yttrium – together with their metals, oxides, alloys and the permanent magnets that contain them. Humanoids felt it within weeks: Elon Musk said that month that Optimus production was being affected and that Tesla was working with Chinese authorities on an export licence for the magnets. Those April controls are still in force.
In October 2025 China went further, with extraterritorial rules reaching foreign-made products that contain Chinese-origin rare earths or were made with Chinese technology. After the Trump–Xi meeting that month, those measures were suspended for one year. Unless the pause is extended, they return on 10 November 2026. The price gap already shows where the pressure sits: in late August, dysprosium and terbium oxide in Rotterdam cost roughly four to five times the Chinese export price, and Fastmarkets reported that no dysprosium or terbium oxide had been exported to Japan at all in 2026.
Chart 1 · Heavy rare earth oxide prices: inside vs outside China
Source: Fastmarkets price assessments, 27 August 2026 (China FOB vs Rotterdam). Bars show the assessed low–high range.
There’s an asymmetry here that matters for the humanoid race. Chinese makers built more than 40,000 humanoids in the first half of 2026, around 97% of reported global shipments, and they buy magnets on the domestic market with no export licence in the loop. A Western OEM designing the same joint pays more, waits longer and carries licensing risk on every shipment.
The volume math: it’s not the tonnage
Humanoids won’t strain the world’s magnet supply by sheer weight any time soon. At 2–4.5 kg each, a million robots a year would need 2,000–4,500 tonnes of NdFeB. Electric cars already used about 30,100 tonnes in 2025 – 21 million vehicles at roughly 1.4 kg each.
Non-Chinese capacity is coming, but late. MP Materials’ 10X campus in Texas is sized at around 10,000 tonnes of magnets a year, with commissioning from 2028. USA Rare Earth’s South Carolina plant targets 6,400 tonnes a year of sintered NdFeB, also from 2028. Even with those, projections leave China with more than 82% of global NdFeB capacity in 2030. For the next two years, a humanoid maker outside China that needs heat-resistant grades is buying from Chinese magnet plants or from the small pool of alternatives in Japan and Europe.
Ways out, none of them quick
Actuator teams have four levers. Each costs something.
- Cut the dysprosium, not the magnet. Grain-boundary diffusion and Dy-lean grades hit the same coercivity with less heavy rare earth, at the price of requalifying every part.
- Run the magnets cooler. Better heat paths from winding to housing let a designer drop a temperature class – from UH to SH, say – and every class down means less Dy or Tb in the alloy.
- Accept ferrite where mass matters least. Torso and waist joints can swallow a 30% heavier motor more easily than an ankle can.
- Wait for new materials. Niron Magnetics’ iron nitride is rare-earth-free by design; its July 2026 agreement with ASPINA covers brushless DC and stepper motors, backed by a Minnesota plant with a conditional $150 million federal loan. No humanoid joint runs on it yet.
The date to watch is 10 November. If the October 2025 rules come back, the licence question stops being about magnets shipped from China and starts covering motors and actuators built elsewhere with Chinese material inside – which today means almost all of them.
Go deeper: The Humanoid Actuation Report
249 pages on motors, reducers, screws, bearings, encoders and drives, the vendors behind every major platform, and the cost curves that decide adoption to 2036.
The Humanoid Actuation Report All joints & actuators articles All humanoid.guide reportsKeep reading · Joints & actuators
Related articles
FAQ
How much rare-earth magnet does a humanoid robot contain?
Analyst estimates from the Foundation for American Innovation put it at 2–4.5 kg of NdFeB magnet per robot, containing roughly 0.57–1.3 kg of neodymium-praseodymium and 35–79 g of dysprosium. The spread comes from robot size, joint count and how much thermal headroom the motors are designed with.
Why do humanoid actuators need dysprosium?
NdFeB magnets lose coercivity as they heat up, and joint motors run hot during torque peaks. Dysprosium and terbium raise coercivity so the magnets survive higher temperatures and opposing fields without permanent demagnetisation. Higher temperature classes such as SH, UH and EH carry more of them.
Can humanoid motors use ferrite or rare-earth-free magnets?
Yes, at a cost. A ferrite motor needs to be about 30% heavier for the same torque, which hurts most in legs and arms. Iron nitride magnets from Niron Magnetics are moving into brushless DC motors through partners like ASPINA, but no humanoid joint is known to use them yet.
What happens on 10 November 2026?
That’s when China’s October 2025 export-control package, suspended for one year after the Trump–Xi meeting, is due to return unless extended. It would extend licensing to foreign-made products containing Chinese-origin rare earths. The April 2025 controls on seven heavy and medium rare earths and their magnets remain in force regardless.
Is the magnet supply big enough for mass-produced humanoids?
By volume, mostly yes: a million robots a year would need 2,000–4,500 tonnes of NdFeB, against roughly 30,000 tonnes already used by electric cars in 2025. The risk is concentration. China makes about 94% of sintered magnets, and new capacity outside China from MP Materials and USA Rare Earth only starts commissioning in 2028.
Sources
- Foundation for American Innovation – The Robot Boom Has a Magnet Problem (NdFeB, NdPr and Dy per robot; EV magnet demand; Dy/Tb supply; ferrite weight penalty)
- Rare Earth Exchanges – Humanoid Robot NdFeB Magnet Supply, China vs US 2026 (refining and sintered-magnet shares; H1 2026 humanoid output; MP Materials and USA Rare Earth capacity)
- Fastmarkets – China’s looming November export controls test rare earth refining, recycling ambitions (10 November 2026 date; Dy and Tb oxide prices)
- Yahoo Finance / GlobalData – China rare earth export pause nears expiry (April 2025 element list; 2025 mining share)
- CSIS – Rare Earth Export Restrictions One Year Later (April and October 2025 controls; one-year suspension)
- TrendForce – Tesla’s Robot Plans Stumble Amid China’s Rare Earth Export Curbs (Musk on Optimus and export licences, April 2025)
- Crosscurrents – NdFeB Magnets: The Material Bottleneck Behind Humanoid Robotics (magnet composition and energy product; separation share; 2030 capacity)
- Katusa Research – Every robot needs 30 magnets (motor count per humanoid)
- Radial Magnet – Neodymium Magnet Temperature Classes (grade temperatures; Dy/Tb and grain-boundary diffusion)
- Mainrich International – Magnets for Humanoid Robot Actuators (grade selection from hot demagnetisation curves; GBD qualification)
- Niron Magnetics – Niron and ASPINA to develop rare-earth-free motor solutions (July 2026)
