A neodymium magnet is not on China’s export control list. A neodymium magnet containing terbium or dysprosium is. What separates the two is heat: the heavy elements are what let a magnet hold its field when it gets hot. So the control is not really written on magnets. It is written on thermal margin, and a humanoid robot’s joints are exactly where that margin runs out.

The line is drawn on composition, not on the part

On 4 April 2025, China’s Ministry of Commerce and the General Administration of Customs published Announcement 18, putting items derived from seven medium and heavy rare earth elements on the dual-use control list. It exists in Chinese only; the wordings below are my translations. Three of its item codes name permanent magnet materials outright: 1C902.a.4 samarium-cobalt, 1C904.a.4 “terbium-containing NdFeB permanent magnet materials”, and 1C905.a.4 “dysprosium-containing NdFeB permanent magnet materials.” An explanatory note extends those three, and only those three, to “magnet bodies or magnet powder.”

Neodymium is not on that list. Neither is iron, nor boron. The control is defined by what the alloy contains.

That sounds like a technicality until you ask what decides the contents.

A robot joint cannot stay on the low grades

NdFeB magnets are sold in grades, and the letters after the number are a temperature rating. Arnold Magnetic Technologies publishes the ladder: standard N grades top out around 60 to 80°C, M at 100, H at 120, SH at 150, UH at 180, EH at 200 and AH at 220. Coercivity is what buys those degrees, and heavy rare earths are what buy coercivity. The industry’s own workaround says so plainly: grain boundary diffusion exists, in Arnold’s words, for “reducing the amount of Dysprosium (Dy) (GBDD) and Terbium (Tb) (GBDT) used.”

Now put a robot joint on that ladder. In a magneto-thermal study of a high torque density joint motor built for humanoids, Zhang and colleagues report a magnet surface temperature of 105°C, with the winding at 119.5°C. Standard and M grades are already out at that number.

The authors do not present that as a problem. They note it “will not cause permanent magnet demagnetization,” which is itself the point. For that to hold at 105°C, the rotor cannot be running a bottom-of-the-ladder grade.

Read the 105 as an illustration rather than a specification. It is one motor, from one group, and calculated rather than measured. Two caveats cut opposite ways. The paper’s table labels these figures a temperature rise; if so, the magnet sits nearer 130°C on a 25°C ambient. Against that, the method was checked against hardware: the same table puts calculated winding temperature at 119.5°C and measured at 115.5°C. What survives either way is the shape: a motor sized for torque density, buried in a limb with no cooling loop, runs far above 80°C. Continuous torque rather than peak is what a duty cycle actually spends, and that gap is one of the numbers demos never show you.

The dependency is not 71%

Rare earth concentration is usually quoted for the category as a whole, and that number is bad enough. China mined 270,000 of the world’s 390,000 tons in 2025, and over 2021 to 2024 China supplied 71% of US imports of rare earth compounds and metals.

The heavy elements are a different order of problem. USGS puts US net import reliance for heavy rare earth compounds and metals at 100% in every year from 2021 through 2025. For terbium, the import source line reads “China, 100%.”

And the same section makes even that a floor. It closes with a caveat: “Import sources do not include heavy rare earths contained in value-added intermediates and finished products.” A magnet arriving inside a finished motor is not in the count.

The price is the wrong instrument

Here is why this is hard to see from a financial statement. The controls did not reprice these materials the way a hard shortage would.

USGS has dysprosium oxide at $239/kg in 2025, down from $410 in 2021, and terbium oxide at $1,010 against $1,340. Both are below where they started, and terbium is half its 2022 level of $2,050.

The control year is not flat, though, and reading only from 2021 hides that. USGS publishes prices for five of the seven April-controlled elements. Four rose in 2025: samarium 40% and gadolinium 7%, terbium 24% and lutetium 14%. Dysprosium alone fell, by 7%. The 2025 figures are USGS estimates. That is a market ticking up inside a range it has covered before, not the step change a physical shortage produces.

A cheap input under a licence regime is not a market. It is an allocation. In the last post I argued that a capacity price pinned to a regulatory cap stops carrying information, so the shortage surfaces as quantity instead. This is the same failure from the opposite side: here the price is free to move, it has moved both ways, and the constraint is real regardless. The binding term is permission, not price.

What it costs

A licence is not a price. It admits no partial answer: the magnet ships or it does not, whatever the buyer will pay. That binary is the exposure.

And the choice is made years before anyone reads a control list. Thermal design fixes the grade, the grade fixes the composition, and the composition decides whether the part needs an export licence. Re-sourcing means redrawing the motor, not renegotiating a contract. Whatever optionality a programme has here, it spent at the drawing board.

The US Government stockpile’s fiscal 2025 potential acquisitions included 450 tons of neodymium-iron-boron magnet block, alongside 60 tons of samarium-cobalt alloy. A stockpile line is not proof of scarcity, but it is one government putting a quantity on the risk.

Where I could be wrong

Licences are being issued. USGS notes that China “began to issue general export licenses to selected exporters”, and China’s stated position is that this is a licensing regime, not a ban. If general licences become routine, the gate stops binding and most of this argument goes with it.

Public money is moving against this. In 2025 the US Department of War lent $150 million to build a heavy rare earth separation facility in Mountain Pass, California, and put $80 million and $5.1 million into two recyclers recovering terbium and dysprosium. In November the US International Development Finance Corporation approved a $465 million loan to a Brazilian heavy rare earth producer. That is roughly $700 million of announced public funding. None of it is capacity today: USGS says no US producer reached sustained commercial-scale output in 2025. But it is the clock this thesis runs against.

I cannot tell you which grades are actually caught. The announcement draws its line at “containing” terbium or dysprosium, and its explanatory notes state no minimum content. Magnet makers do not publish composition by grade. So a datasheet will give you an operating temperature and will not tell you whether the part needs a licence. That opacity cuts against precision in everything above.

The obvious substitutions are also controlled. Samarium-cobalt is the standard answer for high-temperature magnets and sits on the same list at 1C902.a.4. Grain boundary diffusion reduces heavy rare earth content rather than removing it, so it addresses cost, not a control written on containment.

And the timestamp matters. USGS assessed the April controls as still in effect as of December 2025, after a separate set of October additions was suspended for a year in November. Read every claim here against that date.

Every company building humanoid actuators knows what share of its magnets would need an export licence to leave China. None of them publish it. And the price of dysprosium has spent four years falling while the answer got worse.

---

Correction, 6 September 2026. The closing section originally began by calling a magnet “a rounding error in a robot’s bill of materials.” A share of a bill of materials is something you measure, and I had no source for it, so the sentence should not have run. It has been replaced with one that makes the same point about the licence being binary without asserting a cost share. Nothing else changed. The case that these materials are cheap rests on the USGS price series quoted above, which is unaffected.