Forecasters broadly agree that a humanoid robot boom is coming, but they disagree wildly on how many robots will be built. Scenarios range from Goldman Sachs's "bear case" estimate of 700,000 robots per year by 2035 to Elon Musk's prediction of 10 billion humanoid robots by 2040 (implying an average annual production of nearly 1 billion robots through the 2030s).
Many of these forecasts acknowledge the same risk: magnet production and the upstream refining of rare earth elements (REEs) could bottleneck robot production. Robot actuators require high-torque motors that currently rely on REE-dependent neodymium-iron-boron (NdFeB) magnets. To perform at the temperatures robots demand, these magnets need “heavy” REEs such as dysprosium (Dy) or terbium (Tb), both of which are produced at negligible quantities outside of the Chinese sphere of influence.
Is pessimism warranted? It’s complicated. Today’s electric vehicle industry already consumes as much heavy rare earth each year as the production of 26 million humanoid robots—more than all but the most aggressive forecasts. Furthermore, a humanoid boom would shift relative rare earth demand away from the heavy elements China monopolizes and toward light elements that the West already produces. This shift works in the U.S.’s favor, but Western capacity is sparse enough that either input could still bind. To reveal the severity of these bottlenecks, we’ll walk through three questions: how much rare earth goes into one robot, how that demand stacks up against the industry already consuming magnets in bulk, and whether substitutes for rare earth magnets can alleviate the pressure.
Before we can answer these questions, we must clarify the distinction between light and heavy magnet-relevant rare earths. The light rare earths (Nd and Pr) make up over 25 percent of most NdFeB magnets and are crucial for magnetization. The heavy rare earths (Dy and Tb) replace some of the joint NdPr, sacrificing strength for heat resistance. Most NdFeB magnets are under 5 percent heavy rare earth by weight, but Dy and Tb are scarcer in nature and subject to Chinese export controls. This makes them far more expensive than NdPr outside China.
In NdFeB magnets, Dy substitutes for Tb at a rate of 2.5 to 1: you need 2.5 times as much Dy to match the heat resistance Tb provides (throughout this piece, "Dy-equivalent metal" therefore means Dy + 2.5Tb). Because Tb does the same work with less material, it leaves more of the magnet for NdPr. Tb is reserved for motors that need both high torque (from the greater NdPr quantity) and high heat resistance (from the Tb itself), and it commands prices over four times Dy’s despite being only 2.5 times as efficient.
How Much Rare Earth Goes into a Single Robot?
To estimate the quantities of light and heavy rare earths used in the production of a humanoid robot, we need three figures: the quantity of NdFeB magnet required, the grade of these magnets, and the element breakdown of these particular grades.
Published estimates of the per-robot magnet quantity converge on the same range: 2–4.5 kg of NdFeB magnet per robot. Magnet grades are ratings of heat resistance, running from N (standard) up through M, H, SH, UH, EH, and AH (the most heat-tolerant). Robot producers do not disclose the grades used in their actuators, but the middle H–SH range is a reasonable assumption, based on coil winding specifications. That leaves the element breakdown.
A 2022 Department of Energy (DOE) report on rare earth permanent magnets provides a table with exactly what we need. The table shows that 1 kg of H-grade NdFeB magnet requires 28 g Dy and 272 g NdPr, while 1 kg of SH-grade NdFeB magnet requires 42 g Dy and 258 g NdPr. The report notes that a “thrifting” technique called grain-boundary diffusion lets manufacturers reach a given grade with less Dy, often cutting Dy content by 50 percent or more. Our calculations therefore assume magnets use half the Dy listed in the DOE table, with NdPr making up the remaining weight. Assuming an even split between H- and SH-grade magnets, each with 50 percent Dy thrifting, a robot contains roughly 0.57–1.3 kg of NdPr and 35–79 grams of Dy. But these figures mean little in isolation. A better sense of scale comes from comparing them to the industry that already buys these magnets in large quantities.
Scaling Up: Comparing Humanoid Robots to Electric Vehicles
Today's electric vehicles (EVs), like today's robots, use NdFeB magnets. An EV uses a sales-weighted average of 1.4 kg of NdFeB magnet—approximately half of what’s used in a humanoid robot. But the magnets used in EVs must be rated for much higher temperatures; these magnets are typically graded EH or AH (the commercial grades with the highest heat resistance). Even after thrifting, dysprosium can make up between 4.25 percent and 5.5 percent of the magnet by weight. A typical EV thus uses half the magnet quantity of a humanoid robot, but at a grade requiring 2.8 times the heavy REE per kg. This means one EV represents the Dy-equivalent demand of about 1.2 robots.
In 2025, 21 million EVs were sold. These vehicles contained approximately 19.4 million permanent-magnet motors, consuming 30.1 million kg of NdFeB magnets. In conjunction with our conversion factor, this implies that approximately 26 million humanoid robots would need to be produced each year to match the heavy-REE demand presented by the present-day EV industry.
This quantity of EV magnets requires approximately 1.47 million kg of Dy-equivalent metal per year. According to an International Energy Agency report on rare earth elements, 2.7 million kg of Dy and 800,000 kg of Tb were sold in 2024—together, 4.7 million kg of Dy-equivalent metal. EVs alone, then, consume approximately 31 percent of the world’s Dy-equivalent supply.
The Bank of America Institute projects 10 million robots produced annually by 2035, while Goldman Sachs’s “base case” estimate is 1.4 million. The 10 million projection would require approximately 12 percent of existing heavy REE supply and the 1.4 million projection would require 1.7 percent. The upshot is that heavy rare earths would become a real limiting factor under the higher projections, while worlds with lower humanoid demand are less likely to feel this constraint.
Note that these figures compare global demand to global supply. A tightening of export controls on heavy REEs could still be disastrous for both the EV and robotics industries in the West, concentrating the production of both in China itself. Efforts to secure the critical mineral supply chain must remain a top priority.
What about Light Rare Earths?
Dysprosium is only half the story. 10 million robots would also consume 9 million kg of NdPr (0.92 kg per robot). That is a tenth of the approximately 87 million kg of NdPr sold in 2024, and more NdPr than is consumed by a full year of EV production. A humanoid robot boom would thus tilt relative REE demand away from the China-concentrated heavy rare earths and toward light rare earths, which exist in ample reserves outside China—an inversion of the usual story.
Unlike the heavy rare earths, the non-Chinese supply chain already refines NdPr at scale—roughly 7.8 million kg per year today (calculated from Lynas and MP Materials 2025 reports; other producers exist but produce at much smaller quantities). This is the equivalent of 8.7 million robots, but robot producers would still have to compete with other industries for access.
There is far more mining of NdPr outside China and Myanmar than refining (roughly 16 million kg of NdPr mined per year compared to the 7.8 million kg refined). In other words, we currently mine enough NdPr for approximately 17.8 million robots per year. So for the immediate future, refining bottlenecks Western light REE production. And unlike with Dy and Tb, China does not currently export-control refined NdPr. This means NdPr is less likely to bind in the short-run than the heavy REEs, but medium-run strain remains—unless progress in NdFeB magnet alternatives accelerates.
Alternatives to NdFeB Magnets
Faced with incoming scarcity, the robotics industry is exploring ways to reduce its dependence on rare earths. The first option is ferrite magnets. Motors using ferrite magnets need to be about 30 percent heavier to maintain the same torque, and also tend to be physically larger. They can probably work for most joints, especially as further advances in motor engineering mitigate their disadvantages. But for small joints, especially in hands, the downsides of ferrite motors start to bite. Still, hands account for only a modest share of a robot’s magnet mass, so improvements to ferrite motors could still cut NdPr and Dy requirements substantially.
The second option is “magnet-free” motors. These motors have already seen deployment in electric vehicles, but they are unlikely to be the optimal design for robots. Magnet-free motors require continuous current to maintain a magnetic field, while the magnetic field in NdFeB and ferrite motors is permanent. Joints therefore use more energy to hold position, which creates more heat buildup and battery drain. They are much better suited for substitution in EVs, where motors do not hold static positions. Even if magnet-free motors never appear in a robot, their diffusion through the EV fleet would ease REE demand and free up magnets for robots.
The final option is the use of higher gear ratios in robot joints. Higher gear ratios allow for the use of lower-torque motors with smaller magnets, without sacrificing strength. These don’t replace NdFeB magnets, but they can substantially reduce the quantity needed. This can already be seen in the difference in architecture between Western and Chinese humanoids: Western humanoids tend to use high-ratio harmonic drives, while Chinese humanoids tend to use low-ratio quasi-direct-drives (QDDs)—architectures that pair larger, higher-torque motors with minimal gearing. In the estimates of “2–4.5 kg of NdFeB magnet per robot,” the lower end corresponds to robots with high gear ratios, the higher end to QDDs. For Chinese companies with access to cheap magnets, QDDs cut expensive precision reducers out of the bill of materials. But as my colleague Amelia Michael and I argued in a recent paper, precision reducers do not bottleneck the Western robotics supply chain. So these architectural differences naturally emerge from the constraints of the markets that produce them.
The Bottom Line
Rare earths bind the most optimistic scenarios regardless of where the robots are built. Aggressive forecasts imply REE consumption that would require a significant expansion of global supply. In these worlds, even China would struggle to meet short-run demand from existing capacity, especially in light rare earths. Robots increase demand for NdPr faster than demand for heavy rare earths, a shift that slightly favors the light-REE capacity the West already has. But Western light-REE capacity is still limited, and heavy-REE capacity is in its infancy. So despite the favorable shift, either input could end up as the binding constraint. If Chinese export controls tighten, robot production outside China scales only as fast as non-Chinese mining and processing of both light and heavy REEs can grow.
This expansion is still in its earliest stages, but it shows promise. Brazil holds the world’s second-largest rare earth reserves, including deposits rich in heavy REEs. Its Serra Verde mine has been extracting all four magnet rare earths since 2024, making it the only significant heavy-REE source outside Asia. Output is still modest (Serra Verde is targeting an extraction volume of 6.4 million kgs of rare earth oxides annually by the end of 2027), and separation remains a bottleneck in the West. But these heavy-rich deposits also carry substantial light rare earths, while mines on light-dominant deposits hold only traces of heavies. Further buildout of heavy-REE mining therefore increases light-REE capacity as well.
Promising mitigations exist, but they all point the same direction. Ferrite motors, magnet-free designs, and higher gear ratios each reduce China dependence for magnet inputs, but none eliminates it. What limited non-Chinese capacity exists will therefore have to be triaged across industries. NdFeB magnets will flow to the applications that can least substitute away from them, while industries that feel the downsides of alternatives least will be pushed toward ferrite and magnet-free designs first.
