DailySandThe term "rare earth elements" is misleading on two counts. The 17 elements in the rare earth group — ranging from lanthanum to lutetium, plus yttrium and scandium — are not particularly rare in Earth's crust. Cerium, for example, is more abundant than copper. And they are not always in the form of earths (oxides) in their useful applications.
What makes them genuinely rare is economic availability: the geological conditions that concentrate rare earths to mineable grades are uncommon, refining them requires specialized hydrometallurgical processes with significant environmental byproducts, and the global refining capacity is almost entirely located in one country.
That country is China, which controls approximately 85% of global rare earth processing capacity. This concentration — not the geological scarcity — is what makes rare earths a strategic concern for the AI industry.
The Rare Earths Inside Every AI Chip
Rare earth elements appear at multiple points in the AI chip supply chain, some visible and some invisible.
Cerium is used in the chemical mechanical planarization (CMP) slurries that polish silicon wafers between photolithography steps. Every wafer manufactured at TSMC, Samsung, or Intel Foundry undergoes dozens of CMP steps. Cerium oxide is the abrasive of choice because it polishes at rates that are controllable and repeatable at sub-nanometer precision. There is no commercially deployed alternative for advanced node manufacturing.
Yttrium is a component of yttria-stabilized zirconia (YSZ), used as a thermal barrier coating in the high-temperature components of gas turbines — including the turbines that power data center backup generators and the gas peaker plants that provide grid stability when renewable energy output drops. It is also used in phosphors for certain LED lighting applications in cleanroom environments.
Europium and terbium are essential in the phosphor systems of certain display technologies. More relevantly for AI infrastructure, terbium is alloyed with neodymium-iron-boron magnets alongside dysprosium to provide temperature stability in high-performance motor applications.
Neodymium and dysprosium deserve the most attention, as they are the rare earths most directly exposed to AI infrastructure growth.
Neodymium-Iron-Boron: The Magnet That Powers AI
NdFeB permanent magnets are the strongest permanent magnets known to materials science — roughly 10 times more powerful by weight than conventional ferrite magnets. This strength-to-weight ratio makes them the default choice wherever motor efficiency and size matter: hard disk drive actuators, electric vehicle traction motors, wind turbine generators, industrial pump motors, and data center cooling system fans.
Every hard disk drive in a conventional storage system contains NdFeB magnets in its actuator arm. AI training runs generate enormous amounts of data that must be stored during and after training — even solid-state storage arrays, which use less NdFeB than HDDs, contain NdFeB in their cooling system motors.
A single hyperscale data center may contain 50,000 to 200,000 individual fan units, each using NdFeB magnets. As liquid cooling adoption accelerates — driven by the high thermal design power of AI accelerators — pump motors replace or supplement fan motors, and those pumps also use NdFeB.
The production numbers tell the story: global NdFeB magnet demand was approximately 220,000 tonnes in 2024. AI-driven data center construction is projected to add meaningful incremental demand through 2030.
Dysprosium: The Unseen Bottleneck
Dysprosium is added to NdFeB magnets in proportions of 2–5% by weight to prevent demagnetization at elevated temperatures. Without dysprosium, an NdFeB magnet begins to lose its magnetization above approximately 80°C — a temperature routinely reached in motor applications under load.
The supply situation for dysprosium is more concentrated than for neodymium. While neodymium is produced from multiple ore deposits in China, Australia, and the United States, dysprosium is predominantly found in heavy rare earth deposits, which are concentrated in southern China's ionic clay ore deposits. China controls over 90% of global dysprosium supply.
In 2023, China tightened export controls on rare earth processing technology — a precursor control that prevents other countries from building independent refining capacity even if they can mine rare earth ores. This followed the 2010 playbook, when China temporarily restricted rare earth exports to Japan during a diplomatic dispute, causing prices for certain rare earths to increase 2,000% within 18 months before moderating.
The 2023 technology controls are considered more durable than direct export quotas because they constrain the entire supply chain, not just a specific product flow.
What the United States Is Doing
The United States mined zero rare earths domestically from 2002 to 2017. The Mountain Pass mine in California — the only significant U.S. rare earth mine — was closed during that period due to competition from lower-cost Chinese production and environmental permit challenges.
Mountain Pass reopened in 2017 under MP Materials, which has since become the largest rare earth producer outside China. In 2024, MP Materials opened a rare earth magnet manufacturing facility in Fort Worth, Texas — the first NdFeB magnet plant in the United States in decades. The facility has initial capacity to produce magnets for approximately 500,000 electric vehicle traction motors per year.
The Department of Defense has provided over $200 million in funding to MP Materials and related rare earth processing companies under the Defense Production Act. The goal is to establish a complete domestic supply chain — from mining to magnet production — that is independent of Chinese supply.
The challenge is timeline. Building the hydrometallurgical processing capacity to refine heavy rare earths, including dysprosium, takes 7–10 years from funding commitment to full production. The Mountain Pass mine produces light rare earths (neodymium, praseodymium, lanthanum, cerium) but has limited heavy rare earth content. Domestic dysprosium production remains essentially zero.
The Australia Factor
Australia holds the world's second-largest identified rare earth reserves after China. Lynas Rare Earths, an Australian company, is the largest rare earth producer outside China and operates the Mount Weld mine in Western Australia — one of the highest-grade rare earth deposits in the world.
Lynas processes its ore at a facility in Malaysia and has been building a separation plant in Kalgoorlie, Western Australia, to bring more processing onshore. The company also operates a light rare earth separation facility in Texas under a U.S. Department of Defense contract.
However, Lynas has limited heavy rare earth production. The Mount Weld deposit is light-rare-earth-dominant, meaning it produces neodymium and praseodymium at scale but not dysprosium or terbium in meaningful quantities.
The heavy rare earth supply problem — the bottleneck most critical for temperature-stable motor magnets — remains unresolved outside China.
The AI Industry's Exposure
The rare earth supply chain is a second-order risk for AI companies that most have not explicitly addressed in public disclosures. Nvidia, Google, Microsoft, and Amazon do not source rare earths directly — they purchase finished chips and servers from manufacturers who purchase magnets from magnet producers who purchase refined oxides from processors who purchase ore from mines.
The opacity of this supply chain means that supply disruptions propagate slowly and are often not identified until they manifest as component shortages or price increases in finished goods — at which point the lead time to find alternatives is measured in years, not months.
The price signal to watch is the spot price of dysprosium oxide, published daily by Asian Metal and Argus Media. As of early 2026, dysprosium oxide trades at approximately $280–$320 per kilogram, elevated relative to 2021 lows but well below the 2011 peak of over $2,000 per kilogram during the last major supply disruption. The next disruption, if it comes, will look different from 2011 — but the underlying supply concentration that made 2011 possible has not changed.