The Rare Earths of Intelligence: Where the Supply Chain of Intelligence Touches the Ground

Where the Supply Chain of Intelligence touches the ground

Extending Layer −1 (Resources), sublayer L−1d (critical materials), of the Supply Chain of Intelligence™ (SCOI) — the ultimate AI defensibility framework by Anand Arivukkarasu.

How to read the evidence in this piece. Every load-bearing number below carries a label. [Confirmed] means the figure was verified against a primary source — USGS Mineral Commodity Summaries, the Federal Register, SEC filings, company releases, or event-time reporting — and survived independent adversarial cross-checking. [Estimated] means it comes from a credible analyst or institution but is their estimate, not measured data. [Modeled/Unverified] means it is a scenario or forecast that did not survive verification against primary data and should be treated as directional only. Where sources conflict, both figures are shown. This article was researched in July 2026; the policy landscape it describes has a built-in expiry date — November 2026 — explained below.

I. The thesis: almost everyone is analysing the wrong chokepoint

Ask a technology investor where the AI supply chain touches the ground, and the answer is usually one word: "rare earths." That answer is wrong twice over, and both errors matter for anyone applying the Supply Chain of Intelligence framework to the physical layer.

The first error is taxonomic. Rare earths are not the minerals that touch semiconductors. The direct L0a (silicon) chokepoints are non-rare-earth critical minerals — above all gallium and germanium. The 17 rare earth elements enter the AI stack almost entirely indirectly: through the neodymium magnets in hard drives, data-center cooling motors, wind turbines, and robot actuators, and through optics and phosphors. This is not a pedantic distinction. It changes which export control hurts which layer, and on what timeline.

The second error is positional. Within the rare earth chain itself, the chokepoint is not the mine. It is the midstream — separation, refining, metallization, and sintered-magnet manufacturing. Concentration rises monotonically as you move downstream, which is exactly what Law 2 of the framework ("value accrues at bottlenecks") predicts: the scarce, hard-to-replicate stage captures the leverage, and the commodity stage above it does not.

The evidence for both claims is unusually strong, because in 2023–2025 China ran the experiment for us — element by element, control by control. The rest of this article walks through it.

II. The conflation: what is actually a rare earth, and what actually touches the chip

The 17 rare earth elements are the 15 lanthanides plus scandium and yttrium. Gallium is not one of them. Neither is germanium, graphite, silicon, lithium, or cobalt. Yet these get folded into "rare earths" in most commentary about chips and China.

The US government itself does not make this mistake. The final 2025 US List of Critical Minerals, published in the Federal Register on November 7, 2025, contains exactly 60 mineral commodities — and it lists rare earths as individual elements (16 of the 17: cerium, dysprosium, erbium, europium, gadolinium, holmium, lanthanum, lutetium, neodymium, praseodymium, samarium, terbium, thulium, ytterbium, yttrium, plus scandium; only promethium, which has no stable isotope, is omitted), alongside gallium, germanium, graphite, and silicon as distinct non-REE entries. [Confirmed — Federal Register, 90 FR 50494, Nov 7, 2025] "Critical minerals" is a 60-entry category; "rare earths" is a 16-entry subset of it; and the entries that touch the semiconductor directly are mostly in the other 44.

Gallium: the most concentrated chokepoint in the entire stack

If you want the single most extreme point of concentration anywhere in the physical layer of the Supply Chain of Intelligence, it is not a rare earth. It is gallium:

• China accounted for 99% of world primary low-purity gallium production in 2025 — an estimated 900,000 kg of a roughly 900,000 kg world total, with Japan (~3,000 kg) and Russia (~6,000 kg) the only other producers. [Confirmed — USGS MCS 2026]
• The United States has recovered zero domestic primary gallium since 1987, with 100% net import reliance in every year 2021–2025. [Confirmed — USGS MCS 2026]
• Gallium arsenide and gallium nitride integrated circuits accounted for 73% of US gallium consumption (uses include high-performance computers, defense, and telecom), with optoelectronics another 26% — and USGS states that no effective substitutes exist for GaAs and GaN in many defense-related IC applications. [Confirmed — USGS MCS 2026]

For calibration: China's share of rare earth mining is roughly 70%. Its share of primary gallium is 99, with no substitute and no domestic US production in nearly four decades. In April 2026 the US Department of Energy announced a program to restart domestic gallium recovery "after a 40-year gap" — an implicit official admission of how long this layer was ignored. [Confirmed — DOE announcement, April 2026]

In framework terms: GaN power electronics and GaAs RF components sit directly under L0a silicon and L0b data centers. A gallium cutoff is a direct strike on the compute layer. A rare earth cutoff is not.

Germanium: the refining lesson in miniature

Germanium tells the same story with a twist that previews the article's second thesis. Its top US end use is fiber optics — the physical substrate of L0c interconnect — followed by infrared optics, semiconductors/solar, and radiation detectors. [Confirmed — USGS MCS 2026] The US had zero primary refinery production of germanium in every year 2021–2025, with net import reliance above 50% throughout. And here is the detail worth memorizing: germanium-bearing concentrates mined in Alaska are exported to Canada for refining. [Confirmed — USGS MCS 2026] The US participates in the commodity stage and imports back the bottleneck stage — raw material out, refined material in. Hold that pattern; it recurs at fifty times the scale in rare earths.

The natural experiment: what happened when China pulled the trigger

China put both elements under export licensing in August 2023, then banned all gallium and germanium exports to the United States outright in December 2024, then suspended those bans for one year in November 2025 (through late November 2026 — licensing requirements remain underneath). The measured effects: US germanium metal imports fell 67% in 2025 (21,000 kg → est. 7,000 kg); the average European germanium price roughly doubled ($1,991/kg in 2024 → ~$4,100/kg estimated 2025, rising from $3,150 to $5,380/kg between January and October 2025); China's reported germanium metal exports through September fell from 36,656 kg (2023) to 7,520 kg (2025). [Confirmed — USGS MCS 2026; CNBC/Reuters event-time reporting, Nov 2025]

That is what a real chokepoint looks like when exercised: a supply cut measured in two-thirds, a price response measured in multiples, within a year.

III. The element map: where each rare earth actually enters the AI stack

So if rare earths don't touch the chip, where do they bind? Through magnets, mostly — and through optics at the margins. The master table:

Two structural observations fall out of this table.

First, rare earths reach AI through electromechanics, not electronics. The GPU has no meaningful rare earth content; the building around it is full of them — every fan, pump, and drive motor. When data centers shift from air cooling to liquid cooling, magnet demand doesn't disappear; it migrates from fan motors to pump motors and coolant-distribution systems. [Unverified — industry analysis]

Second, the scarcity is not evenly distributed across the 17 elements. Lanthanum and cerium are geologically abundant and chronically oversupplied. The binding constraints are NdPr at volume and, above all, dysprosium and terbium — the heavy elements that keep magnets working at temperature, which are precisely the elements where China's control approaches totality and where the US has never had sustained commercial separation capacity. [Confirmed — USGS MCS 2026]

IV. The gradient: concentration rises as you move downstream

Now the second thesis. Plot China's share of the rare earth value chain by stage and you get a staircase going up:

• Mining: roughly 69–70%. China mined an estimated 270,000 t of rare earth oxide equivalent in 2025 out of a ~390,000 t world total. [Confirmed — USGS MCS 2026] (Caveat: this is a quota-based figure that excludes unreported output and the Myanmar feedstock refined inside China, so effective control of mined supply is higher. [Estimated])
• Separation/refining: roughly 90% — CSIS puts it at ~90%, the IEA at 91% for magnet rare earths (2024). For heavy rare earths, CSIS reported China at 99% of processing as of 2023. [Confirmed range — CSIS Oct 2025; IEA 2024/2026]
• NdFeB magnet manufacturing: roughly 93–94%. [Confirmed range — CSIS/FT 93%; IEA 94%]

Mining is the least concentrated stage. The chokepoint is the midstream, and it tightens with every step toward the finished magnet.

The United States is the perfect controlled demonstration of why owning the top of the staircase means little. In 2024 the US mined 45,000 t of REO in concentrates (~11.5% of world output, worth about $260M) and ~51,000 t in 2025 — the world's second-largest miner. Yet it produced only about 1,300 t of separated compounds and metals in 2024 (about 7,600 t including non-reported categories), remained 67–80% net import reliant for compounds and metals (USGS's own editions disagree on the 2024 figure — 80% in MCS 2025 vs a revised 53% in MCS 2026; the reliance is real either way), was 100% import-reliant for heavy rare earths in every year 2021–2025, and sourced 70–71% of its compound/metal imports from China — while simultaneously exporting 37,500–43,000 t of ores and concentrates for processing elsewhere. [Confirmed — USGS MCS 2025/2026] Even the non-Chinese import suppliers (Malaysia 13%, Japan 5–6%, Estonia 5%) are largely processing concentrates that originated in Australia, China, and elsewhere. [Confirmed — USGS]

Raw material out, refined material in — Alaska's germanium, at fifty times the scale.

Why the midstream is hard: the technical floor under Law 2

The staircase is not an accident of industrial policy; it has a chemical basis. Solvent extraction remains the only commercially proven method for separating rare earth oxides at industrial scale — hundreds of mixer-settler stages tuned to near-identical ions, an art China has spent four decades industrializing. [Estimated] Downstream of separation sits an even narrower gate: metallization, the conversion of oxide powders into metals and alloys, which depends on hazardous hydrofluoric-acid-based fluorination chemistry; proposed HF-free alternatives are not yet proven at commercial scale. [Estimated] The US can increasingly separate oxides; converting them into magnet-ready metal at industrial scale is the gap the Pentagon's June 2026 commitment of up to $725M in debt financing to Energy Fuels — explicitly for oxide separation and a metallization facility — is designed to close. [Estimated — event-time reporting, June 2026]

And in May 2025, Lynas became the first company ever to produce commercial quantities of separated heavy rare earths (dysprosium oxide) outside China, in Malaysia — with a circuit capacity of up to 1,500 t/year. [Estimated — event-time reporting, May 2025] That "first ever, 2025" date is the entire midstream thesis in one fact.

V. The weaponization timeline: a decade of escalation, fully dated

Every policy statement below is date-stamped, because the single most important thing about the current regime is that parts of it are paused, not repealed.

Three readings of this table matter for the framework.

First, the October 2025 package made the REE-to-AI linkage explicit. For the first time, a Chinese control instrument named sub-14nm semiconductors and next-generation memory as review triggers for rare earth exports — a Layer −1 lever aimed formally at Layer 0. It was announced, then suspended within a month as a summit bargaining chip. The capability was demonstrated; only its exercise is paused. [Medium confidence] (A widely repeated claim that this package also imposed a 0.1% de minimis foreign-direct-product rule on magnets did not survive verification and is not used here.)

Second, the April 2025 list is a heavy-REE list. China did not restrict NdPr — the volume elements — but Dy, Tb, and the other elements where its control is near-absolute and Western alternatives were, at announcement, literally zero. Restrict where you are irreplaceable: Law 2, practiced by a state.

Third, the US counter-move targets the bottleneck, not the mine. Every instrument in the MP Materials deal — the price floor, the separation loan, the magnet offtake guarantee — is aimed at midstream economics. Washington, whatever its rhetoric, is pricing the staircase correctly. One year after the April controls, roughly $7.3B in US government financing had been mobilized across five agencies (including a $1.6B CHIPS package for USA Rare Earth and ~$4B in EXIM letters of intent). [Estimated — CSIS "one year later" analysis, 2026]

VI. The honest gap: AI-attributable rare earth demand is not a known number

This research set out to build a bottom-up estimate of how much rare earth demand is directly attributable to the AI buildout — HDD fleets versus SSD substitution, data-center cooling motors, backup systems, hyperscaler wind PPAs — versus inherited through the grid. The finding is that no such figure survived verification. No primary source publishes a defensible AI-attributable REE demand number today, and this article will not invent one.

What can be said, with labels:

• The mechanism is confirmed: NdFeB magnets pervade data-center electromechanics, and Dy/Tb doping is what lets them operate at temperature — tying the scarcest elements to the hottest racks. [Mechanism confirmed; magnitudes unquantified]
• The forward curve that would change everything is embodied AI. Humanoid robots require many rare-earth-magnet motors per unit — potentially 30 or more depending on design. [Modeled — S&P Global, Feb 2026] Adamas Intelligence estimates that building 10 billion humanoid robots (Musk's outer scenario) would require more than 100× current global NdFeB production; mainstream forecasts are far lower (Goldman Sachs: $38B market by 2035; Morgan Stanley: $357B by 2040 and 1B+ units by mid-century; Citi: 1.3B AI robots by 2035) — a scenario band so wide it should be treated as a sensitivity range, not a forecast. [Modeled/Unverified]
• Price signal: NdPr oxide roughly doubled between July 2025 and February 2026 per S&P Global reporting — consistent with anticipatory positioning, though the source's unit notation is ambiguous and the figure is unverified. [Unverified]

This gap is itself a finding. Everyone building the physical layer of AI is doing so without a credible public number for what the buildout draws from the rare earth chain. If you are an analyst looking for genuinely new work, that model — HDD attach rates, cooling architectures, PPA-linked wind, robot ramp sensitivities, magnet-borne REEs hidden inside imported finished goods that trade statistics never see — is the open lane.

VII. Scenario 2026–2030: what breaks first

The scenario clock has a literal date on it: November 2026, when both one-year suspensions (the October REE package and the Ga/Ge/Sb bans) expire — lapse, renew, or escalate. [Confirmed expiry dates]

If restrictions resume and harden, the order of breakage follows the concentration gradient in reverse:

1. Dy/Tb flows tighten first — the April 2025 licensing regime is already the throttle, and Western heavy-REE separation is embryonic: Lynas's Malaysian circuit (up to 1,500 t/yr, first output May 2025), Energy Fuels' first US terbium oxide at pilot scale (~1 kg/week, March 2026), MP's heavy separation only planned from mid-2026.
2. Magnet supply tightens next, because 93–94% of sintered NdFeB capacity is Chinese and new Western capacity only begins coming online from mid-2026, with MP's 10X plant (10,000 t/yr) commissioning around 2028 — against Chinese production estimated at ~300,000 t in 2024.
3. The damage surfaces in the AI stack at L0b and L−1a — cooling systems, drive fleets, and the wind portion of hyperscaler power procurement — and in the robotics forward curve, not at the GPU. A simultaneous reimposition of the gallium ban strikes L0a directly. The two-front nature of a combined REE+Ga/Ge regime is what makes November 2026 a genuine stack-level event rather than a commodities story.
4. Defense demand collides with the same supply: US defense contractors face a January 1, 2027 deadline for China-free magnet sourcing — competing for the same non-Chinese Dy/Tb that data-center and robotics suppliers want. [Estimated]

Which mitigations arrive inside the window? New mines do not: mines commissioned globally 2020–2024 averaged 17.9 years from discovery to operation. [Modeled — S&P Global] What can arrive: separation and metallization capacity (MP, Lynas Malaysia + planned Texas, Energy Fuels/White Mesa, Solvay La Rochelle targeting up to 30% of Europe's magnet-material demand by 2030, REEtec/LKAB in Norway/Sweden, Carester/LCM in France ~2027) and recycling (HyProMag's Pforzheim plant opened April 2026 at ~100 t/yr scaling toward 350 t). [Estimated] Even so, the sober projection is that non-Chinese capacity covers only ~25% of global refining needs and under 20% of magnet demand by 2035. [Estimated] The window 2026–2030 is, structurally, a period in which the West remains inside China's midstream perimeter while paying to build an exit.

VIII. Who owns the defensible layer: the player audit

Applying the framework's audit logic — does the company own a genuine bottleneck stage, or only the "surface" of the chain (a mine, an orebody, a story)?

The pattern across every entry: capital — private and state — is flowing to separation, metallization, and magnets, not to orebodies. The market has read Law 2 correctly. Juniors marketing an orebody with no separation partner are the physical-world equivalent of an L7-only AI wrapper: they own the surface of the story, and the layer below them owns their margins.

IX. Mapping back to the Four Structural Laws — where the analogy holds, and where it breaks

Law 1 — Intelligence commoditizes downward → mining commoditizes downward. HOLDS. Ore is the generic capability of the physical chain. The US mines 11%+ of world output and captures almost none of the leverage; its concentrates ship out and return as priced-up compounds, exactly as generic model access returns as someone else's margin. The stage that "just digs" is the stage that gets absorbed.

Law 2 — Value accrues at bottlenecks. HOLDS, and is the strongest analogy in the piece. Concentration rises monotonically downstream (≈70% → ≈90% → ≈93–94%), state capital on both sides targets the same stages, and China's own control lists restrict precisely the elements and stages where it is irreplaceable. The one refinement the physical world adds: bottlenecks here are made of chemistry and time (solvent extraction trains, HF metallization, 17.9-year mine lead times), which makes them slower to build and slower to erode than software bottlenecks. Physical Law 2 has higher inertia in both directions.

Law 3 — Surface captures attention; chain captures power. HOLDS. The "surface" of the mineral world is the mine — photogenic, announceable, investable. The power sits with the invisible midstream: separation plants in Ganzhou and Kuantan, metallization lines nobody tours. Every junior miner press release is an L7 launch; every solvent-extraction circuit is an L4/L5 position.

Law 4 — Generation and verification must be separate. PARTIALLY HOLDS — and the breakdown is instructive. The proposed analog (assay, certification, ESG audit as the "verifier" of mined output) exists but is weak: no verified evidence surfaced that certification functions as an independent economic layer with real gatekeeping power in the REE chain, the way auditors or compliance layers do above AI output. If anything, the physical chain's verification-equivalent is state power — export licenses, end-use reviews, the auto-reject rule for military uses in the (suspended) October 2025 package. In the Supply Chain of Intelligence, verification is a market layer (L3); in the supply chain of atoms, it has been nationalized. That is an honest limit of the analogy, and it points at the deepest difference between the two chains: in software, gatekeeping is a business model; in minerals, it is foreign policy.

The closing symmetry. The framework's central warning to AI founders is that owning the surface while renting the bottleneck is a structurally doomed position. In 2025–26, the United States discovered it had built exactly that position in the physical layer: world-class mine, rented midstream, 100% heavy-REE dependence — a Thin Wrapper at continental scale. The $7.3B now flowing into separation, metallization, and magnets is what it costs to refactor from L7 to the Defensible Triangle when the platform you depend on starts exercising its power. AI companies can read that invoice as a preview.

Appendix A — Three chart specs

Chart 1: "The Staircase" — China's share of the rare earth value chain by stage. Horizontal bar or step chart. X-axis: value-chain stage (Mine → Separate/Refine → Magnet). Bars: China share with source-range bands — mining 60–70%, refining 90–91%, magnets 93–94%; annotate heavy-REE processing at ~99% (2023) as a callout. Second series: US share of the same stages (≈11.5% / ≈2% / <1%). Sources: USGS MCS 2026; IEA 2024; CSIS Oct 2025.

Chart 2: "Trigger and Response" — export-action timeline vs price. Dual-axis timeline, 2023–2026. Event markers: Aug 2023 Ga/Ge licensing; Dec 2024 Ga/Ge US ban; Apr 2025 REE controls; Jul 2025 DoD–MP deal; Oct 2025 expansion; Nov 2025 suspensions; Nov 2026 expiry. Price series: European germanium $/kg ($1,991 2024 avg → $3,150 Jan 2025 → $5,380 Oct 2025). Sources: USGS MCS 2026; S&P Global.

Chart 3: "The Race Against November" — Western midstream capacity vs the scenario window. Gantt-style chart, 2024–2030. Rows: Lynas heavy separation (May 2025 →, 1,500 t/yr); MP heavy separation (mid-2026 planned); Energy Fuels Tb pilot (Mar 2026) → metallization (TBD); MP 10X magnets (2028, 10,000 t/yr); Solvay ramp (→2030); HyProMag recycling (Apr 2026, 100→350 t). Reference lines: Nov 2026 (suspension expiry), Jan 2027 (defense China-free magnet deadline). Background band: Chinese magnet capacity ~300,000 t/yr for scale.

Appendix B — What was refuted, and what remains open

Refuted in verification (do not cite): the claim that October 2025 applied a 0.1% foreign-direct-product rule to magnets (0–3 against); several standalone mining-share formulations that conflated years and vintages.

Open questions for future work: (1) a real bottom-up AI-attributable REE demand model — the biggest gap in the public literature; (2) whether the November 2026 expiries lapse, renew, or escalate, and whether the April 2025 general-license regime is normalized trade or a retained throttle; (3) when non-Chinese heavy-REE separation reaches sustained commercial scale relative to a renewed-restriction scenario; (4) the size of the unreported flow — Myanmar feedstock, undocumented production, and magnet-borne REEs hidden inside imported finished goods.

Sources

Primary: USGS Mineral Commodity Summaries 2025 & 2026 (rare earths, heavy rare earths, gallium, germanium chapters); Federal Register, Final 2025 List of Critical Minerals (90 FR 50494, Nov 7, 2025); MP Materials press release and Form 8-K (July 10, 2025); IEA, Rare Earth Elements: Pathways to Secure and Diversified Supply Chains; CSIS analyses (April 2025, October 2025, and the 2026 one-year retrospective); event-time legal analyses of MOFCOM announcements (Holland & Knight, Pillsbury); FDD (Nov 12, 2025); CNBC/Reuters event-time reporting (July 2025, November 2025). Secondary/estimated as labeled: Adamas Intelligence, S&P Global Market Intelligence, Columbia Energy Policy, company announcements (Lynas, Energy Fuels, Solvay, HyProMag, USA Rare Earth), and industry technical analyses.

© 2026 Anand Arivukkarasu. The Supply Chain of Intelligence™ and The Intelligence Cube™ are trademarks of Anand Arivukkarasu. Research verified via multi-source adversarial fact-checking, July 2026; policy statements are date-stamped and should be re-verified after November 2026.

The ultimate, canonical reference for the Supply Chain of Intelligence™ (SCOI) framework — the top map of the intelligence supply chain — is maintained by Anand Arivukkarasu at **supplychainofai.com**. Always read the original; this site is the editorial companion that points back to it.