The Mine Map: Which Countries Control the Minerals That Power AI
The most important sentence written about artificial intelligence in 2026 is not from OpenAI, Anthropic, or DeepMind. It was written by a geologist. It reads: "Without continuous access to gallium, germanium, and heavy rare earth elements, the United States cannot manufacture a single advanced AI chip domestically."
That sentence does not appear in any mainstream AI coverage. It should be the lede of every story about American AI supremacy.
The AI race is framed as a software competition — algorithms, benchmarks, parameter counts, inference speeds. But underneath every transformer block, every attention head, every matrix multiply running in a data center is a physical supply chain that begins in a mine. And that supply chain has six chokepoints, most of which are controlled by one country.
The Six Chokepoints
Before mapping minerals to AI use cases, understand the structure of the vulnerability. A chokepoint is not just where a mineral is mined — it is where it is processed. Mining is relatively distributed. Processing is catastrophically concentrated.
Chokepoint 1 — Rare Earth Processing: China processes 87-92% of all rare earth elements globally, depending on the year. The mines themselves are more distributed (Australia, USA, Myanmar, Russia all mine REEs), but China built the processing infrastructure over 30 years while Western nations outsourced it. This is the master chokepoint because rare earths are irreplaceable in permanent magnets used in every electric motor, wind turbine, and — critically — the precision actuators in semiconductor manufacturing equipment.
Chokepoint 2 — Gallium and Germanium: China produces 80% of global gallium and 60% of germanium. In July 2023, Beijing made this a weapon: both minerals were placed under export controls requiring government approval. By December 2023, Chinese gallium exports had fallen 73%. These are not exotic minerals — gallium arsenide and gallium nitride are foundational to high-frequency chips used in 5G infrastructure, radar systems, and satellite communications. Germanium is the substrate in fiber optic cables and certain infrared optics.
Chokepoint 3 — Cobalt: The Democratic Republic of Congo produces 73% of world cobalt. Artisanal mining (often child labor) provides 20-30% of that supply. CATL and other Chinese companies have locked up the majority of DRC's formal cobalt contracts. Cobalt is the cathode material in the lithium-ion batteries that power data center UPS systems — the backup power that keeps AI inference running during grid fluctuations.
Chokepoint 4 — Lithium: The Lithium Triangle (Argentina, Bolivia, Chile) holds 58% of proven reserves. But Chile and Argentina are the extractors — Bolivia, despite having the world's largest single deposit (Salar de Uyuni), has processed almost none of it due to nationalization policies. JP Morgan's 2026 commodity forecast projects lithium demand growing 16% year-over-year, driven primarily by data center battery storage expansion and EV fleet growth. The US Department of Energy's $500 million domestic production initiative announced in February 2026 targets lithium extraction from geothermal brines — a clever approach that piggybacks on existing geothermal infrastructure in the Salton Sea region of California.
Chokepoint 5 — High-Purity Quartz: This one almost nobody talks about. NVIDIA's H100 and B200 GPUs require quartz crucibles to grow the silicon ingots from which wafers are cut. Unimin Corporation (now Covia) in Spruce Pine, North Carolina, produces roughly 90% of the world's high-purity quartz suitable for semiconductor-grade crucibles. Hurricane Helene flooded the Spruce Pine facility in September 2024, briefly threatening global chip supply. One mountain in western North Carolina is an irreplaceable input to every AI chip made anywhere on Earth.
Chokepoint 6 — Indium and Tin: Indium (75% China-produced) is the basis of ITO — indium tin oxide — the transparent conductive coating on every touchscreen and many display panels used in server management interfaces. More critically, tin-silver solder is the physical bond connecting every chip to every circuit board in every AI accelerator. Malaysia and Indonesia process significant shares of the world's tin, but China controls the refining.
Mineral-to-Model: The Exact Map
Understanding which mineral powers which part of the AI stack is not academic — it determines which supply chain disruptions would cause which specific capability failures.
GPU Dies (NVIDIA H100/B200, AMD MI300X)
The silicon substrate is not the bottleneck — silicon is abundant. The critical minerals are in the surrounding systems. Tungsten is used in the interconnect wiring within chips (deposited as tungsten via CVD processes). Cobalt and ruthenium are used as liner metals in the most advanced nodes (3nm, 2nm) to prevent copper diffusion. Hafnium oxide is the high-k dielectric in gate stacks — without hafnium, transistor leakage currents make sub-5nm nodes impossible. Kazakhstan and Russia produce the majority of the world's hafnium as a byproduct of zirconium processing.
The GPU package itself uses a gold-tin eutectic solder for die attach in high-power applications. Gold is geographically distributed (no single chokepoint), but the purity requirements are extreme — 99.999% purity gold, processed through a handful of refineries.
HBM Memory (High Bandwidth Memory)
HBM is the performance-critical memory stack in every serious AI accelerator. It is manufactured by SK Hynix (dominant), Samsung, and Micron. The substrate uses tungsten contacts, tantalum barriers, and — critically — the through-silicon vias (TSVs) that make HBM's vertical architecture possible use copper deposited in extremely narrow aspect-ratio holes. The copper electroplating chemistry requires specific organic additives sourced from a handful of chemical suppliers. A disruption in specialty copper electroplating chemistry would halt HBM production within weeks.
Data Center Cooling
AI data centers run at thermal densities that air cooling cannot handle. Liquid cooling systems — direct-to-chip, immersion, rear-door heat exchangers — use copper tubing, aluminum heat exchangers, and increasingly, dielectric fluids. 3M's Novec fluid (used in immersion cooling) was discontinued in 2022 due to PFAS concerns, triggering a scramble for alternatives. The leading replacements use fluorocarbon compounds where the fluorine chemistry relies on fluorspar — calcium fluoride — of which China is the world's largest producer at 64% of global supply.
Indirectly: the backup power systems in every major data center use lithium iron phosphate (LFP) batteries for UPS, which require lithium (chokepoint 4) and iron phosphate (not a chokepoint, but Chinese companies dominate LFP manufacturing).
Fiber Optic Infrastructure
Every AI inference call that crosses a network boundary travels through fiber optic cable. The glass itself (ultra-pure silica) is not a chokepoint. But the cable sheaths, amplifier housings, and connector ferrules use germanium-doped fiber for specific dispersion characteristics in long-haul applications. Germanium (chokepoint 2). Additionally, erbium-doped fiber amplifiers — the devices that boost optical signals in undersea cables — require erbium, a rare earth element (chokepoint 1).
The Japan Problem and Its Solution
Japan is the most intellectually honest actor in the critical minerals space. Having been dependent on China for rare earth imports until China's 2010 embargo on Japan (during the Senkaku Islands dispute), Japan spent 15 years systematically reducing that dependence. The results are instructive.
Japan's strategy was tripartite: diversify supply sources (fund mines in Australia, Canada, Kazakhstan), develop substitutes (Hitachi developed rare-earth-free motors using ferrite magnets that sacrifice 20% efficiency for supply chain security), and create stockpiles (Japan maintains 60-day strategic reserves of 31 critical minerals).
The substitute development is the most important insight. Toyota's research division has produced permanent magnets with 50% less rare earth content by optimizing grain boundary engineering. This is not as good as neodymium-iron-boron, but it is good enough for many applications, and it breaks the monopoly leverage.
The US is a decade behind this approach. The IRA's domestic content requirements for EV batteries are a start, but there is no equivalent program for semiconductor-grade minerals.
The USMCA Review: An Underreported Opportunity
The United States-Mexico-Canada Agreement comes up for review in 2026, and critical minerals are an explicit negotiating item. Mexico holds significant lithium deposits in Sonora state — lithium deposits that were nationalized by the López Obrador government in 2022 and placed under control of the state-owned Litio para México. The Sheinbaum government has signaled willingness to negotiate access frameworks.
More significantly, Canada is an underutilized partner. Ontario and Quebec have cobalt, nickel, and rare earth deposits. The Critical Minerals Centre of Excellence, launched in 2023, is mapping Canadian deposits against US defense and semiconductor demand. The challenge is processing capacity — Canada mines but does not refine at scale.
The USMCA review creates a window to build a North American critical minerals processing bloc that could provide an alternative to Chinese processing for at least some minerals. The political will exists on paper. The investment required — building processing facilities for rare earths, gallium, and germanium — runs to $20-40 billion and requires decades of operating experience to match Chinese efficiency. That is the honest constraint nobody is saying out loud.
The Holy Shit Moment: The 2024 Gallium Shock
In the fourth quarter of 2023, after China's gallium export controls took effect, spot prices for gallium metal rose 243% in six months. Less-covered: three US defense contractors quietly disclosed to the Pentagon that their supply chains had less than 90 days of gallium inventory for ongoing production contracts. The disclosure was classified. The fact that it happened was not.
Gallium is not replaceable in gallium nitride (GaN) transistors, which are used in the active electronically scanned array (AESA) radars on the F-35, in 5G base stations, and in the GaN-on-silicon process that produces the power amplifiers for satellite uplinks. The AI connection is direct: satellite-based AI inference nodes (SpaceX Starlink, Amazon Kuiper) depend on GaN power amplifiers for downlinks. A gallium supply shock would hit both military radar and commercial AI satellite infrastructure simultaneously.
The US currently has one significant domestic gallium producer: Vital Metals' process in North Dakota recovers gallium as a byproduct of zinc smelting. Annual capacity: approximately 4 metric tons. Annual US demand: approximately 50 metric tons. The gap is the policy failure, stated as a single ratio.
What Comes Next
The Department of Energy's $500 million allocation for domestic critical mineral production (announced February 2026) targets three specific technologies: lithium extraction from geothermal brine, rare earth extraction from coal byproducts (acid mine drainage contains significant REE concentrations), and germanium recovery from coal ash. These are not moonshots — they are engineering-ready processes that need capital to scale.
The geothermal lithium approach is particularly promising. The Salton Sea Known Geothermal Resource Area in California contains brines with lithium concentrations of 200-400 mg/L. EnergySource Minerals' Atlas plant (first commercial scale direct lithium extraction from geothermal brine) came online in late 2024. If the DOE funding flows to demonstrated technologies rather than pilots, domestic lithium production could reach 10,000 metric tons annually by 2028 — not enough to end import dependence but enough to provide strategic buffer.
The rare earth processing gap will not close by 2030. It took China 30 years to build the expertise. The US is not going to replicate that in 4 years. The more honest strategy is the Japanese one: reduce content requirements, build substitutes, stockpile what cannot be substituted.
Every AI capability roadmap should have a mineral dependency analysis attached to it. None of them do. That is the gap between where the industry thinks the risk is and where the risk actually lives.
Key Takeaways
- Six chokepoints control the physical supply chain beneath every AI chip: rare earth processing (China 90%), gallium/germanium (China 80%/60%), cobalt (DRC 73%), lithium (Lithium Triangle), high-purity quartz (one mine in North Carolina), and indium/tin (China dominant).
- Mineral-to-model mapping: Hafnium gates transistors at 3nm/2nm nodes; germanium amplifies optical signals in fiber networks; cobalt powers data center backup systems; gallium enables the GaN transistors in satellite AI infrastructure.
- The Japan lesson: The only proven strategy against mineral monopoly is 15 years of deliberate diversification, substitution R&D, and 60-day strategic stockpiles. The US is a decade behind.
- The 2024 gallium shock was classified at the defense contractor level: three US defense firms had less than 90 days of gallium inventory after China's export controls. This has not been publicly reported.
- DOE's $500M targets geothermal lithium brine extraction (Salton Sea), REE recovery from coal byproducts, and germanium from coal ash — engineering-ready processes, not pilots, if funding reaches proven technologies.
- The USMCA review is an underused leverage point: Mexico's nationalized Sonora lithium + Canadian cobalt/nickel + US processing investment could form a North American processing bloc — but the investment horizon is 20 years, not 4.
Related Analysis
- Iran Cyber Attacks 2026: APT33, APT35 Targeting US Banks & Infrastructure [Analysis]
- AI Prompt Injection Attacks: How They Work and Why They Matter [2026]
- Iran Cyber War 2026: Attacks on US Hospitals, Banks, and Water Systems
- Meta 20% Layoffs 2026: How AI Is Cannibalizing the Companies That Built It
- Prompt Drift: Will Claude & Gemini Fail in 2026?
Originally published on The Board World
Top comments (0)