Critical Minerals and Components for AI

Artificial intelligence looks weightless from the outside, a model answering a question in a browser, yet every answer runs on a vast physical machine. Training and serving large models takes racks of accelerators, high-bandwidth memory and optical networking, housed in data centres that draw as much power as a small city and need industrial cooling to stay online. Each layer of that machine is built from critical minerals.

At the heart of the stack are the chips, made on silicon wafers and wired with copper, tantalum, cobalt and tungsten, with hafnium in their transistors. Around them sit the components that move data and power: indium phosphide and germanium in optical links, gallium nitride in power supplies, copper and aluminum in busbars, cables and cold plates. Behind the building stands the grid, the transformers and, increasingly, dedicated generation, from gas turbines built on nickel superalloys to nuclear plants fuelled by uranium.

That is why the race towards ever larger models, and the talk of superintelligence that comes with it, is also a race for materials. Stockatomics maps 48 critical minerals to AI, from the familiar to the obscure.

AI needs critical minerals at three levels: the chips that do the computing, the data centre that houses and cools them, and the power system that keeps them running around the clock. Each level has its own bottlenecks, and a shortage at any one of them can hold back the others.

Chips and memory

Accelerators and memory are made on silicon wafers, yet the interesting materials sit on top. Hafnium oxide insulates the gates of the transistors, tungsten and cobalt fill the tiniest contacts, tantalum lines the copper wiring, and ruthenium is emerging for the smallest interconnects. Tin, silver and bismuth solders join the chips to their packages, while palladium and nickel plate the contacts.

Networking

Training a frontier model means moving huge volumes of data between thousands of chips. The optical transceivers that do it use indium phosphide lasers and germanium photodetectors, and long-distance links between data centres rely on erbium-doped fibre amplifiers.

Power and cooling

Inside the building, copper and aluminum carry power through busbars and cables, gallium nitride and silicon carbide chips convert it efficiently, and backup batteries built on lithium or lead take over when the grid falters. Liquid cooling, now standard for the densest racks, relies on copper cold plates and heat exchangers.

Generation

Data centres increasingly arrange their own power. Gas turbines use nickel, cobalt and rhenium superalloys, while nuclear plants and small modular reactors depend on uranium, zirconium and hafnium.

The minerals inside AI hardware are among the most concentrated in the world. China produces nearly all primary gallium and most germanium, and since 2023 it has restricted exports of both, later banning shipments to the United States. It also dominates the refining of rare earths used in hard drives and cooling fans, and has extended its controls to tungsten, indium and other minor metals used in chips.

The physical build-out of data centres faces a different kind of bottleneck. Copper supply is constrained by slow mine development, and transformers and grid connections have become some of the longest-lead items in a data centre project, with waiting times measured in years. Power itself is now a strategic resource, which is why technology companies are signing long-term deals with nuclear and gas generators.

What that means for AI

• Chip and equipment makers outside China are building stockpiles and qualifying new suppliers of gallium and germanium
• Data centre developers increasingly secure power, transformers and cabling years in advance
• The cost of the physical layer, from copper to cooling, is rising alongside the cost of the chips

As of October 2026, the build-out of AI infrastructure is the fastest-moving source of new demand for several critical minerals. The largest technology companies are spending at a historic pace on data centres, and the ambition behind it, from ever larger frontier models to the pursuit of superintelligence, shows little sign of slowing.

What to watch

• Power first: grid connections and generation have become the limiting factor for new data centres, pulling forward demand for copper, aluminum, transformers and on-site power, including nuclear
• Higher voltages: the shift to higher-voltage power distribution inside data centres favours gallium nitride and silicon carbide power chips
• Optics everywhere: as clusters grow, more of the data moves as light, raising demand for indium phosphide and germanium components
• Liquid cooling: dense racks are moving from air to liquid cooling, adding copper and other materials to every rack
• Export controls: US limits on AI chip exports and Chinese limits on gallium, germanium and other minerals are pushing both sides to build separate supply chains