What an AI data center is made of
Accelerators
Nvidia GPUs dominate; AMD, Google TPUs, and custom silicon from hyperscalers compete. All fabbed at TSMC.
HBM memory
High-bandwidth memory (SK Hynix, Samsung, Micron) is the scarcest input — stacked DRAM that feeds the GPUs.
Networking
Optical interconnect, switches, and copper cabling tie thousands of accelerators into one machine.
Power & grid
The binding constraint is now electricity — transformers, switchgear, and generation. Copper-intensive.
Cooling
Liquid cooling, pumps, and heat exchangers — increasingly the difference between a site that scales and one that can't.
Backup & storage
Battery storage and UPS smooth the enormous, spiky loads AI clusters place on the grid.
Why the materials matter
AI inherits the entire semiconductor supply chain — so gallium, germanium, silicon, and the EUV chokepoint all apply. On top of that, the data-center build-out is a massive new draw on copper (wiring, busbars, transformers), rare earths (generator and cooling-fan magnets), and grid metals. Power, not chips, may become the real ceiling on scaling.
The chokepoints
- HBM & advanced packaging — CoWoS capacity and HBM supply gate GPU output.
- Leading-edge fab — same TSMC/Taiwan dependency as all advanced chips.
- Electricity & transformers — multi-year lead times on grid equipment.
- Copper — structurally tight as electrification and AI demand collide.
Critical materials powering AI compute & power (…)
The chips, the interconnect, and the electricity behind them. Click any element for its full profile. Explore all 68 →