Study: copper is ~83% of AI data-center mineral demand
A new demand model confirms what grid engineers already suspected: AI's mineral footprint is mostly copper, and mostly outside the server hall.
What happened
Researchers published a bottom-up model of the minerals needed to build AI data centers from 2025 to 2035, covering 20 materials across computing, cooling and power systems. Copper makes up roughly 82–83% of the modeled mineral mass, and nearly two-thirds of that copper is tied to grid transmission.
The study also flags grain-oriented electrical steel for transformers as a potential constraint, and finds that for gallium, germanium, graphite and rare earths the risk lies in processing concentration rather than geology.
Our take
This is a useful corrective to the chip-centric way AI supply chains are usually discussed. The GPUs get the attention, but the tonnage — and a lot of the scheduling risk — sits in substations, transformers and transmission lines.
It also separates two kinds of risk that often get blurred together. Copper and transformer steel are volume problems: there may simply not be enough capacity. Gallium and germanium are concentration problems: there's enough in the ground, but one country controls the processing.
Why it matters for AI infrastructure
Transformer lead times already stretch for years in many markets. If electrical steel becomes the bottleneck, it could delay data-center energization regardless of how many chips are available.
What we're watching
- Transformer and electrical-steel capacity announcements
- Utility interconnection queues for large AI campuses
- Follow-up studies on optical-network mineral demand