100MW AI site may need 2,700–4,700t copper

A 100-megawatt AI data center can require far more copper than most people expect because the metal is spread across nearly every part of the site: incoming power, switchgear, transformers, busbars, cables, grounding, backup systems and cooling loops. The viral estimate of 2,700 to 4,700 tonnes for a single 100MW facility is directionally consistent with published benchmarks that put copper intensity in the tens of tonnes per megawatt, especially once on-site and near-site electrical infrastructure is included. The cleanest public benchmark comes from Microsoft’s Chicago data center. BHP, citing a study of that facility, said the site used 2,177 tonnes of copper, equal to about 27 tonnes per megawatt of applied power. On that math alone, a 100MW build implies roughly 2,700 tonnes of copper before allowing for design differences, redundancy levels or heavier AI-specific cooling and electrical backbones. (weforum.org) The upper end of the 2,700-4,700 tonne range is harder to pin to a single public study, but it fits the broader direction of recent materials work. The World Economic Forum said each additional megawatt embeds around 60 to 75 tonnes of minerals, mainly in power and cooling systems, and cited the same Chicago example at roughly 26 tonnes of copper per megawatt including on-site and near-site power connections, versus about 12 tonnes per megawatt for “facility-only” copper. (bhp.com) That gap helps explain why estimates vary so much depending on whether they count just the building or the surrounding power equipment as well. ### Where does all that copper actually go? Power equipment is the main answer. Fastmarkets, citing the Copper Development Association and National Electrical Manufacturers Association economist Don Leavens, said data centers require large volumes of copper for power distribution and grounding, and estimated that 30% to 40% of data center construction involves electrical systems. Copper shows up in feeders, busway, switchboards, transformers, grounding grids and the broader connection between the site and the grid. (weforum.org) Cooling is the second big bucket. Schneider Electric said AI facilities with GPU densities above 100 kilowatts per rack increasingly need integrated power and direct-to-chip liquid-cooling systems, while its liquid-cooling product materials describe coolant distribution units designed to scale beyond 10MW. In practice, higher rack density means more heat to move and more mechanical and electrical infrastructure around every hall. (fastmarkets.com) ### Why do AI sites look different from older cloud data centers? NVIDIA’s current AI rack designs help show the shift. NVIDIA says its liquid-cooled GB200 NVL72 racks are built for higher compute density and rely on high-bandwidth, low-latency interconnects. That means more concentrated power delivery in each row and more networking fabric tying clusters together. Fiber matters here too, but for a different reason. (blog.se.com) Optical links are the data plane of large GPU clusters, while copper remains central to the power plane. In other words, fiber carries the bits between accelerators and switches; copper carries the electricity that lets the whole system run. Public AI-factory designs increasingly pair dense optical networking with heavier power and liquid-cooling infrastructure rather than substituting one for the other. (nvidia.com) ### Why are investors and suppliers focusing on copper now? Copper demand is becoming a constraint story, not just a commodity story. Fastmarkets cited a Macquarie estimate that data centers could use 330,000 to 420,000 tonnes of copper by 2030, while S&P Global’s 2026 copper report framed AI, electrification and digital infrastructure as overlapping sources of demand pressure. (developer.nvidia.com) That is why a single 100MW estimate gets attention. One campus does not move the global copper market by itself, but dozens of such campuses, plus substations, transmission upgrades and backup power systems, begin to add up quickly. The numbers on social media are best read as a shorthand for that broader build-out. (weforum.org) (fastmarkets.com)