Sustainability
Every model in production depends on physical resources that rarely make it into the conversation — power, water, materials, and the hardware itself. A high-level look at what it actually takes to keep this running, and what that costs.
Power
A single large AI data center campus can draw several hundred megawatts to multiple gigawatts of continuous power at the largest sites — comparable to a mid-sized city, or more. Global electricity demand from data centers is on track to more than double by 2030, to roughly 945 terawatt-hours, according to the International Energy Agency (about 485 terawatt-hours in 2025, up 17% in a year). Utilities are responding with new generation built specifically for AI campuses and, in some cases, by delaying coal plant retirements to keep the grid stable. For how that plays out at a single site, see Power comes first.
Water
Cooling a data center takes water — a large hyperscale campus can use up to 5 million gallons a day, according to the Environmental and Energy Study Institute. Closed-loop liquid cooling reduces this on-site draw compared to traditional evaporative cooling towers, but it doesn't eliminate the water footprint: generating the electricity itself carries an indirect water cost, estimated at around 211 billion gallons in 2023 for U.S. data centers alone, since much of that power still comes from thermoelectric plants that use water to cool their own generators. Most data centers get their water from local municipal water utilities or county water authorities — not investor-owned, publicly traded companies.
Cabling and connectivity
Inside and between data centers, cabling does two different jobs that call for different materials. Copper handles power delivery and the short, extremely high-speed links between GPUs and servers within a rack — a hyperscale AI campus can use up to 50,000 tons of copper for this alone, versus 5,000 to 15,000 tons for a conventional data center, according to the Copper Development Association. Fiber optic cable handles the long-distance, high-bandwidth links between data centers and into the broader internet backbone; a 2025 Fiber Broadband Association study estimated the U.S. will need roughly 214 million additional fiber miles by 2029 just to connect the new wave of AI data centers. Rising demand from AI buildout is now a recognized driver of a global “copper supercycle”: S&P Global estimates data centers used about 1.1 million metric tons of copper in 2025 and could need about 2.5 million metric tons a year by 2040.
Materials
Every chip in this supply chain depends on mined materials — silicon, copper, gold for connectors, and rare earth elements used in some components — extracted, refined, and shipped globally before a single transistor is etched. Each campus also takes up hundreds of acres of land; why that land ends up where it does is covered in How AI Data Centers Get Built.
Hardware lifecycle and e-waste
GPUs are typically replaced every two to four years as new, more efficient generations arrive, meaning the physical hardware — and everything it took to make it — gets retired quickly. Globally, only about 22% of e-waste is formally collected and recycled, according to the 2024 Global E-waste Monitor from the UN and ITU; the rest is often exported, landfilled, or informally processed in ways that recover far less of the valuable material inside. A small industry of companies now specializes in securely decommissioning retired data center hardware and recovering the components and metals inside before they're lost to landfill.
Read next How AI Data Centers Get Built — where these campuses go, who owns them, and why.