AI infrastructure could generate between 395 million and 617 million tonnes of electronic waste between 2025 and 2050, according to a new report from the Basel Action Network (BAN). That volume would fill 15 million to 23 million shipping containers; placed end to end, 20 million of them would stretch roughly 244,000 km — about six times Earth's circumference.

The 46-page paper, titled "How Big Is the AI Waste Wave?", argues that the most widely cited academic projections drastically understate the coming scrap stream. Those estimates focus chiefly on servers and accelerators such as GPUs, which account for just 13 percent of a datacenter's equipment by weight. Once power distribution, networking, cooling and other infrastructure enter the calculation, BAN puts the total at 40 to 60 times higher than the standard projections.

The material breakdown

BAN's model identifies five equipment categories: networking, power distribution, backup power systems, servers plus accelerators, and cooling. Together they total about 7,000 metric tons for a reference 100 MW AI datacenter — much of it destined for replacement when the facility upgrades.

The copper intensity alone is striking. A single 100 MW facility contains about 2,700 tonnes of copper in its cabling and busbars, the report finds. When an operator replaces racks drawing 5–15 kW with the 50–140 kW racks that AI workloads require, much of the existing power infrastructure must come out and be replaced.

Replacement cycles drive the model. GPUs and other AI accelerators turn over every two to three years, BAN claims, against five to seven years for traditional general-purpose servers. The model assumes owners scrap entire servers when moving to the latest generation of Nvidia GPUs — plausible, BAN argues, because OEMs supplying hyperscalers tend to ship complete, preconfigured systems. Networking infrastructure cycles in three to four years, power distribution in eight, and backup power and cooling systems in five.

Built on stated buildout intentions

BAN's model starts from what the industry says it intends to build and calculates the waste implied by that expansion. A central assumption is that the copper, steel, cooling and power distribution equipment, networking gear and computing hardware in current-generation AI datacenters will be decommissioned within years rather than decades as the technology becomes obsolete.

The scale of the projected stream exposes a governance gap. No hyperscaler, government or international body has published a plan for handling AI-driven e-waste at the projected scale, the report states. It adds that processing infrastructure is already insufficient for current volumes.

"To date the environmental debate around AI has focused on electricity, carbon and water while largely overlooking what happens to the hardware itself," said BAN founder and chief of strategic direction Jim Puckett. "If companies and governments do not begin planning for this new waste tsunami, today's AI buildout could become an even more cataclysmic toxic waste crisis than we are already experiencing."

What comes next

The report is the first in a four-part series. Part two will examine how reuse, refurbishment and repurposing could cut e-waste volumes. Part three will address potential toxicity, including contamination from perfluoroalkyl and polyfluoroalkyl substances (PFAS). Part four will cover risk mitigation.

BAN, a nonprofit named after the Basel Convention that governs international movements of hazardous waste, frames the findings as a call for planning rather than prediction. The numbers that matter for the recycling sector are the replacement cycles: a two-to-three-year accelerator turnover compressed against a 7,000-tonne material footprint per 100 MW facility defines the inbound tonnage the industry will have to permit, sort and process. Whether regulators extend e-waste frameworks to datacenter decommissioning streams — and whether any hyperscaler publishes a takeback or recycling plan — is the milestone that will determine whether the projected wave becomes managed material flow or uncontrolled scrap export.