Basel Action Network Warns AI Hardware Could Create 617 Million Tonnes of E-Waste

The AI industry’s next environmental problem may begin when its hardware is retired. The Basel Action Network estimates that AI-related equipment could create 395 million to 617 million tonnes of e-waste between 2025 and 2050. That total includes not only accelerators and servers, but also the power, cooling and networking systems around them.

BAN’s newly published analysis argues that past estimates have focused too narrowly on the computing hardware most visible in the AI race. Servers and accelerators represent about 13% of the relevant equipment in a data center, the group says. The remaining 87% includes networking gear, power distribution, storage and backup systems, and cooling equipment.

The estimate starts with the equipment around the chips

That difference changes the scale. BAN estimates that each added gigawatt of data-center capacity needs about 70,000 tonnes of electronics. The group also expects 100 gigawatts of new capacity worldwide by the end of the decade. Much of that equipment will eventually need replacement, but many e-waste forecasts do not fully count it.

The container comparison illustrates BAN’s estimate; it is not a count of waste already produced. Even so, it shows how much physical equipment data-center expansion requires beyond the processors commonly used to represent the AI boom.

Faster replacement turns expansion into a disposal issue

The estimate assumes a shorter refresh cycle for AI infrastructure. BAN expects AI accelerators, servers and racks to be replaced every 2.5 years, compared with five to seven years for general-purpose servers. That is a forecast, not a count of actual retirements. In the model, faster hardware turnover adds to the effect of capacity growth.

The estimate comes as infrastructure spending is also expected to surge. Global data-center capital spending could reach $7 trillion from 2025 through 2030, including $4.3 trillion for electronic hardware. Those figures provide context, but they do not show that the long-range waste estimate will come true.

The proposed response is to keep equipment in use longer

• Design AI data-center equipment for longer service lives.
• Expand recycling practices for retired equipment.
• Increase second-use options for hardware leaving AI deployments.

The report does not assess whether AI-specific hardware can move into lower-performance roles after leaving AI deployments. That leaves a practical question around second use unresolved.