Between 395 million and 617 million metric tons of electronic equipment could be retired between 2025 and 2050 in connection with artificial intelligence, according to a new report from the Basel Action Network (BAN). That range is triple the current volume of equipment in service, and it lands well above the NGO's earlier projections for the AI-driven e-scrap stream.

The report, "The Coming AI Waste Wave, Part 1," differs from previous estimates because it counts the entire ecosystem of material tied to data center buildout — not just servers. By mass, BAN breaks a data center's electro-mechanical infrastructure down as follows: servers, accelerators and racks account for 13%; power supply and distribution for 34%; cooling systems for 35%; backup power systems for 15%; and networking equipment for 3%.

For context, the UN Global E-Waste Monitor estimated global generation at 62 million metric tons in 2022, with only 22.3% properly recycled. Scaled on that trajectory, annual generation would reach roughly 165 million metric tons by 2050 — before adding any AI-driven demand. BAN now puts the 2050 ceiling at up to 211 million metric tons per year, 28% above its own previous estimate and triple today's annual total.

The revision comes from broadening the system boundary. BAN's model takes three inputs: how fast compute capacity is growing, how much equipment each added gigawatt requires, and how long that equipment lasts. It counts networking gear, power distribution, data storage and backup, and cooling infrastructure alongside the compute hardware itself.

The 'contagion' factor

The report's most novel category is what BAN calls "AI waste contagion" — electronics outside data centers retired early because personal devices and telecommunications networks need upgrades to work with AI.

"This AI contagion obsolescence as we call it, will not take place in data centers but with our own IT equipment in our homes, our businesses, and our gaming rooms," the report states.

BAN projects that contagion alone could add between 16.2 million and 30.7 million metric tons of retired equipment annually to the global e-scrap total by 2050.

Upside scenarios

The numbers could climb higher still. As more countries pass data sovereignty laws, operators may need to build redundant capacity, especially if data centers become critical infrastructure and conflict targets. A wholesale replacement of aged telecom infrastructure would also lift the tonnage. And a "rip-and-replace" transition — data centers switching to superconducting power systems, liquid cooling and new battery chemistries — could force early retirement of otherwise functional equipment.

BAN acknowledges countervailing forces too. Equipment may outlast projected lifespans. Servers could be modularly upgraded rather than swapped. A secondary market for reuse and refurbishment may emerge. Software efficiency gains could cut hardware demand, and AI-related data center buildout could plateau before reaching BAN's 2050 levels.

The advocacy aim

BAN founder Jim Puckett framed the white paper series, of which three further releases are planned, as an attempt to slow AI expansion, not halt it.

"I earnestly hope we can slow the horserace, the freight train, the juggernaut that is the 'AI revolution' down," Puckett said by email. "From what I have learned in just the last few months, and only looking at this one segment of its impact — toxic e-waste — I am scared for that outcome, while realizing there are so many other results far more harrowing."

He conceded the irony that he used Anthropic's Claude model to help draft the white paper. "Given the speed at which things are moving with the AI horserace, I felt I could not afford to worry about this moment's ironies in order to match that sense of urgency," he said, adding that he is "harnessing the power of AI in the most positive way now in the hope of being able to rein in its likelihood of it creating havoc."

Puckett noted the same tension confronts the sector's largest players: "It's the same irony even the masters of AI are facing down when they remarkably beg governments to control their product."

The remaining installments of the series will assess whether reuse, refurbishment and repurposing can blunt the projected volumes, characterize the toxicity of the AI-related stream, and lay out BAN's recommended policy responses. For processors and regulators tracking this material, the near-term milestones are those follow-up releases and whether demand-side signals — data center buildout rates, equipment lifespans and modular upgrade adoption — confirm or undercut BAN's tonnage bands.