The AI data center buildout could generate between 395 and 617 million metric tons of electronic waste by 2050, according to a report published today by the Basel Action Network. That volume would fill 23 million shipping containers — enough 40-foot boxes to circle the planet six times end to end. The estimate is the most expansive to date because it counts the full stack of equipment inside a modern AI facility, not just the servers and GPUs.
BAN, a nonprofit focused on the international trade of hazardous waste, argues that previous AI e-waste forecasts missed roughly 87% of a data center's electro-mechanical infrastructure. The group's methodology adds power supply and distribution, cooling systems, backup power, and networking equipment to the count, alongside a category it calls AI Waste Contagion — telecom infrastructure and consumer devices likely to be replaced earlier as AI capabilities advance.
“AI may feel weightless, but every model depends on an enormous amount of highly specialized, cutting edge hardware.”— Jim Puckett, Founder and chief of strategic direction, Basel Action Network
The math produces a headline figure of 70,000 metric tons of e-waste per gigawatt of data center capacity. Applied to a McKinsey projection of up to 219GW of total data center capacity by 2030, and extended out to 2050, the annual retirement rate lands between 8.6 million and 13.1 million metric tons of AI-linked equipment per year.
Key facts
- 01Basel Action Network projects 395 to 617 million metric tons of AI-related e-waste accumulated between 2025 and 2050.
- 02The report attributes 15 to 20 percent of global e-waste by 2050 to AI, out of a projected 211 million metric tons per year.
- 03BAN estimates 70,000 metric tons of e-waste per gigawatt of data center capacity, against a McKinsey forecast of 219GW by 2030.
- 04Prior AI e-waste studies missed roughly 87% of data center electro-mechanical infrastructure by focusing only on servers and GPUs.
- 05Less than a quarter of the 68.3 million tons of e-waste generated annually worldwide is formally collected and recycled.
For scale, the world already generates 68.3 million tons of e-waste annually across all categories, and less than a quarter of it is formally collected and recycled. BAN expects that global total to nearly triple to 211 million metric tons a year by 2050, with AI responsible for 15 to 20 percent of the pile.
Earlier estimates were considerably smaller. A 2024 study projected AI e-waste in the range of 1.2 million to 5 million tons by 2030. A separate February analysis put AI-server e-waste at 131,000 to 225,000 tons a year by the end of the decade — comparable to the total e-waste output of a country the size of Denmark, but a fraction of what BAN now expects once full-facility infrastructure is included.
The report's author frames the missing count as a planning failure rather than a fixed constraint of the technology.
“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.”— Jim Puckett, Founder and chief of strategic direction, Basel Action Network
The disposal pipeline is the second half of the problem. The United States hosts more data centers than any other country and has not ratified the Basel Convention, the treaty designed to limit cross-border shipments of hazardous waste. Investigations have repeatedly found US recyclers exporting e-waste abroad, where much of it moves into informal collection channels. The World Health Organization has documented health risks to millions of children living near or working in backyard recycling operations that burn or bury equipment containing lead and chromium.
The counterweight is that hyperscalers have been investing in refurbishment and component-level recycling for years, and specialized secondary markets exist for retired GPUs, memory, and networking gear. BAN's projections assume today's collection and disposal patterns roughly persist through 2050 — an assumption that will be tested as the volumes climb and as recovering high-value chips becomes commercially attractive on its own merits. The report also does not distinguish between equipment retired to landfill and equipment redeployed at lower tiers of compute, where GPU depreciation cycles can extend hardware useful life by years.
Still, the scale gap between prior estimates and this one is the news. If BAN's methodology holds up to scrutiny, the industry has been under-accounting for the physical footprint of AI infrastructure by nearly an order of magnitude, and the retirement wave is front-loaded into the 2030s as first-generation hyperscale AI facilities hit end-of-life.
For AI operators, the practical implication is that reverse-logistics and end-of-life planning are shifting from a compliance line item to a capacity-planning input. The vendors positioned to matter here — chip refurbishers, closed-loop recyclers, and the hyperscalers building in-house component recovery — stand to capture value from a waste stream that BAN now sizes in the hundreds of millions of tons. The AI infrastructure story has spent two years focused on power and water; hardware retirement is the third constraint, and it is arriving on the same timeline.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
