A new report says the AI boom will leave behind far more electronic junk than anyone has counted so far. The nonprofit Basel Action Network (BAN) puts AI-related e-waste at 395 to 617 million metric tons by 2050, according to The Verge AI. That’s enough to fill 23 million 40-foot shipping containers, or a line of containers long enough to wrap around the planet six times.
The headline number is dramatic, but the interesting part is how BAN got there. Previous studies counted servers and GPUs. BAN counted everything else in the building too.
What the researchers actually measured
Most prior estimates of AI’s e-waste footprint stopped at the racks: the accelerators, the CPUs, the memory. BAN argues that approach misses roughly 87 percent of a data center’s electro-mechanical infrastructure. So the new report widens the scope to include:
- Power supply and distribution equipment
- Cooling systems
- Backup power (think battery banks and generators)
- Networking gear
- What BAN calls “AI Waste Contagion”: telecom infrastructure and personal devices that get replaced earlier because AI makes them obsolete faster
From that inventory, BAN lands on a rule of thumb: about 70,000 metric tons of e-waste per gigawatt of data center capacity. It then applies a McKinsey projection that global data center capacity could hit 219GW by 2030, and models equipment retirements from 2025 through 2050.
The numbers, side by side
Here’s how BAN’s estimate compares with earlier work, as detailed in The Verge AI:
- BAN (2026): 395 to 617 million metric tons cumulative by 2050, or 8.6 to 13.1 million metric tons per year
- A 2024 study: 1.2 to 5 million tons total by 2030
- A February 2026 study: 131,000 to 225,000 tons per year from AI servers by 2030, roughly Denmark’s entire annual e-waste output
BAN also projects that global e-waste of all kinds will triple to as much as 211 million metric tons a year by 2050, with AI responsible for 15 to 20 percent of it.
“AI may feel weightless, but every model depends on an enormous amount of highly specialized, cutting edge hardware,” said Jim Puckett, BAN’s founder and chief of strategic direction. He warned that without planning, “today’s AI buildout could become an even more cataclysmic toxic waste crisis than we are already experiencing.”
Why the gap between studies is so wide
What stands out here is that all three studies could be right, depending on what you’re asking. If the question is “how much GPU scrap will AI produce,” the smaller numbers hold. If the question is “what does a data center leave behind when it’s torn out,” BAN’s number is closer to reality. Chillers, transformers, switchgear and UPS batteries don’t vanish when a facility upgrades.
The contagion category is the softest part of the methodology. Attributing a replaced smartphone or router to “AI” involves judgment calls, and BAN gives a wide range (395 to 617 million tons) partly for that reason. Readers should treat the upper bound as a scenario, not a forecast.
The recycling problem underneath
The volume matters because the world already handles e-waste badly. Less than a quarter of the 68.3 million tons produced each year gets formally collected and recycled. The rest often flows into informal operations where workers burn or bury equipment, exposing themselves and nearby communities to lead and chromium. The World Health Organization says millions of children are affected.
The US, home to more data centers than any other country, still hasn’t ratified the Basel Convention that restricts international trade in hazardous waste. Investigations have found US recyclers shipping e-waste abroad, where it ends up in “backyard recycling.”
What practitioners can do with this
If you’re buying or building compute, this report gives you a defensible number to plug into sustainability planning: 70,000 tons per gigawatt. A few practical moves follow from it:
- Ask cloud and colocation providers about decommissioning plans, not just PUE and renewable energy mix.
- Favor hardware programs with certified take-back and refurbishment (e-Stewards and R2 are the common labels).
- Push for longer refresh cycles where workload allows. Not every rack needs the newest accelerator.
- Track infrastructure retirements, not just server retirements, in ESG reporting.
Expect this to become a regulatory conversation quickly. Energy and water use already draw scrutiny for data centers. Hardware disposal is the next line item, and BAN just handed policymakers a very large number to work with. The full report and The Verge AI’s coverage have the complete methodology.