AI data centers could dump 2.5 million metric tons of e-waste yearly

Key Takeaways

• AI data centers are projected to generate 2.5 million metric tons of e-waste annually, equivalent to 250 Eiffel Towers
• Over 1,500 data centers are in various development stages in the US alone, amplifying the waste problem
• Toxic metals from discarded electronics contaminate water, soil, and food chains, affecting communities far from disposal sites

The AI boom's hunger for compute is creating a waste problem that will outlast any chatbot. A June 2026 report from the United Nations University estimates that hardware turnover from AI data centers could generate 2.5 million metric tons of e-waste annually. That's the equivalent of 250 Eiffel Towers' worth of discarded chips, servers, cables, and networking gear.

Electronic waste is already the fastest-growing category of refuse on the planet. Humans are on track to produce 82 million metric tons of it annually by 2030. The data center buildout now underway threatens to compound that figure significantly.

Where does all this hardware go?

More than 1,500 data centers are in various stages of development across the US alone, according to the Pew Research Center. Hyperscale facilities, the kind Microsoft, Google, and Amazon operate, can cover millions of square feet and pack in at least 5,000 servers each, per IBM data.

The lifespan of that hardware is contested. CNBC reported that Google, Oracle, and Microsoft claim their servers can last up to six years. But some industry analysts suspect these numbers are inflated to make depreciation expenses look better in quarterly earnings.

"The basic issue is that the way current AI systems, especially generative AI models, are developed, reward scale, acceleration, and rapid turnover," said Golestan (Sally) Radwan, chief digital officer of the UN Environmental Programme. The incentive structure prioritizes newer, faster chips over longevity.

The recycling problem no one wants to solve

Spinning media drives, commonly used for archival storage in hyperscale facilities, present a particular challenge. Tony Harvey, vice president and analyst at Gartner, noted these drives are among the hardest electronics to recycle effectively. Erasing massive drives to meet compliance standards takes significant time, and even then, there's no guarantee sensitive data has been completely destroyed.

The result: many drives end up shredded rather than refurbished. That approach solves the data security problem but creates mountains of unrecyclable material.

The communities paying the price

Solomon Njoroge started picking waste at the Dandora dumpsite outside Nairobi, Kenya, as a child. He's watched other waste pickers suffer asthma, miscarriages, cancer, and respiratory diseases from exposure to toxic materials. The 40-acre landfill receives waste from across the region, including electronics shipped from wealthier nations.

“Inasmuch as we are working for development or greatness to change the world… [hyperscalers] should also consider that there are people somewhere, suffering from what they are calling greatness.”

— Solomon Njoroge, founder of the Dandora Recyclable Waste CBO

Toxic metals and chemicals in e-waste, including arsenic, lead, cadmium, and mercury, don't stay local. They contaminate water systems, seep into soil, and enter food chains. A server that once processed your AI queries in Virginia might end up poisoning groundwater in West Africa.

Who bears responsibility?

The UN report and advocates like Njoroge are pushing for waste pickers to have a voice in policy discussions. These informal workers often know more about e-waste flows than regulators do. They're also the ones most directly affected by disposal practices.

Hyperscalers have made sustainability commitments, but those pledges rarely address end-of-life hardware in detail. Microsoft, Google, and Amazon publish carbon neutrality goals; they're less forthcoming about what happens to a GPU that's no longer efficient enough for inference workloads.

What can enterprises do?

Companies procuring AI infrastructure have leverage. They can demand transparency about hardware lifecycle management from cloud providers. They can prioritize vendors with certified e-waste recycling programs. And they can push for longer hardware refresh cycles where performance requirements allow.

The economics of AI encourage churn. Every generation of chips offers better performance per watt, creating pressure to upgrade. But that calculus ignores the externalized costs dumped on communities like Dandora.

Frequently Asked Questions

How much e-waste will AI data centers generate annually?

According to a June 2026 UN University report, AI hardware turnover could generate approximately 2.5 million metric tons of e-waste per year, equivalent to 250 Eiffel Towers.

What toxic materials are found in data center e-waste?

Discarded electronics contain arsenic, lead, cadmium, and mercury, all of which can contaminate water systems, soil, and food chains.

How long do hyperscale data center servers last?

Google, Oracle, and Microsoft claim their servers can last up to six years, though some analysts believe these figures may be optimistic.

Why are data center storage drives hard to recycle?

Large spinning media drives take significant time to erase to compliance standards, and even thorough wiping can't guarantee all sensitive data is destroyed, leading many to be shredded instead of refurbished.

Who is most affected by e-waste from AI data centers?

Informal waste pickers and communities near landfills, particularly in developing nations, face the greatest health impacts from toxic e-waste exposure.

Source: Latest news