AI’s Data-Center Waste Problem May Be Far Bigger Than Expected

Introduction

AI may appear to live in the cloud, but every model depends on physical infrastructure: accelerators, servers, cooling systems, power equipment, and networking hardware. As companies race to expand data-center capacity, the environmental cost of AI is no longer limited to electricity and water. What happens when that hardware is retired is becoming an equally important question.

A new report from the Basel Action Network (BAN) argues that earlier estimates of AI-related electronic waste were too narrow. The nonprofit projects that equipment retired because of AI between 2025 and 2050 could generate between 395 million and 617 million metric tons of e-waste. That translates to roughly 8.6 million to 13.1 million metric tons per year. In a more visual comparison, BAN says the upper end could fill about 23 million shipping containers by 2050.

Key points

• The accounting is broader. Earlier studies often focused on servers and GPUs. BAN also includes power supply and distribution equipment, cooling systems, backup power, and networking hardware. It argues that a server-only approach overlooks much of a data center’s electromechanical infrastructure.
• The impact may spread beyond data centers. BAN uses the term “AI Waste Contagion” for telecommunications infrastructure and personal devices that could be replaced earlier as AI capabilities advance. This is not a conventional hardware category, but including it makes the estimate more comprehensive—and more dependent on assumptions.
• Growth projections matter. The report cites a McKinsey projection that global data-center capacity could reach as much as 219 gigawatts by 2030. BAN estimates about 70,000 metric tons of e-waste per gigawatt of data-center capacity. Actual totals will depend on hardware lifetimes, upgrade cycles, reuse, and the pace of AI deployment.
• Formal recycling remains limited. Less than one-quarter of the world’s 68.3 million metric tons of annual e-waste is formally collected and recycled. The rest can enter informal channels, where burning or burying equipment exposes workers and nearby communities to hazardous substances such as lead and chromium.

Why it matters

The report’s central message is less about one definitive number than about what conventional accounting leaves out. AI upgrades can shorten the useful life of expensive equipment, while each new wave of capacity also requires supporting electrical, cooling, and networking systems. If companies report the compute used by a model but not the lifespan, reuse, or final destination of the hardware, part of AI’s environmental cost remains invisible.

The policy implications are significant. Data-center approvals and AI infrastructure plans need to include retirement and hazardous-waste management, not just energy and water requirements. The United States has more data centers than any other country but has not ratified the Basel Convention, which is intended to limit international trade in hazardous waste. Investigations have also found that US e-waste can be shipped overseas and enter poorly regulated informal recycling operations.

For operators, extending hardware life, refurbishing components, and building traceable take-back programs could reduce the long-term burden. BAN’s wider methodology should not be compared directly with studies that count only servers and accelerators. Still, every methodology points to the same issue: AI’s physical footprint does not end when a model is deployed. Waste capacity is becoming a necessary part of scaling AI responsibly.

Source: The Verge AI