AI E-Waste May Be 40–60 Times Higher Than Previously Estimated

A new analysis puts AI e-waste at a scale no earlier study came close to. According to a white paper released on 16 September by the Basel Action Network (BAN), between 395 and 617 million tonnes of AI-driven electronic equipment could be retired from service between 2025 and 2050. Packed into shipping containers and placed end to end, they would circle the Earth about six times.

The environmental debate around artificial intelligence has so far revolved around three things: electricity, carbon and water. Almost nobody has asked what happens to the metal. Yet a data centre is not a cloud: it is thousands of tonnes of copper, steel, batteries, chillers and printed circuit boards, most of which becomes waste within a handful of years.

Why is this number so much higher than earlier ones?

The answer comes down to a single figure: 13 percent.

Previous studies — including the most widely cited ones by Wang et al. (2024) and de Vries-Gao (2026) — counted essentially only servers and accelerators (GPUs). By mass, these represent just 13 percent of a data centre’s electro-mechanical infrastructure.

BAN’s model breaks down a 100 MW reference facility containing roughly 7,000 tonnes of equipment:

• cooling systems – 35% (about 2,450 tonnes)
• power supply and distribution – 34% (about 2,380 tonnes: transformers, switchgear, UPS units, busbars, cabling)
• backup power – 15% (battery banks and diesel generators)
• accelerators, servers, racks – 13%
• networking equipment – 3%

In other words, earlier estimates simply left out 87 percent of the mass. The paper’s analogy is apt: it is like estimating the scrap value of a car factory by weighing only the engines.

This is not a semantic quibble. Under the Basel Convention’s e-waste amendments, in force since 1 January 2025, any whole unit or component containing electronic circuits counts as controlled e-waste — as do the residues from processing them. An integrated cooling unit or an intelligent power distribution unit is therefore e-waste in exactly the same way a retired GPU is.

AI e-waste in numbers

BAN did not model demand. It took the industry at its word about what it intends to build. McKinsey puts capital expenditure at close to $7 trillion between 2025 and 2030 just to construct the infrastructure that will power AI. In energy terms that means roughly 219 GW of total data centre capacity by 2030 — 80 percent more than existed in 2025.

Behind every gigawatt of capacity sits an average of 70,000 tonnes of physical equipment. Applying category-specific lifespans, that works out at roughly 14,000 tonnes of retired equipment per gigawatt per year.

The results:

The 2030 figure of 8.6–13.1 million tonnes is roughly 40 to 60 times the most widely cited academic projection. The 2050 total of 196–211 million tonnes is more than triple today’s roughly 67 million tonnes — and 19 to 28 percent above what the UN Global E-Waste Monitor’s own trajectory implied. The UN report mentions artificial intelligence exactly once, in the foreword; it plays no part in the model.

The contagion that leaves the data centre

The most striking — and by the authors’ admission the least researched — section concerns what BAN calls the AI Waste Contagion. This waste is not generated inside the data centre. It is generated in our homes, our offices and the mobile network.

It works through three channels:

1. Hard hardware thresholds. Microsoft Copilot+ requires a neural processing unit capable of 40 trillion operations per second and 16 GB of RAM. Apple Intelligence needs an iPhone 15 Pro or newer. These are not recommendations; they are cut-offs. Gartner projects that 100 percent of enterprise PC purchases will be AI PCs by the end of 2026, and IDC that 93 percent of PCs shipped will be AI-capable by 2028. The consumer is not deciding: the non-AI option is disappearing from the shelf.
2. Edge AI devices. Smart cameras, sensors with on-device inference, vehicle computing modules — a $26 billion market in 2025, $59 billion by 2030. Four-year lifecycles, and effectively no established end-of-life pathway.
3. Telecommunications infrastructure. Every step of the 400G → 800G → 1.6T transition retires the previous generation, and the optical networking market is growing at roughly double its historical rate.

The key finding: this downstream wave will be larger than the data centre waste itself. By 2050, 16.2–30.7 million tonnes a year, against 15.5 million tonnes from inside data centres. The pattern is familiar from the Windows 10 end-of-support analyses — just at a far smaller scale. There are ways to push back: initiatives like the ChromeOS Flex pilot show that older laptops can be given a second working life rather than retired on a vendor’s schedule.

Importantly, there is no double counting here. The model does not ask “does this device use AI?” but “was it replaced earlier than it otherwise would have been?” There are plenty of historical precedents: the analogue-to-digital TV switchover, the 3G shutdown, Windows 11’s processor requirements.

Why is this hardware so short-lived?

The industry has a metaphor that has shaped its thinking since around 2012: “cattle, not pets.” In the old world a server was a pet, with a name, maintenance and repairs. In the new world it is a number, and when it breaks it is replaced.

That culture is the precise opposite of the circular economy — essentially the return of “take, make, dispose” with a modern vocabulary.

The lifespans BAN applies:

• accelerators, servers, racks: 2.5 years (NVIDIA ships a new architecture roughly every two years: Hopper → Blackwell → Rubin)
• networking equipment: 3.5 years
• backup power: 5 years
• cooling: 5 years
• power distribution: 8 years

For comparison, a conventional server used to run for 5–7 years. For accelerators the binding constraint is not physical but economic life: the next generation is fast enough that keeping the working predecessor makes no commercial sense.

Modularity will not save the situation either. NVIDIA’s GB200 NVL72 rack ships as a single factory-integrated unit in which liquid cooling, compute boards and network fabric are physically built together. You cannot swap “just the chip”, because the cooling manifolds and power delivery are designed around that specific chip’s thermal and electrical profile. A new generation means the whole assembly goes.

Three “rip-and-replace” waves arriving at once

The model assumes orderly, scheduled replacement. Reality will be rougher, and the paper identifies three transitions that will discard perfectly functional equipment:

• The superconductor transition. VEIR’s commercial high-temperature superconducting cables begin deployment in 2027 and are 15 times lighter than copper. Copper cabling and busbars are the heaviest single element in a data centre — a 100 MW facility holds around 2,700 tonnes of copper for cabling and busbars alone.
• The shift to liquid cooling. In 2026, 70 percent of data centres are still air-cooled, but above 40–50 kW per rack there is no alternative. The conversion scraps not only CRAC units and raised floors, but the servers too.
• Battery chemistry transitions. The move from lead-acid to lithium-ion is already under way; sodium-ion and solid-state arrive between 2027 and 2030. Two full replacement cycles inside a single decade.

The historical analogy is display waste: nobody built the treatment capacity before the CRT-to-LCD switchover either — the billions of leaded-glass tubes just arrived.

How toxic will it be?

Part 3 of the white paper will address this in depth, but the warning signs are already visible. The dielectric fluids used in two-phase immersion cooling — 3M’s Novec line, discontinued in 2025, among them — are PFAS, the so-called forever chemicals. In April 2024 the EPA designated PFOA and PFOS as hazardous substances under CERCLA, and PFOS is listed under Annex B of the Stockholm Convention. The EU is developing a broad PFAS restriction under REACH.

On top of that come refrigerants (R-410A, R-134a — potent HFC greenhouse gases), lead-acid batteries, the thermal runaway risk of lithium cells, and the fuel, oil and exhaust fluid waste streams of backup diesel generators.

The uncomfortable part: the system already can’t cope

What gives the analysis its weight is that all of this lands on a system that is failing at today’s volumes.

According to the Global E-Waste Monitor, the world manages barely one fifth of its e-waste responsibly. The rest ends up in informal channels, export chains and dumps. BAN’s own investigations have found that much of it reaches countries where processing poisons workers and local environments — often shipped by companies certified as responsible recyclers.

Jim Puckett, BAN’s founder, made the point that the time to prevent toxic waste is before it is created, not twenty years later. His remark targets the planning gap: to their knowledge, no hyperscaler — not AWS, Azure, Google, Meta, Oracle or Alibaba — and no government or international body has published a plan for managing AI-driven e-waste. There are plenty of carbon-neutrality and water-stewardship commitments. On waste, there is nothing.

What can be done?

The report is explicit that its figures are a floor. Data sovereignty rules forcing duplicated capacity, geopolitical risk and the under-counting of telecom waste all push the numbers up. The steepest stretch will be the 2027–2035 window, when hardware installed during the current build-out reaches end of life in bulk.

Realistic levers:

• Design for reuse. Part 2 will examine how much reuse, refurbishment and repurposing can actually shave off these totals. The ITAD industry sees a gold mine in surplus hardware — the open question is what can be done with tightly integrated, purpose-built racks.
• Requiring modular architecture. The current trend runs towards more integration, meaning more waste per upgrade. Regulation and procurement terms can influence this.
• Building treatment capacity in advance. The CRT lesson: if capacity follows the waste, it is already too late.
• The procurement side. Communities hosting data centres are not only getting jobs and tax revenue — they are getting a point source of hazardous waste that produces every two to three years. That belongs on the siting agenda.

At the individual level the room for manoeuvre is smaller but real: extend device lifetimes, repair where possible, and when there is no more life left, hand equipment in at a proper collection point rather than leaving it in a drawer.

AI genuinely feels weightless. A prompt weighs nothing. But behind every answer sit thousands of tonnes of copper, coolant, batteries and silicon — and so far nobody has a plan for what happens to them when they leave service.

Frequently Asked Questions About AI E-Waste

What is AI e-waste and how does it differ from conventional e-waste?

AI e-waste is electronic equipment retired from artificial intelligence infrastructure. It covers far more than servers and GPUs: cooling systems, power distribution, backup batteries and networking gear make up 87 percent of a data centre’s equipment mass. Under the Basel Convention’s 2025 amendments, any unit containing electronic circuits qualifies as controlled e-waste.

How much AI e-waste could be generated by 2030 and 2050?

The Basel Action Network model projects 8.6 to 13.1 million tonnes of AI-driven equipment retired annually by 2030, rising to 31 to 46 million tonnes a year by 2050. Cumulatively between 2025 and 2050 the figure reaches 395 to 617 million tonnes. Total global e-waste could hit 196 to 211 million tonnes annually, against roughly 67 million today.

Why is this 40 to 60 times higher than earlier projections?

Because earlier studies counted essentially only servers and accelerators, which represent just 13 percent of a data centre’s equipment mass. BAN’s model includes all five equipment categories and adds the AI-driven device replacement happening outside data centres. The gap is therefore not a methodological dispute but a difference in scope.

What is the AI Waste Contagion?

It describes equipment outside data centres forced into early replacement by AI requirements: laptops that cannot run Copilot+, phones excluded from Apple Intelligence, networking gear retired by bandwidth transitions. BAN estimates this at 16.2 to 30.7 million tonnes a year by 2050, exceeding the 15.5 million tonnes projected from inside data centres themselves.

Why is AI hardware so short-lived?

The model assumes 2.5 years for accelerators and 3.5 years for networking equipment, against 5 to 7 years for conventional servers. The driver is economic rather than physical obsolescence: each new chip generation is fast enough that keeping the working predecessor makes no commercial sense. Factory-integrated racks also rule out chip-level upgrades.

Source:

Basel Action Network, The Coming AI Waste Wave – Part 1: How Big Is The AI Waste Wave? (Jim Puckett, September 2026), and BAN’s press release of 16 September 2026. Full white paper: wiki.ban.org