Almost every argument about the environmental cost of artificial intelligence has been an argument about flows: how many terawatt-hours a training run consumes, how many liters of water evaporate through a cooling tower, how much carbon a gigawatt campus adds to a regional grid. Those are the numbers that show up in sustainability reports, and they share a convenient property — they stop when you stop. Turn the machines off and the flow ends.
A white paper released on Wednesday by the Basel Action Network argues that the industry has been measuring the easy half. The hard half is stock, not flow: the physical hardware itself, which does not disappear when a model is retired. It gets unbolted, wheeled onto a loading dock, and becomes somebody's problem. According to the group's projections, that problem is between forty and sixty times larger than the most widely cited academic estimate.
What Happened
The Basel Action Network, a nonprofit that monitors compliance with the 1989 UN Basel Convention on transboundary movements of hazardous waste, published Part 1 of a four-part series titled The Coming AI Waste Wave. The opening paper asks a deliberately blunt question — how big is it? — and answers with a methodology that departs from earlier work in one specific way.
Previous quantitative studies counted servers and accelerators. BAN's analysts measured what a data center is actually made of, by mass, and found that servers, accelerators and racks account for roughly 13 percent of a facility's electromechanical infrastructure. The other 87 percent is cooling at about 35 percent, power distribution at 34 percent, backup power at 15 percent, and networking at 3 percent. Nearly all of it meets the legal definition of electronic waste when it is removed. Almost none of it has appeared in any published AI e-waste projection.
Working from a 100-megawatt reference facility, the group lands on a figure of roughly 70,000 metric tonnes of electromechanical hardware per gigawatt of capacity, with a stated range of 62 to 77 kilotonnes. It then applies that density to the buildout the industry says it intends to complete — McKinsey's estimate of nearly $7 trillion in capital expenditure between 2025 and 2030 — and adds an "AI Waste Contagion" term for equipment outside data centers that gets retired early as AI spreads into PCs, phones, edge devices and telecom networks.
The resulting numbers are large enough to require unit conversion to stay legible. By 2030, BAN projects 8.6 to 13.1 million tonnes of AI-driven equipment retired annually. By 2050, 31 to 46 million tonnes a year from data centers alone, with the contagion term adding another 16.2 to 30.7 million tonnes. Cumulatively between 2025 and 2050: 395 to 617 million tonnes, which the group converts into 15 to 23 million forty-foot shipping containers — a chain long enough to circle the equator roughly six times. Added to the conventional e-waste trajectory, total global generation would reach 196 to 211 million tonnes a year by mid-century, more than triple today's figure of about 67 million.
Why It Matters
The projection would be an abstraction if the world had a functioning system for absorbing it. It does not. By the Global e-Waste Monitor's count, roughly one fifth of electronic waste is currently managed responsibly. The remaining four fifths goes somewhere else, and BAN's own field investigations — published under the title Brokers of Shame — have traced a significant share of it to informal processing operations in lower-income countries, frequently routed there by firms holding recycler certifications.
This is where the geography of AI becomes uncomfortable. The capability is built in wealthy jurisdictions with the capital to finance gigawatt campuses and the regulatory infrastructure to permit them. The residue tends to travel in the opposite direction, toward places with neither the equipment nor the labor protections to handle it. Copper recovery by burning insulation off cable is not an exotic scenario; it is standard practice across a number of informal recovery sites, and cabling is one of the five categories BAN says the previous models omitted.
"AI may feel weightless, but every model depends on an enormous amount of highly specialized, cutting edge hardware," said Jim Puckett, the organization's founder. His argument is less about the precision of any single projection than about sequencing: waste prevention is a design decision made before manufacture, not a remediation exercise undertaken two decades after the fact.
There is also a contamination question the series intends to take up later. BAN flags that much of the incoming hardware is likely to carry PFAS — the fluorinated compounds used in everything from cable insulation to cooling loops, now under accelerating regulatory pressure in both the EU and the United States. Part 3 of the white paper series is slated to examine that in detail.
Reaction
Analysts who reviewed the paper broadly accepted the framing and then spent most of their commentary on the arithmetic. The split is instructive: the part of the model built from physical measurement held up well, and the part built from behavioral assumptions did not.
Frank Dickson, principal analyst at Dickson Research, told Network World that the per-gigawatt mass estimate was constructed reasonably — cross-checked against World Economic Forum mineral-intensity figures and Microsoft's disclosed copper consumption at a Chicago facility — and that he had no strong basis to dispute it. His objection was to what happens next in the model: an 8.8 percent compound annual growth rate in capacity sustained for 26 consecutive years, and a 2.5-year retirement cycle for accelerators. He called the headline multiplier the least defensible part of the report.
The lifespan assumption is the sharpest point of contention, because the accounting record runs the other direction. Between 2022 and 2025 every major hyperscaler extended its assumed useful life for servers — Microsoft from four years to six, Alphabet from 4.5 to six, Meta to 5.5, Oracle from five to six. The reasoning is that a chip's working life does not end when it rotates out of a frontier training cluster; it cascades down to inference, then to lower-intensity batch work, for years afterward. BAN acknowledges this literature, calls it untested for AI accelerators specifically, and keeps 2.5 years anyway.
Independent consultant Steven Eric Fisher made a related point about the difference between retirement and disposal. Nvidia introduced the V100 in 2017 and Google Cloud still lists V100 instances; the A100 arrived in 2020 and remains an actively offered AWS platform in 2026. A new generation entering the market does not automatically render the previous one economically useless. Fisher also noted that BAN assigns power distribution an eight-year life while its own text concedes that traditional power infrastructure can last 15 to 20 years — a modeling choice that, multiplied across hundreds of gigawatts, moves the totals substantially.
Nidhi Luthra, executive advisor at Acceligence, offered the reading that seems to have the most consensus behind it: the long-range tonnage may not survive contact with reality, but the measurement framework probably should. "IT may be measuring the wrong thing," she said, and the specific wrong thing is treating physical obsolescence as the trigger. Equipment can work perfectly well and still become commercially unviable because the next generation demands different power density, different cooling, different rack architecture.
What's Next
The report identifies three architectural transitions that could retire functioning infrastructure ahead of any replacement schedule: a shift to superconducting power systems, a shift to liquid cooling, and new battery chemistries for backup power. Each would strand equipment that has nothing wrong with it. In BAN's framing these are "rip-and-replace" events, and they fall disproportionately on the 87 percent that nobody was counting.
The liquid cooling transition is the one already underway. Rack power densities that made sense for air handling in 2019 do not work at current accelerator thermal loads, and operators retrofitting for direct-to-chip or immersion systems are removing air handling gear that has years of service left. That is exactly the dynamic Luthra described: economic obsolescence arriving well ahead of physical failure.
On the policy side, the report's recommendations point toward lifecycle transparency rather than emissions caps. Darin Stahl, distinguished analyst at Info-Tech Research Group, argued that oversight frameworks for large AI developments should require disclosure of equipment lifespans, material turnover, reuse pathways, batteries, refrigerants and fire suppression agents — so that regulation reflects the full infrastructure footprint rather than electricity and water alone. The EU is weighing extended producer responsibility requirements for data center equipment; several US states are examining e-waste mandates for large facilities.
For operators, the practical ask is more modest and arrives sooner. Luthra's version is to put a lifecycle model next to the capacity model: what gets redeployed into lower-tier workloads at refresh, what residual value remains, whether a system is modular enough to upgrade selectively, and what a vendor has actually committed to on take-back and recovery. None of that requires resolving the dispute over 2050 tonnage.
Closing Thoughts
The most useful thing about this report may be that its weakest number is also its most quotable one. Forty to sixty times will travel; the 2.5-year assumption underneath it probably will not survive peer scrutiny. But the 13 percent figure — the observation that the entire public conversation about AI hardware waste has been conducted about one eighth of the hardware — does not depend on any forecast at all. It is a description of what a building contains.
There is something clarifying about the shipping container as a unit. Terawatt-hours are hard to picture and easy to offset on paper. A container is a fixed volume that has to sit somewhere, and 23 million of them is a claim about physical geography rather than accounting. Whether the real number turns out to be 15 million or five, the shape of the obligation is the same: the material has to go somewhere, and at present the world processes about a fifth of what it already produces.
What BAN is really proposing is a change in the tense of the conversation. Energy and water are measured in the present tense — this quarter's consumption, this year's intensity. Hardware is a future-tense problem that behaves like a deferred liability, invisible on the balance sheet until the decommissioning order goes out. The industry spent the last three years learning to account for the flows. The stock has been accumulating the entire time.
Sources: Eurasia Review · Network World · The Coming AI Waste Wave, Part 1 (PDF) · Basel Action Network
한글 요약
국제 환경 감시 단체 바젤액션네트워크(BAN)가 9월 16일 공개한 백서 ‘다가오는 AI 폐기물 물결’ 1부는, AI 데이터센터가 만들어낼 전자폐기물 규모가 기존 학계 추정치의 40~60배에 이를 수 있다고 본다. 차이는 계산 범위에서 나온다. 지금까지의 연구는 서버와 GPU만 셌지만, 서버·가속기·랙은 데이터센터 전기기계 설비 질량의 13%에 불과하다. 나머지 87% — 냉각 35%, 전력 분배 34%, 백업 전원 15%, 네트워킹 3% — 는 어떤 추정치에도 들어간 적이 없다. BAN은 1기가와트당 약 7만 톤이라는 밀도를 산출한 뒤 2025~2030년 약 7조 달러 규모의 인프라 투자 계획에 적용해, 2050년까지 누적 3억 9,500만~6억 1,700만 톤, 40피트 컨테이너 1,500만~2,300만 개 분량을 전망했다.
분석가들의 반응은 둘로 갈렸다. 기가와트당 질량 추정은 세계경제포럼의 광물 집약도 자료와 마이크로소프트가 공개한 시카고 시설 구리 사용량으로 교차 검증돼 비교적 견고하다는 평가를 받았다. 반면 가속기 수명을 2.5년으로 잡은 가정은 가장 약한 고리로 지목됐다. 2022~2025년 사이 마이크로소프트는 서버 감가상각 기간을 4년에서 6년으로, 알파벳은 4.5년에서 6년으로, 오라클은 5년에서 6년으로 늘렸다. 최전선 학습 클러스터에서 빠진 칩이 추론과 저강도 배치 작업으로 내려가 수년 더 쓰인다는 논리다. 엔비디아 V100은 2017년 제품이지만 구글 클라우드에 아직 남아 있고, 2020년 A100도 2026년 현재 AWS에서 현역이다.
숫자 논쟁과 별개로, 측정 대상 자체를 바꿔야 한다는 데에는 대체로 동의가 모였다. 설비가 물리적으로 고장 나기 전에 경제적으로 쓸모없어지는 속도가 진짜 문제라는 것이다. 액침 냉각 전환, 초전도 전력 설비, 새로운 배터리 화학 같은 예고된 아키텍처 교체는 멀쩡한 장비를 조기 퇴역시킨다. 현재 전 세계 전자폐기물 중 적정 처리되는 비율은 5분의 1 남짓이고, 나머지 상당량은 저소득국의 비공식 처리장으로 흘러간다. EU는 데이터센터 장비에 대한 생산자책임재활용제 확대를 검토 중이며 미국 여러 주도 대형 시설 규제를 살피고 있다. 전문가들의 조언은 단순하다. 용량 계획 옆에 수명주기 계획을 나란히 놓으라는 것.