AI data center e-waste has just been re-measured, and the new number is not a little higher than the old one. It is roughly forty to sixty times higher than the last serious attempt to size AI data center e-waste. On 15 September 2026 the Basel Action Network published The Coming AI Waste Wave, a white paper that projects between 395 and 617 million tonnes of AI-driven electronic equipment retired between 2025 and 2050 — enough to fill 15 to 23 million forty-foot shipping containers.
The Verge’s Justine Calma reported the findings on 16 September 2026 under a headline that summarises the situation exactly: the problem is huge, and it is getting bigger. What makes the report worth reading rather than skimming is not the container analogy. It is the methodology. Coverage of the AI buildout has mostly argued about electricity and water — Al Gore recently argued the real risk is neither — and hardly ever about mass. Every previous estimate of AI data center e-waste counted servers and accelerators. This one counts the building.
That single change of scope is responsible for almost all of the gap between the old AI data center e-waste numbers and the new ones. Servers and accelerators are only about 13 percent of the electro-mechanical mass inside a data center. The other 87 percent — power distribution, cooling plant, backup batteries, networking gear — has never appeared in a published AI data center e-waste projection before. This article walks through what the report actually models, where its arithmetic comes from, which parts are measured and which are estimated, and what any organisation buying AI compute should take from it.
Table of contents
- What the AI Data Center E-Waste Report Actually Claims
- Why Earlier AI Data Center E-Waste Estimates Came In So Much Lower
- Inside the Building: Where AI Data Center E-Waste Mass Actually Sits
- How Quickly the Hardware Dies, and Why That Drives AI Data Center E-Waste
- The AI Waste Contagion Outside the Fence
- Checking the AI Data Center E-Waste Arithmetic
- Where the AI Data Center E-Waste Would Actually Go
- What Could Make the AI Data Center E-Waste Forecast Wrong
- What This Means for Anyone Buying AI Compute
- AI Data Center E-Waste: Frequently Asked Questions
- References
What the AI Data Center E-Waste Report Actually Claims
The white paper is Part 1 of a four-part series, written by BAN founder Jim Puckett. It is deliberately narrow: it asks only how much retired electronic equipment the announced AI data center buildout implies, and leaves the questions of toxicity, recycling and environmental justice to later parts.
The headline AI data center e-waste figure
Cumulative AI-driven electronic equipment retired between 2025 and 2050 comes to 395–617 million tonnes. Packed into forty-foot containers and placed end to end, that is roughly six times the circumference of the Earth.
The annual AI data center e-waste figure
By 2050 the model projects 31 million tonnes a year of AI data center e-waste on conservative assumptions and 46 million tonnes a year on aggressive ones. Total global e-waste from all sources reaches 196–211 million tonnes a year, against roughly 67–68 million tonnes today.
The near-term AI data center e-waste figure
By 2030 — four years away — the report puts AI-driven retirement at 8.6 to 13.1 million tonnes a year. That is the number that generates the “40 to 60 times” headline, and the comparison is worked through later in this article.
The mass-intensity constant
Every gigawatt of installed data center capacity implies roughly 70,000 tonnes of physical equipment across five categories. This is the load-bearing assumption of the whole model, derived from a reference 100 MW facility carrying about 7,000 tonnes.
The quote that frames it
“AI may feel weightless, but every model depends on an enormous amount of highly specialized, cutting edge hardware,” Puckett said. “If companies and governments do not begin planning for this new waste tsunami, today’s AI buildout could become an even more cataclysmic toxic waste crisis than we are already experiencing.”
Why Earlier AI Data Center E-Waste Estimates Came In So Much Lower
Three prior studies are the benchmark, and the report is explicit that it is not accusing any of them of error. It is saying they measured a smaller object.
What earlier AI data center e-waste studies counted
Wang et al. (2024) projected 1.2 to 5 million tonnes of AI e-waste by 2030. De Vries-Gao, published in February 2026, projected 131,000 to 225,000 tonnes a year from AI servers by the end of the decade — comparable, as The Verge put it, to all the e-waste produced by a country the size of Denmark. Gröger et al. added storage. None included power, cooling, networking or batteries.
The arithmetic behind “40 to 60 times”
Divide BAN’s 2030 range by de Vries-Gao’s. The low end is 8.6 million tonnes against 225,000 tonnes, or about 38 times. The high end is 13.1 million tonnes against 225,000 tonnes, or about 58 times. Round those and you get the report’s headline claim, so the AI data center e-waste multiplier is a scope ratio, not a disagreement about growth rates.
The author’s own caveat
De Vries-Gao described the method of scaling hardware mass to power draw as “crude,” and the figure excludes networking gear, cabling, batteries and cooling entirely — which means it understates rather than overstates. BAN notes the newer report was reviewed pre-publication by researchers in the field including de Vries-Gao.
The analogy the report uses
“Counting servers alone is like estimating the scrap value of an automobile factory by weighing only the engines.” It is a fair description of what changed between the studies.
| Study | Published | Scope counted | Projection |
|---|---|---|---|
| Wang et al. | 2024 | AI servers | 1.2–5 Mt total by 2030 |
| de Vries-Gao | Feb 2026 | AI servers, power-scaled | 131,000–225,000 t/yr by 2030 |
| Gröger et al. | 2026 | Servers plus storage | Literature synthesis |
| Basel Action Network | Sep 2026 | Full infrastructure plus contagion | 8.6–13.1 Mt/yr by 2030 |
Inside the Building: Where AI Data Center E-Waste Mass Actually Sits
The report’s facility-level analysis is the most useful section for anyone who has never priced a decommissioning. It breaks a reference 100 MW AI facility into five categories totalling about 7,000 tonnes.
Cooling is the heaviest AI data center e-waste category
About 35 percent of mass, roughly 2,450 tonnes. Cooling is therefore the largest single contributor to AI data center e-waste by weight, and in legacy air-cooled rooms that means chillers, CRAC and CRAH units, cooling towers, ductwork and raised-floor plenums. In AI facilities it is coolant distribution units, piping and manifolds.
Power supply and distribution is a close second
About 34 percent, roughly 2,380 tonnes: transformers, switchgear, UPS units, PDUs, cabling and busbars. A single 100 MW facility needs approximately 2,700 tonnes of copper for cabling and busbars alone.
Backup power adds another 15 percent
Roughly 1,050 tonnes. Hyperscale facilities each require 10 to 50 MWh of battery storage; a 10 MWh lead-acid installation weighs 250–330 tonnes of lead and sulfuric acid, and a lithium equivalent about a third of that. A 100 MW facility also runs 40 to 56 diesel generators at 16–18 tonnes each.
Compute is only 13 percent of AI data center e-waste
About 910 tonnes. At roughly 1,360 kg per fully populated AI rack across about 650 racks in a 100 MW facility, the arithmetic gives 884 tonnes — which is the cross-check that makes the 13 percent figure credible. Dense and valuable per kilogram, but light in aggregate.
Networking is the lightest and the fastest-cycling
About 3 percent, roughly 210 tonnes of switches, optics and cabling. Individual switches weigh 10–30 kg; cabling adds 225–360 kg of copper per rack.
How Quickly the Hardware Dies, and Why That Drives AI Data Center E-Waste
Mass alone does not produce waste. Mass divided by lifespan does, and that ratio is what converts installed capacity into AI data center e-waste. The report assigns a replacement cycle to each category and is careful to label which figures are sourced and which are its own estimates.
Accelerators and servers: 2.5 years of useful life
Compressed from the traditional five to seven years for general-purpose servers by Nvidia’s biennial architecture cadence — Hopper, then Blackwell, then Rubin — where each generation makes the last economically uncompetitive for frontier work. Sourced to Gartner, Dell’Oro, Resource Recycling and de Vries-Gao.
Networking: 3.5 years
Driven by the 400G to 800G to 1.6T to 3.2T transition: four generations of switch and optical transceiver technology in about four years. Equipment is retired well before physical failure.
Power distribution: 8 years, and this one is an estimate
Traditional power infrastructure lasts 15 to 20 years. BAN compresses it for two reasons: density upgrades from 5–15 kW racks to 50–140 kW force wholesale replacement, and high-temperature superconducting cable from VEIR — fifteen times lighter than copper, with a Microsoft partnership announced — begins commercial rollout in 2027. Eight years is the midpoint of a stated 7–10 year range.
Backup power: 5 years
A blended average across the VRLA-to-lithium transition, with sodium-ion and solid-state batteries arriving 2027–2030 potentially triggering a second rip-out inside the decade.
Cooling: 5 years, also an estimate
A blend of air-cooled plant scrapped early during liquid conversion — 70 percent of data centers were still air-cooled in 2026 — and liquid systems themselves lasting five to seven years.
| Category | Share of mass | Assumed life | Sourced or BAN estimate |
|---|---|---|---|
| Accelerators, servers, racks | 13% | 2.5 years | Sourced |
| Networking equipment | 3% | 3.5 years | Sourced |
| Power supply and distribution | 34% | 8 years | BAN estimate |
| Backup power systems | 15% | 5 years | Sourced |
| Cooling systems | 35% | 5 years | BAN estimate |
Working the model by hand
Take the five shares and divide each by its lifespan: 0.34/8 plus 0.35/5 plus 0.15/5 plus 0.13/2.5 plus 0.03/3.5 gives 0.203. In plain terms, about 20 percent of a facility’s mass turns over every year, an effective blended life of just under five years. Multiply that by 219 GW of 2030 capacity at 70,000 tonnes per GW — 15.3 million tonnes installed — and the inside-the-fence AI data center e-waste flow lands near 3.1 million tonnes a year. The rest of BAN’s 8.6–13.1 range comes from somewhere else entirely.
The AI Waste Contagion Outside the Fence
That “somewhere else” is the report’s most original and most contestable idea, and it is the larger half of the total.
What contagion means
Electronic equipment retired outside data centers, forced into early replacement because AI capability became the baseline. Microsoft’s Copilot+ requires a neural processing unit rated at 40 trillion operations per second and 16 GB of RAM. Apple Intelligence requires an iPhone 15 Pro or later. Google’s Gemini requires recent Tensor silicon.
Why the report calls it supply-side
Gartner projects that 100 percent of enterprise PC purchases will be AI PCs by the end of 2026; IDC projects 93 percent of all PCs shipped will be AI-capable by 2028. The report’s point is that this is a forecast of what manufacturers will offer, not what buyers want. The non-AI option disappears from the shelf.
Conservative versus aggressive
The conservative band counts only devices technically unable to run the features demanded of a basic user experience: 16.2 million tonnes a year by 2050. The aggressive band adds the cultural replacement effect — the smartphone that still works but no longer feels adequate — bringing it to 30.7 million tonnes a year.
The two uncounted channels
Edge AI devices, an installed base MarketsandMarkets puts at $26 billion in 2025 rising to $59 billion by 2030, with typical four-year lifecycles and no established end-of-life pathway. And telecoms: optical networking is growing at 10.5 percent a year, roughly double its historical rate, and the Fiber Broadband Association projects a 2.3-fold increase in fibre miles by 2029. The report concedes no published study has quantified AI-driven telecom waste in tonnes at all.
Why contagion dominates
Compare the bands directly. Inside-the-fence retirement runs about 15.5 million tonnes a year by 2050. Contagion runs 16.2 to 30.7. On the report’s own numbers, the AI data center e-waste generated outside data centers exceeds the waste generated inside them in both scenarios.
Checking the AI Data Center E-Waste Arithmetic
Several of the report’s comparisons can be verified with a calculator, which is a reasonable test of whether the headline numbers are doing honest work.
The AI data center e-waste container claim
A forty-foot container is 12.19 metres long. Fifteen million of them end to end is about 182,850 km; twenty-three million is about 280,370 km. Earth’s circumference is 40,075 km, so the range spans 4.6 to 7.0 laps. “About six times” is the midpoint, and it is a fair reading of the range rather than the top of it.
The payload cross-check
617 million tonnes divided across 23 million containers is about 26.8 tonnes each, which sits just inside the typical 26–28 tonne payload limit of a forty-foot box. The two figures are consistent with each other, which is more than can be said for many waste analogies.
The cumulative-to-annual check
395–617 million tonnes spread across the 26 years from 2025 to 2050 averages 15.2 to 23.7 million tonnes a year. That is higher than the 8.6–13.1 figure for 2030 and lower than the 31–46 figure for 2050, which is what a compounding growth model should produce.
The tripling claim
The world produces roughly 68.3 million tonnes of e-waste a year now. The projection of 211 million tonnes by 2050 is 3.09 times that. AI’s 15 to 20 percent share of 211 million tonnes works out at 31.6 to 42.2 million tonnes, which brackets the stated 31–46.
Where the growth rate comes from
Global installed capacity was about 122 GW in 2024. McKinsey projects 171–219 GW by 2030, JLL 200 GW by 2030, ABI Research 277 GW by 2035. The average implied compound growth rate is about 8.8 percent, and the report applies it through 2050 while explicitly flagging that sustaining it for 25 years “might be viewed as a provocative assumption.”
Where the AI Data Center E-Waste Would Actually Go
The disposal side is what turns a tonnage forecast into a public health question, and here the existing baseline is poor rather than uncertain.
Only a fifth of e-waste is handled properly
Less than a quarter of the 68.3 million tonnes of e-waste created worldwide each year is formally collected and recycled. The rest enters informal channels where burning or burying equipment exposes workers and local environments to lead and chromium.
The treaty gap
The United States has more data centers than any other country — 5,388 of roughly 12,259 worldwide as of August 2026, with 4,871 more in the pipeline — and has still not ratified the Basel Convention that governs cross-border movement of hazardous waste. Investigations have repeatedly found US recyclers shipping abroad.
The materials inside AI data center e-waste
Backup batteries are hazardous under the Basel Convention in both lead-acid and lithium chemistries. Air-cooled plant carries R-410A and R-134a refrigerants, potent greenhouse gases already scheduled for phasedown. Two-phase immersion cooling uses fluorocarbon dielectrics that are PFAS; the EPA designated PFOA and PFOS as CERCLA hazardous substances in April 2024.
The people at the other end
The World Health Organization has documented health threats to millions of children working in or living near informal recycling operations. That is the endpoint the AI data center e-waste forecast is describing, not an abstract landfill, and it is why the AI data center e-waste question is a public health question rather than a reporting exercise.
What Could Make the AI Data Center E-Waste Forecast Wrong
The report devotes two full sections to arguing against itself, which is unusual enough to be worth reading on its own terms.
Arguments that the AI data center e-waste model overstates
Hyperscalers including Google and Microsoft have published research on extending server life to five or six years. A refurbishment and resale market could absorb a large share of retired accelerators for inference and lower-tier workloads. Both would cut the AI data center e-waste flow materially if they hold.
Why the report doubts them
Published lifespan-extension research covers general-purpose fleets, not accelerator clusters. And modular upgrade is moving backwards: Nvidia’s GB200 NVL72 ships as a single factory-integrated unit where cooling manifolds, compute boards and networking fabric are built together around one chip’s thermal and electrical profile. You cannot swap the GPUs and keep the rack.
The efficiency objection
If models get cheaper to train, surely fewer servers are needed. The report invokes Jevons’ paradox and points at DeepSeek-R1 in January 2025, trained for about $5.6 million, after which Nvidia lost roughly $589 billion in market value in a single day — and the buildout continued regardless.
Arguments that it understates
Data sovereignty rules in the EU, China, India and Indonesia force the same capability to be duplicated across jurisdictions, so capacity scales faster than demand. Frontier training runs continuously for weeks and cannot be consolidated the way general-purpose workloads can. And conversion waste — ripping out functional power and cooling plant to reach AI density — is a one-time dump the stock-and-flow model does not capture at all.
The historical analogy
The CRT-to-LCD transition left the world with billions of leaded-glass tubes nobody had planned to dispose of. The report’s position is that the industry did not build disposal capacity before building flat screens, and is not doing so now either.
What This Means for Anyone Buying AI Compute
Most readers will never decommission a chiller. The AI data center e-waste question still reaches them through procurement, reporting and refresh policy.
Ask where the AI data center e-waste mass goes, not just the carbon
Sustainability questionnaires for cloud and data center operations have been built almost entirely around energy and water. On this report’s numbers, mass turnover deserves its own line, and 87 percent of the AI data center e-waste mass is not the servers.
Treat refresh policy as a waste decision
The 2.5-year accelerator cycle is an industry norm, not a law of physics. Organisations running their own inference hardware set that cadence themselves, and disciplined IT asset management is the difference between a resale and a skip.
Resist the contagion where it is optional
The forced half of the contagion band is genuinely forced. The aggressive half of the AI data center e-waste contagion is not. Fleet refreshes driven by on-device AI features nobody has asked for are the cheapest tonnage in the entire model to avoid.
Read the Basel definition carefully
The convention now treats whole equipment containing electronic circuits, components from that equipment, and residues from processing it as controlled e-waste. That sweeps in integrated cooling units and power gear that many disposal contracts still treat as ordinary scrap metal.
Watch Part 2
BAN has said the next instalment examines whether reuse and refurbishment can absorb the wave. That is the single assumption with the most leverage over the AI data center e-waste total, and it is currently unresolved.
AI Data Center E-Waste: Frequently Asked Questions
Who published the AI data center e-waste report?
The Basel Action Network, a Seattle-based non-profit focused on toxic trade, on 15 September 2026. The author is founder Jim Puckett and it is Part 1 of a four-part white paper series.
Is the 40-to-60-times claim about growth or scope?
Scope, not growth. Both AI data center e-waste models describe the same buildout; BAN counts power, cooling, networking, batteries and downstream device replacement that the earlier work did not attempt to count.
How much of the AI data center e-waste total is actually inside data centers?
Less than half. About 15.5 million tonnes a year by 2050 comes from inside the fence, against 16.2 to 30.7 million tonnes from the downstream contagion effect.
Which figures are measured and which are assumed?
The AI data center e-waste mass breakdown is built from facility-level component data. The accelerator, networking and battery lifespans are sourced. The power distribution and cooling lifespans, the contagion susceptibility factors and the aggressive band are explicitly labelled BAN estimates.
Does the report say AI should stop?
No. It argues that the AI data center e-waste impacts of the buildout have been ignored while carbon and water received scrutiny, and that planning for disposal should start now rather than after the fact.
What is the single most load-bearing number?
70,000 tonnes per gigawatt. If that constant is wrong, every downstream AI data center e-waste figure moves with it, which is why the report spends a full section deriving it from a reference 100 MW facility.
References
The AI data center e-waste problem is huge — and getting bigger, The Verge
AI E-Waste May Be 40 to 60 Times What Has Been Estimated, Basel Action Network
The Coming AI Waste Wave, White Paper Part 1 (PDF)
AI has an environmental problem, UN Environment Programme
Electronic waste fact sheet, World Health Organization
PFAS guidance, US Environmental Protection Agency
Al Gore Says the Real AI Risk Isn’t Data Centers, Progressive Robot
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.