AI Data Center E-Waste Statistics 2026
By Axis Intelligence Research
Co-author: Aidan Jad | Last updated: September 25, 2026 | License: CC BY 4.0
AI-driven electronic equipment retirement could reach 8.6–13.1 million tonnes a year by 2030, according to the Basel Action Network’s September 2026 white paper. Axis Intelligence Research calculates that 74.4% of the hardware a data center retires each year is not compute at all — it is cooling plant, power distribution, batteries and networking.
Quick Answer
How much e-waste do AI data centers produce? The Basel Action Network (BAN) projects that data center infrastructure alone will retire 2.8 million tonnes (Mt) of equipment in 2030, rising to 15.2 Mt in 2050. Counting AI-induced device replacement outside the data center, the 2030 total is 8.6–13.1 Mt per year. Servers-only estimates are far lower: 131–224.8 kilotonnes in 2030 (de Vries-Gao, 2026). Axis Intelligence Research finds the difference is scope, not arithmetic: a gigawatt of AI capacity retires roughly 14,215 tonnes of hardware a year, and only a quarter of it is servers.
Key Findings
- Axis Intelligence Research estimates that each gigawatt of AI data center capacity retires about 14,215 tonnes of equipment per year, derived from BAN’s 70,000 t/GW mass intensity and category lifespans.
- Axis Intelligence Research calculates that 74.4% of a data center’s annual retired hardware mass is non-compute equipment — cooling, power, backup batteries and networking.
- Axis Intelligence Research’s Retirement Velocity Factor (RVF™) shows accelerators, servers and racks leave the building 1.97 times faster than their share of installed mass, while power gear exits at 0.62 times.
- Axis Intelligence Research finds that BAN’s 2030 data-center-only estimate (2.8 Mt) is 12.5 to 21.4 times the leading servers-only estimate, a gap explained almost entirely by what each study counts.
- The UN Global E-waste Monitor 2024 reports 62 million tonnes of e-waste generated in 2022, with only 22.3% documented as properly collected and recycled.
How Much E-Waste Will AI Data Centers Generate by 2030?
The honest answer comes in three sizes, and which one you quote depends on where you draw the fence around “AI e-waste.”
| Estimate | Scope | 2030 figure | Source |
|---|---|---|---|
| Supply-constrained servers-only | AI servers | 131–224.8 kt/yr | de Vries-Gao, Resources, Conservation and Recycling (2026) |
| Data center infrastructure (Band 2) | Servers + power + cooling + backup + networking | 2.8 Mt/yr | Basel Action Network, Sept 2026 |
| Data center + AI “contagion” | Band 2 + early device/telecom retirement | 8.6–13.1 Mt/yr | Basel Action Network, Sept 2026 |
| Cumulative LLM hardware | LLM-related hardware, 2020–2030 total | 1.2–5.0 Mt (cumulative) | Wang et al., Nature Computational Science (2024) |
The narrowest number comes from Alex de Vries-Gao at Vrije Universiteit Amsterdam, who capped server output at what the chip supply chain can physically produce. VU Amsterdam’s announcement of the study puts AI server e-waste at 131–224.8 kilotonnes per year in 2030 — roughly the annual e-waste of Denmark, Norway or Austria.
The widest comes from the Basel Action Network’s The Coming AI Waste Wave, Part 1 of a four-part white paper released in September 2026. BAN counts the whole building and projects 8.6–13.1 Mt of AI-driven equipment retired per year by 2030.
The original academic anchor, Wang et al. in Nature Computational Science, modeled LLM hardware only and found cumulative e-waste of 1.2–5.0 Mt over 2020–2030. That paper is cumulative; the other two are annual. Mixing them in a headline — which much of the coverage has done — compares a decade to a year.
AI data center e-waste by year, 2024–2050
| Year | Installed capacity (GW) | Data center infrastructure retired (Mt/yr) | Source |
|---|---|---|---|
| 2024 | 122 | 1.7 | BAN Table 1 |
| 2030 | 202 | 2.8 | BAN Table 1 |
| 2035 | — | 4.3 | BAN Table 1 |
| 2040 | — | 6.6 | BAN Table 1 |
| 2045 | — | 10.0 | BAN Table 1 |
| 2050 | 1,093 | 15.2 | BAN Table 1 |
Aidan Jad: Read this table the way you’d read a charger rollout plan: the megawatts are the plan, the tonnage is the bill that arrives later. BAN’s curve is almost entirely driven by one input — capacity compounding at 8.8% a year from 122 GW. Drop that to 6% and BAN’s own sensitivity table cuts the 2050 figure from 15.5 to 7.9 Mt. The steepness is a buildout assumption, not a hardware one.
What Is Inside an AI Data Center That Becomes E-Waste?
BAN’s reference 100 MW AI facility carries about 7,000 tonnes of equipment across five categories. Servers are the smallest heavy category.
| Equipment category | Share of facility mass | Assumed replacement cycle | Source |
|---|---|---|---|
| Cooling systems | 35% | 5 years | BAN (lifespan: BAN estimate) |
| Power supply & distribution | 34% | 8 years | BAN (lifespan: BAN estimate) |
| Backup power (batteries, generators) | 15% | 5 years | BAN |
| Accelerators, servers, racks | 13% | 2.5 years | BAN |
| Networking | 3% | 3.5 years | BAN |
Two physical details anchor the numbers. A single 100 MW facility needs roughly 2,700 tonnes of copper for cabling and busbars, and a fully populated GB200 NVL72 rack weighs about 1,360 kg, per BAN’s component inventory. The rack is dense and valuable; the copper is heavy and everywhere.
Retirement Velocity Factor (RVF™): Which Hardware Leaves the Data Center Fastest?
Installed mass tells you what a building weighs. It doesn’t tell you what the loading dock ships out every year. For that, mass has to be divided by lifespan — and nobody had published that split by category. So Axis Intelligence Research built it.
Definition. The Retirement Velocity Factor (RVF™) is a category’s share of annual retired mass divided by its share of installed mass. An RVF above 1.0 means the category leaves the building faster than its weight on the floor would suggest; below 1.0, slower.
Formula. Annual retired mass per GW for category c = 70,000 t/GW × mass share(c) ÷ replacement cycle(c) RVF(c) = [retired mass(c) ÷ total retired mass] ÷ mass share(c)
All inputs are BAN’s published parameters (70,000 t/GW, five mass shares, five lifespans). Anyone can recompute every row below.
| Category | Retired per GW per year (t) | Share of annual retired flow | RVF™ | Inputs source |
|---|---|---|---|---|
| Cooling systems | 4,900 | 34.5% | 0.98 | BAN parameters; Axis calculation |
| Accelerators, servers, racks | 3,640 | 25.6% | 1.97 | BAN parameters; Axis calculation |
| Power supply & distribution | 2,975 | 20.9% | 0.62 | BAN parameters; Axis calculation |
| Backup power | 2,100 | 14.8% | 0.98 | BAN parameters; Axis calculation |
| Networking | 600 | 4.2% | 1.41 | BAN parameters; Axis calculation |
| Total | 14,215 | 100% | — | Axis calculation |
Three things come out of this that the source report does not state directly:
- Compute is overrepresented in the waste stream, but still a minority of it. Servers are 13% of what’s installed but 25.6% of what’s retired each year. Even so, 74.4% of annual retired mass is non-compute.
- Cooling is the single largest annual retirement stream at 4,900 t per GW — more than servers, on BAN’s 5-year cooling assumption.
- The RVF™ ranks where lifespan assumptions matter most. Accelerators (1.97) and networking (1.41) are the categories where a change in refresh policy moves tonnage fastest. Power distribution (0.62) is heavy but slow, so its tonnage is set by the buildout rate rather than by refresh decisions.
Our total of 14,215 t per GW per year matches BAN’s own rounded statement of about 14,000 tonnes retired annually per gigawatt, which is the check we ran against their text.
Aidan Jad: The RVF™ is the number an ITAD buyer or a site EHS lead should care about. A factor of 1.97 on compute means the server cage is where your triage throughput and data-sanitisation capacity get stress-tested first. A factor of 0.62 on power gear means the copper wave is lagged — it arrives with retrofits and superconductor swaps, not with every GPU generation.
Why Do AI E-Waste Estimates Differ by 40 to 60 Times?
BAN states that its 2030 figure is roughly 40–60 times the most widely cited academic projection. That comparison includes the contagion bands — devices outside the data center. Strip those out and compare like for like.
Axis Intelligence Research scope-gap calculation: BAN data-center-only 2030 (2.8 Mt) ÷ de Vries-Gao 2030 high (0.2248 Mt) = 12.5× BAN data-center-only 2030 (2.8 Mt) ÷ de Vries-Gao 2030 low (0.131 Mt) = 21.4×
So roughly one order of magnitude of the headline gap is building scope, and the rest is contagion. The two studies are not rivals on the same question; they answer different ones.
One further check: if compute is 25.6% of the annual retired flow, BAN’s model implies about 0.72 Mt of server-category retirement in 2030 — still 3.2 to 5.5 times de Vries-Gao’s figure. That residual is the lifespan disagreement: BAN assumes 2.5 years for accelerators, while de Vries-Gao argues that a three-year lifespan is already too pessimistic. BAN’s own sensitivity table shows that stretching accelerator life to 4 years only trims 2050 Band 2 from 15.5 to 14.0 Mt — because 87% of the mass isn’t accelerators.
Aidan Jad: Both camps are right about their own fence. The servers-only work is tighter on chip supply; BAN is right that nobody was weighing the chillers. Where I’d push back on BAN is the 8-year power-distribution and 5-year cooling lives — both are explicitly BAN estimates, and traditional power plant runs 15–20 years. If those stretch, the curve flattens.
How Fast Are Hyperscalers Retiring AI Servers?
Depreciation schedules in SEC filings are the closest thing to a disclosed refresh cycle — and the hyperscalers don’t agree with one another.
| Company | Change | Effective | Financial effect | Source |
|---|---|---|---|---|
| Amazon | Useful life of a subset of servers & networking cut from 6 to 5 years | Jan 1, 2025 | +$1.4B depreciation in 2025 | Amazon Form 10-K FY2025 |
| Oracle | Useful life of servers & networking raised from 5 to 6 years | Fiscal 2025 | −$733M operating expenses | Oracle Form 10-K FY2025 |
Amazon’s filing attributes its shorter schedule to the increased pace of technology development in AI and machine learning. Oracle moved the other way in the same window. An accounting life is not a physical life, but when a filer shortens it, the balance sheet is telling you equipment is expected to leave service sooner.
Aidan Jad: Put these two filings next to BAN’s 2.5-year accelerator assumption and you get the real range: 2.5 years at the aggressive end, 5–6 years on the books. The truth for a frontier training cluster is probably nearer the short end; for inference fleets, nearer the long one.
How Does AI E-Waste Compare With Global E-Waste?
The UN Global E-waste Monitor 2024 (ITU and UNITAR) reports:
| Metric | Value | Year | Source |
|---|---|---|---|
| Global e-waste generated | 62 Mt | 2022 | Global E-waste Monitor 2024 |
| Documented as properly collected & recycled | 22.3% | 2022 | Global E-waste Monitor 2024 |
| Projected e-waste generated | 82 Mt | 2030 | Global E-waste Monitor 2024 |
| Projected documented recycling rate | 20% | 2030 | Global E-waste Monitor 2024 |
For scale only: BAN’s 2030 AI-driven range (8.6–13.1 Mt) equals 10.5–16.0% of the Monitor’s 82 Mt projection. We state this as a sense of proportion, not an addition — BAN notes that the Monitor’s placed-on-market method does not isolate data center infrastructure, and the two datasets are built differently.
Looking further out, BAN projects total global e-waste of 196–211 Mt per year by 2050, of which 31–46 Mt is AI-driven, and cumulative AI-driven retirements of 395–617 Mt over 2025–2050.
How Much Hardware Is the 2026–2030 Buildout Locking In?
BAN expects roughly 100 GW of new data center capacity between 2026 and 2030. At 70,000 t per GW, Axis Intelligence Research estimates that this buildout embeds about 7.0 Mt of electro-mechanical equipment — every tonne of which enters a replacement cycle the day it is energised.
| Input | Value | Source |
|---|---|---|
| New capacity 2026–2030 | ~100 GW | BAN |
| Mass intensity | 70,000 t/GW (range 62,000–77,000) | BAN |
| Embedded equipment mass | ~7.0 Mt | Axis calculation |
In the United States alone, BAN counts 4,871 facilities in the pipeline as of late August 2026, citing Aterio’s database.
Can Circular Strategies Cut AI Data Center E-Waste?
Wang et al. modeled circular-economy strategies along the generative-AI value chain — reuse, lifetime extension, component recovery — and found they could reduce e-waste generation by 16–86%. That range was estimated for LLM hardware only, so it does not transfer automatically to chillers and switchgear. BAN plans to examine reuse and refurbishment potential in Part 2 of its series.
The RVF™ suggests where to aim first: categories with a factor above 1.0 (accelerators, networking) respond most directly to lifespan extension, while power and cooling respond mainly to retrofit design — whether a hall can be uprated instead of gutted.
Methodology
Collection. Every figure was retrieved on September 24, 2026 from the documents listed in the CSV’s source_url column: BAN’s The Coming AI Waste Wave Part 1 (PDF), Wang et al. (2024, Nature Computational Science), VU Amsterdam’s release on de Vries-Gao (2026), the UN Global E-waste Monitor 2024, and the FY2025 Form 10-K filings of Amazon and Oracle on SEC EDGAR.
Axis calculations.
- Annual retired mass per GW = 70,000 × mass share ÷ replacement cycle, summed across five categories (BAN parameters).
- Retirement Velocity Factor (RVF™) = share of annual retired flow ÷ share of installed mass.
- Scope-gap multiple = BAN Band 2 (2030) ÷ de Vries-Gao (2030), low and high.
- Embedded mass of new capacity = 100 GW × 70,000 t/GW.
- Share of Monitor projection = BAN 2030 AI-driven range ÷ 82 Mt (illustrative only). All arithmetic was run twice — floating point and exact rational (Python
fractions).
Scope. The RVF™ inherits BAN’s steady-state stock-and-flow structure: it describes the retirement mix of a mature installed base, not the retrofit surges BAN flags separately. Two of the five lifespans (power distribution, cooling) are BAN estimates rather than sourced figures, and RVF™ readings for those categories move with them. We do not merge BAN’s and de Vries-Gao’s figures into a single number because they measure different scopes.
About This Dataset
File: ai-data-center-e-waste-statistics.csv — 102 rows, one observation per row, with source_org, source_document, source_url, retrieved_date, is_primary, axis_calculated and method_note on every row. License: CC BY 4.0. Reuse freely with attribution: “Axis Intelligence Research, AI Data Center E-Waste Statistics 2026, axis-intelligence.com/ai-data-center-e-waste-statistics/”.
Cite This Page
APA: Axis Intelligence Research & Jad, A. (2026, September 24). AI data center e-waste statistics 2026: The 87% nobody counted. Axis Intelligence. https://axis-intelligence.com/ai-data-center-e-waste-statistics/
MLA: Axis Intelligence Research, and Aidan Jad. “AI Data Center E-Waste Statistics 2026: The 87% Nobody Counted.” Axis Intelligence, 24 Sept. 2026, axis-intelligence.com/ai-data-center-e-waste-statistics/.
Chicago: Axis Intelligence Research and Aidan Jad. “AI Data Center E-Waste Statistics 2026: The 87% Nobody Counted.” Axis Intelligence. September 24, 2026. https://axis-intelligence.com/ai-data-center-e-waste-statistics/.
More From the AI Data Center Cluster
- AI data center statistics: power, capacity and investment
- How much electricity AI data centers consume
- AI data center water usage by hyperscaler
- The carbon footprint of AI data centers
- Data center capacity, occupancy and density
- Europe’s data center market in numbers
FAQ: AI Data Center E-Waste
Is “5 million tonnes of AI e-waste by 2030” an annual figure?
No. The 1.2–5.0 Mt range from Wang et al. (2024) is cumulative over 2020–2030 and covers LLM hardware only. The annual 2030 estimates range from 131–224.8 kilotonnes (servers only, de Vries-Gao) to 2.8 Mt (full data center, BAN).
Are cooling systems and transformers legally e-waste?
Under the Basel Convention’s e-waste amendments that took effect in 2025, BAN reads controlled e-waste as any equipment or component containing electronic circuits. A mixed decommissioning load is treated as controlled until sorting and testing show otherwise.
Which part of an AI data center produces the most e-waste each year?
Cooling. Axis Intelligence Research calculates cooling at 4,900 tonnes per gigawatt per year, or 34.5% of annual retired mass, ahead of accelerators and servers at 3,640 tonnes (25.6%), using BAN’s parameters.
What is the Retirement Velocity Factor (RVF™)?
It is an Axis Intelligence Research metric: a hardware category’s share of annual retired mass divided by its share of installed mass. Accelerators, servers and racks score 1.97; power distribution scores 0.62.
How long do AI servers last before they are retired?
Estimates diverge. BAN assumes 2.5 years for accelerators. Amazon depreciates a subset of servers over 5 years since January 2025, and Oracle over 6 years since fiscal 2025. De Vries-Gao argues three years is too short.
How much equipment does one gigawatt of AI capacity contain?
About 70,000 tonnes across five categories, with a range of 62,000–77,000 tonnes, per BAN’s bottom-up inventory of a reference 100 MW AI facility scaled to a gigawatt.
Does liquid cooling reduce e-waste?
Not in mass terms during the transition. BAN reports that converting air-cooled halls to liquid cooling scraps CRAC units and air handlers, and that liquid-cooled rack assemblies integrate cold plates with compute, so they retire together.
How much of today’s e-waste is actually recycled?
The UN Global E-waste Monitor 2024 documents 22.3% of 2022’s 62 million tonnes as properly collected and recycled, and projects that rate falling to 20% by 2030.