The expansion of artificial intelligence (AI) could leave a far bigger material bill than early estimates of e-waste suggest, adding a sharper edge to the broader debate over the environmental footprint of AI data centers. A new report from the Basel Action Network (BAN) calculates that the equipment retired every year because of AI development could reach between 8.6 and 13.1 million tonnes by 2030, a figure 40 to 60 times higher than one of the leading previous academic projections. The difference comes mainly from the fact that BAN includes equipment that’s usually left out of the count, from electrical and cooling systems to networking gear and devices that could be replaced ahead of schedule.
AI e-waste in 30 seconds
- BAN estimates between 8.6 and 13.1 million tonnes of equipment retired per year due to AI in 2030.
- The calculation includes data centers, but also PCs, phones, edge devices, and telecom networks.
- In a reference AI data center, servers and accelerators account for only 13% of the equipment’s mass.
- By 2050, the report puts total e-waste between 196 and 211 million tonnes per year.
- AI alone would account for between 31 and 46 million tonnes per year in that scenario.
The 40-to-60-times figure doesn’t mean AI is going to generate 60 times all the e-waste on the planet. BAN compares its estimate with earlier work that focused mainly on servers and accelerators for AI workloads. Its own model widens that scope and folds in much of the physical infrastructure that keeps those data centers running, plus the accelerated replacement of equipment outside them.
The report, titled The Coming AI Waste Wave and prepared by Jim Puckett in September 2026, presents these figures as a projection tied to infrastructure growth continuing at rates similar to the ones used in its model. It is not, therefore, a fixed prediction of how much waste will necessarily be produced.
The data center weighs a lot more than its GPUs
One of the main differences from earlier estimates shows up when you look at what’s actually inside a data center built for AI. The usual image is racks packed with GPUs, but BAN estimates that accelerators, servers, and racks make up roughly 13% of the total equipment mass at a reference 100 MW facility — a very different picture from the chip-dominated cost breakdown CloudNews has reported for gigawatt-scale AI facilities.
Most of the weight comes from other systems. Cooling accounts for around 35%, electrical distribution for 34%, backup systems for 15%, and networking for about 3%. The result is an estimate of roughly 7,000 tonnes of equipment per 100 MW, equivalent to around 70,000 tonnes per gigawatt of installed capacity. The table on page 15 of the report lays out this breakdown and shows why the math changes so much once you stop counting only the compute hardware.
The electrical section includes transformers, distribution systems, uninterruptible power supplies (UPS), power distribution units, busbars, and large amounts of copper cable. BAN estimates that a 100 MW data center can require around 2,700 tonnes of copper in cabling and busbars alone.
Cooling adds another significant amount of material. Traditional data centers use air conditioning and air cooling systems, while high-density racks built for AI are increasingly turning to liquid cooling. When an existing facility is retrofitted to support these workloads, part of the previous infrastructure can be removed even if it still has useful life left.
Backup systems also generate their own waste stream. BAN accounts for lead-acid and lithium-ion batteries, as well as the diesel generators used to keep facilities running during power outages.
On top of all that comes networking. AI data centers need high-speed connections between servers and accelerators, in a shift that is moving facilities from 400G toward 800G and later generations. For BAN, those changes can shorten the replacement cycles for switches, optical transceivers, and cabling.
Hardware lifespans are shrinking with every generation
BAN’s model doesn’t just calculate how much equipment gets installed, but also how long it stays in service. The report assigns different replacement cycles to each category.
For accelerators, servers, and racks, it uses 2.5 years. For networking it uses 3.5 years, while cooling and backup systems get five years. Electrical distribution gets an eight-year lifespan, though the report itself notes that this last figure is BAN’s own estimate rather than a number published directly by an outside source.
The reasoning is that AI data centers go through faster renewal cycles than conventional computing. An accelerator can keep working physically and still stop being competitive for certain training workloads once a new, higher-performance generation arrives.
The report calls this phenomenon “hyper-obsolescence,” referring to the simultaneous replacement of large volumes of specialized hardware. The problem grows when components are tightly integrated. BAN points to NVIDIA’s GB200 NVL72 as an example, where compute, cooling, and networking are part of a jointly designed architecture — the same tight coupling that CloudNews has seen complicate virtualization on GB200 NVL72 systems.
That limits the option of swapping out just one component while leaving the rest of the installation intact. The report argues that this integration can increase the amount of material removed with every technology upgrade.
There are, however, factors that could bring the numbers down. BAN acknowledges that some operators might extend the useful life of certain servers, that retired hardware can find buyers in secondary markets, and that data center capacity growth could come in below the 8.8% annual rate used in its model. The study itself includes a sensitivity analysis: under a combination of especially conservative assumptions, it calculates around 3.8 million tonnes a year of retired data center infrastructure in 2050, compared with 15.5 million in its baseline scenario.
The second wave could reach PCs, phones, and telecom equipment
The most novel part of the report is what BAN calls “AI Waste Contagion.” The idea is to count the equipment that sits outside data centers but could be retired earlier than planned because of the expansion of AI capabilities.
That covers PCs, mobile phones, tablets, edge devices, and telecom equipment. The mechanism can be technical: certain AI services require a neural processing unit (NPU), a minimum amount of memory, or local capabilities that older devices don’t have.
BAN cites the requirements of devices compatible with Copilot+ and Apple Intelligence as examples. It also factors in the growth of edge devices, such as cameras, sensors, and computing systems for vehicles, along with the renewal of the networks needed to connect this new volume of processing.
The report separates two scenarios. The conservative one estimates that the extra obsolescence driven by AI’s technical requirements would reach 16.2 million tonnes a year by 2050. The aggressive scenario adds replacements driven by demand for new features, reaching 30.7 million tonnes. BAN acknowledges there isn’t enough published data to measure this second effect and uses it as an upper bound for its model, not as a confirmed forecast.
Putting all the pieces together, the report’s conservative scenario puts global e-waste at 196 million tonnes a year in 2050, while the aggressive one reaches 211 million. Of that total, between 31 and 46 million tonnes would come from AI-related activity.
The chart on page 31 shows how the gap widens over the years: the conventional e-waste forecast grows on its own, but BAN adds data center infrastructure and the replacement effect outside of it. The study also calculates that between 2025 and 2050, between 395 and 617 million tonnes of electronic equipment attributable to AI would be retired, depending on the scenario used.
The comparison with earlier studies has to be handled carefully. BAN argues that Wang et al., from 2024, and de Vries-Gao, from 2026, focus mainly on servers and accelerators, while its methodology folds in five infrastructure categories and the obsolescence effect outside data centers. That’s why the figures don’t represent exactly the same scope.
The study also leaves open a question it doesn’t resolve in this first part: what percentage of the retired equipment can be reused, repaired, refurbished, or sold on secondary markets. BAN says that question will be covered in the second part of its series. It’s also saving a third part to study the potential toxicity of this waste and a fourth for the measures needed to manage it.
What’s left on the table, then, is less straightforward than a “60 times more trash” headline. The report’s main contribution is widening the count: AI infrastructure isn’t just GPUs and servers. It’s also copper, transformers, batteries, cooling systems, switches, cabling, and potentially millions of devices that could fall outside new hardware requirements. The final scale will depend on how much capacity grows, how long each piece of equipment stays in service, and how much material can be kept in circulation through reuse and repair.
Frequently asked questions
How much e-waste could AI generate by 2030?
BAN estimates between 8.6 and 13.1 million tonnes a year of AI-related electronic equipment retired by 2030. The calculation includes both data center infrastructure and the replacement effect on equipment outside of it.
Why does BAN calculate more waste than earlier studies?
The report incorporates five broad categories of data center infrastructure: compute, networking, electrical distribution, storage and backup, and cooling. It also adds what it calls “AI Waste Contagion,” tied to the accelerated replacement of devices and networks outside data centers.
How much total e-waste could there be by 2050?
BAN’s model puts global e-waste generation between 196 and 211 million tonnes a year by 2050, depending on the scenario. Between 31 and 46 million tonnes would correspond to the AI-attributed component.
Is the “60 times more” figure confirmed?
No. It’s a comparison BAN makes between its own estimate and an earlier academic projection focused on a narrower scope. The report acknowledges several uncertainties, especially around how fast infrastructure will grow and how quickly devices outside data centers will be replaced.
via: The AI Waste report
