17 kg per GPU: the e-waste wave of 2030

ecogpt.coPublished Oct 23, 2026 · 2 min read

A Nature Computational Science study projects generative AI will produce up to 2.5 million tonnes of e-waste a year by 2030, a thousandfold increase on 2023. The index turns that into a figure per company: tonnes of servers that will be retired within five years.

Wang and colleagues modelled the server fleet needed for generative AI using the DGX H100 as the benchmark unit.
The high-adoption scenario reaches 2.5 million tonnes a year by 2030, including 1.5 million tonnes of circuit boards.
A DGX H100 weighs 130 kg for eight GPUs; a GB200 NVL72 rack about 1,360 kg for 72. The index uses 17 kg per accelerator.
Circular measures, longer lives and component reuse, could cut the total by up to 86%.

The study

In October 2024 Peng Wang and colleagues published a projection of generative AI's e-waste in Nature Computational Science. Their method was to model how many servers the sector would need through 2030 and when those servers would be retired, using NVIDIA's DGX H100 as the reference unit. The result: annual e-waste rising from about 2,600 tonnes in 2023 to as much as 2.5 million tonnes a year by 2030, a thousandfold increase, with a cumulative 1.2 to 5 million tonnes depending on how fast adoption runs. The aggressive scenario includes 1.5 million tonnes of printed circuit boards and 0.5 million tonnes of batteries, both containing hazardous materials.

UC Riverside's Shaolei Ren, who was not involved, called the DGX-based projection a reasonable gauge, while noting that predicting future hardware is hard.

The weight of a GPU

NVIDIA's user guide lists the DGX H100 at 130.45 kg for a system containing eight GPUs, about 16.3 kg each including its share of chassis, CPUs, memory, storage and cooling. A fully populated GB200 NVL72 rack weighs about 1,360 kg for 72 GPUs, or 18.9 kg each. The index uses 17 kg of server per accelerator and multiplies by each company's fleet to estimate the hardware that will become waste within roughly five years: about 87,000 tonnes for Google's fleet, 60,000 for Microsoft's, 300,000 across the index.

Why it will be waste

AI accelerators are retired fast. Training economics favour the newest chip; a three-year-old GPU is worth a fraction of its cost and costs the same electricity to run. Unlike a conventional server, which may serve a decade in a lesser role, a frontier GPU's second life is short. The study's authors found lifespan extension and module reuse the most effective mitigations, cutting the projected total by 16 to 86%.

Who reports anything

Two companies in the index. Microsoft reports that 90.9% of its retired servers and components were reused or recycled in 2024, through six Circular Centers that handled 3.2 million components. Google reports harvesting about 8.8 million components from decommissioned hardware in 2024 for reuse or resale. Neither publishes a tonnage, and no one else publishes a rate. The ninth item on the index's disclosure checklist, hardware reuse or e-waste reported, is the least met.

The number on the profile

Each profile's hardware block shows "future e-waste" in tonnes, labelled derived, with an everyday comparison: Google's 87,000 tonnes is about 480 empty Boeing 747s by weight. The figure is an estimate built from an estimate. It is also, for 28 of 30 companies, the only one that exists.