A new Basel Action Network report models 395 to 617 million tonnes of AI-related electronic waste between 2025 and 2050, and argues that servers and accelerators are only 13 percent of a datacentre's equipment by weight. The interesting part for buyers is the replacement cycle it assumes.
What was announced
- The Basel Action Network report How Big Is the AI Waste Wave? estimates 395 million to 617 million tonnes of e-waste from AI infrastructure between 2025 and 2050.
- Servers and accelerators such as GPUs account for about 13 percent of a datacentre's equipment by weight, so estimates focused on them understate total waste by 40 to 60 times against widely cited academic projections.
- A reference 100 MW AI facility holds roughly 7,000 metric tonnes of equipment across five categories: networking, power distribution, backup power, servers plus accelerators, and cooling.
- A single 100 MW facility contains about 2,700 tonnes of copper in cabling and busbars alone.
- Moving from 5 to 15 kW racks to the 50 to 140 kW racks AI workloads need means much of the existing power infrastructure is ripped out rather than reused.
- Replacement cycles assumed in the model: 2 to 3 years for GPUs and accelerators, 5 to 7 years for general-purpose servers, 3 to 4 years for networking, 8 years for power distribution, 5 years for backup power and cooling.
- The report is the first of four; the second will look specifically at reuse, refurbishment and repurposing.
- The report says no hyperscaler, government or international body has published a plan for handling AI-driven e-waste at this scale.
The ETON view
Strip out the environmental framing and this is a procurement document. The modelled numbers only hold if every operator throws whole systems away on the vendor's cadence, and that is exactly the assumption an infrastructure buyer should be arguing with. A 2 to 3 year accelerator cycle with a 5 to 7 year server cycle means the chassis, memory, drives, NICs and switching around the GPU still have years of useful life when the accelerator moves on, and that hardware has a market value today.
For hosting providers, cloud platforms and enterprises the practical read is twofold. First, plan the exit at purchase: know what a platform is worth on the secondary market and who will take it, because a buyback or trade-in line changes the real cost of a refresh far more than a few percent off list. Second, do not assume a density jump forces a full rebuild. Plenty of inference, storage, backup and general compute work runs perfectly well in a 6 to 12 kW rack, which is what most UK colocation space actually delivers, and that is where current-generation and previous-generation systems are cheapest and available now rather than on a lead time.
We buy, refurbish, test and resell exactly the categories the report counts as waste, across HPE, Dell, Supermicro, Lenovo, GIGABYTE and ASUS. If you are decommissioning, tell us what is coming out before it is scrapped. If you are buying, ask us what the same workload costs on hardware one generation back with a warranty, and compare that to a new quote with a delivery date in 2027.
Related infrastructure
Category: Market · Vendor: NVIDIA · Technology: Datacentre lifecycle, high-density racks, ITAD · Last verified: 20 Sep 2026 · ~2 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
