After the AI Upgrade: Which Servers Deserve a Second Life?

Imagine a refurbishment business offered a pallet of servers that an AI company no longer wants. Some machines still work. Some are missing parts. All of them take up space. The seller wants a collection date; a prospective buyer wants to know what the machines can run, how much power they will draw and what happens if one arrives faulty. There is a business opportunity somewhere between those requests, but it is not the same thing as a warehouse full of cheap hardware.

A server leaves its first job for a reason. It might be too slow or costly for the next workload, or awkward to maintain in a fleet that is moving to a different design. Those reasons matter to the second buyer. The commercial question is whether there is a specific job the equipment can still do well enough to justify testing it, moving it and keeping it in service.

A waste projection is a reason to investigate

In an October 2024 study in Nature Computational Science, Peng Wang and colleagues modelled the electronic waste associated with generative AI. Their estimates ranged from 1.2 million to 5.0 million tonnes cumulatively over 2020–2030 across different development scenarios. These are modelled totals for a particular stream of waste. They are neither a count of equipment already discarded nor a forecast of how much a refurbisher can sell. A large waste problem can coexist with a small pool of machines suitable for a given buyer.

The Open Compute Project’s circularity workstream distinguishes reuse of complete equipment from recovering parts or recycling materials. That gives a buyer or owner a practical starting point. Inspect the machine before picking its destination. A complete working server with a supported configuration may have a second job. Another unit might make more sense as a source of compatible spares. A damaged assembly with no viable use belongs in a different channel. Each route calls for different evidence and different work.

Sell a tested capability rather than an attractive specification

Suppose a small studio wants an internal machine for occasional rendering and a supplier offers a used compute server. A list of processors and memory gives the studio a starting point. It does not tell the buyer whether the machine can fit the available space, whether its fans are tolerable or whether the studio can maintain its software. Ask for a representative job to be run on the proposed configuration. Record whether it completes, how long it takes and what it draws at the wall during the task and when idle. Compare with the studio’s actual alternative, which might be an existing workstation or rented compute when a project arrives.

Do that comparison before calling reuse the greener choice. Keeping a working machine in service can avoid buying another machine if the used one actually replaces that purchase. But a server that sits powered on all week to run one short job has a different operating profile from a machine used intensively. The expected remaining life matters too. A few good weeks are poor compensation for a machine that cannot be repaired when a part fails. A used device can be the right choice for one workload and a poor one for another.

For a refurbisher, the product may be the evidence as much as the hardware. A buyer can compare offers when the same sheet identifies the exact configuration, condition, test workload, observed power draw, missing parts and support terms. Clarify which measurements apply to the individual unit and which come from a representative sample. If the machine contains storage, document how it was handled. The Open Compute Project’s sustainability resources include data sanitisation among its areas of work. Agree who owns that step and what record the next owner receives.

Try a small batch before buying a full lot

• Start with a buyer and a job. Find out what the buyer needs to run, what would make the equipment unusable and who will maintain it after delivery.
• Separate the routes in advance. Identify candidates for whole-unit reuse, parts recovery and recycling. Keep a record of where each unit goes, including ones that never sell.
• Count the work between collection and sale. Include testing time, replacement parts, storage, transport and some allowance for returns. A low purchase price tells you little about those costs.
• Ask what the next owner would otherwise do. Check whether the machine displaces a new purchase, replaces other used equipment or simply adds capacity. Those outcomes support different claims about the benefit.

MaryChuks.com’s coverage of data centres as compute infrastructure looks at the investment required to build capacity. The other end of that investment deserves the same scrutiny. A second-hand market is useful when it can turn a retired asset into equipment that someone can put to work, with a credible account of what happens to everything else.

For the business offered that pallet, the next sensible move is to ask for a small, representative batch. Test it against a real buyer’s needs and keep track of the units that fail. If there is a repeatable business here, it will show up in those records. The size of the global waste projection cannot do that work for you.

Sources

• Peng Wang and colleagues, E-waste challenges of generative artificial intelligence, Nature Computational Science (October 2024).
• Open Compute Project, Circularity workstream and Sustainability Project resources.