The price is on the listing. The history isn’t.
A 300-rack AI cluster was just listed on eBay for $59.7 million.
According to PC Gamer, IT parts reseller Core 4 Solutions put up a complete system: four compute nodes per rack, eight Nvidia A100 40 GB GPUs per node. That comes to about 9,600 GPUs in total, a figure PC Gamer takes from Tom's Hardware. The listing ended on 28 October, and it is unclear whether it sold.
Divide the price by the GPU count and you get roughly $6,200 per GPU, with racks, CPUs and networking included. That number is easy to calculate. What it buys is harder to know.
Same part number, different asset
Two A100s with the same part number can be very different assets. One has run cool and clean for years. Another has been remapping failing memory rows, logging Xid errors, dropping off the bus and throttling under heat. On a spec sheet they are identical. In a training job they are not: one weak GPU can stall a job that spans hundreds of others.
Some of that history stays on the card. Most of it lived in the operator's kernel logs, BMC event logs and telemetry, and it rarely travels with the hardware. When a fleet is decommissioned, the buyer gets the silicon and the seller keeps the story.
A service record for GPUs
A used car comes with a service record and a mileage reading. Retired GPUs, now reaching the market by the thousand as fleets move to newer generations, mostly come with none. Until they do, the second-hand market will keep pricing hardware by its spec sheet. Buyers will discount everything to cover the unknown, and sellers of well-kept fleets will be paid the same as everyone else.
This is the gap Vitali de Mille and I are working on at GlyphAI: a health record for each GPU, built from its own logs, that can travel with the card. More at gpu.glyphai.ie.
Source: James Bentley, "$60 million AI rack cluster listed on eBay…", PC Gamer, 29 October 2026.