In late October 2026, an IT hardware reseller named Core 4 Solutions listed a large decommissioned AI supercomputing cluster on eBay, priced at $59.7 million (about RMB 430 million) and accepting "best offers." The system reportedly includes 300 racks, 9,600 NVIDIA A100 GPUs and 4,200 AMD EPYC processors, and offers free FedEx Ground shipping; expedited overnight delivery is available for an additional $50. The listing quickly drew attention across the tech community.
According to reports from VideoCardz, Tom’s Hardware and other tech outlets, the cluster originated from a hyperscale cloud provider and uses the Ampere‑architecture A100 40GB SXM4 accelerator introduced in 2020. Each rack is said to house four compute nodes, with each node hosting eight A100 GPUs (interconnected via NVSwitch), two 64‑core AMD EPYC Rome processors, 1TB of DDR4 memory, and multiple NVIDIA ConnectX‑6 200Gb/s network adapters. In total the system comprises roughly 268,800 Zen 2 CPU cores, 384TB of GPU memory and about 1.4PB of system memory.
Based on the listing price, the blended cost per A100 comes to about $6,219, a figure that includes CPUs, memory, networking equipment and rack infrastructure. The seller also offered split‑sale options, such as rack units with 32 GPUs priced at $199,000. Notably, the complete bulk listing was later marked as "no longer available," and there is currently no public confirmation whether a sale was completed.
Hardware specifications and market positioning
The A100, NVIDIA’s previous‑generation flagship data‑center GPU, supported a large share of AI training and inference workloads from 2020 through 2023. Its 40GB HBM2e memory and high compute density were industry‑leading at the time. However, with the introduction of Hopper (H100) and Blackwell architectures, the A100 has gradually been displaced from first‑tier training clusters and is increasingly used for inference, research or cost‑sensitive deployments.
In scale terms, this 9,600‑A100 cluster surpasses the 6,080‑A100 configuration of Meta’s early AI Research SuperCluster (RSC) first phase. For small and medium AI startups, research institutions or secondary cloud providers, purchasing decommissioned hardware in bulk could theoretically reduce upfront capital expenditure significantly. Some analyses estimate the secondhand market value of the core components—GPUs, CPUs and memory—at roughly $28 million, with the remaining price gap attributable to racks, power, networking and integration costs.
Practical challenges are considerable. Secondhand data‑center hardware often lacks full factory warranty and requires specialist teams for redeployment, cooling optimization, network reconfiguration and software‑stack adaptation. Moreover, the A100’s power efficiency and interconnect bandwidth lag newer products, so per‑unit compute cost for large‑scale training may not be competitive in the long term. Additionally, logistics for transporting, installing and provisioning power for hyperscale clusters are complex operations; even if "shipping is included," on‑site deployment costs remain substantial.
Industry context and implications
Amid surging AI compute demand, constrained supply and high prices for next‑generation GPUs have prompted some organizations to pay greater attention to the secondhand and decommissioned hardware market. While such listings are uncommon, they illustrate the accelerating turnover cycle of compute hardware: cloud providers typically refresh large portions of their fleets within three to five years to stay at the cutting edge, and retired equipment then flows to secondary markets.
Objectively, these transactions offer buyers a lower‑cost entry to significant compute capacity and enable sellers to recover residual asset value. They also highlight structural features of the AI infrastructure market—leading‑edge compute remains concentrated among a few cloud giants, leaving smaller participants to either bear the high cost of new hardware or seek compromises in the used market. In many cases, software ecosystem support, driver updates, security patches and long‑term operations capability determine practical outcomes more than raw hardware alone.
At present, Core 4 Solutions’ listing has ended and the transaction outcome has not been disclosed. Whether sold or not, the episode vividly illustrates one facet of today’s AI hardware market: as new‑generation chips continue to arrive, previous‑generation flagship devices can still appear on public marketplaces at astonishing scale, offering a distinctive window into compute supply, demand and cost structure.
▶️ Enjoy this article? Creating content is hard work!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com