$600 Million AI Supercomputer Listed on eBay: 9,600 A100s Sold as a Bundle, Free Shipping Included

When you browse eBay, you're usually hoping to snag a used keyboard or a cheap old graphics card. But recently, one listing redefined expectations: an entire decommissioned AI supercomputer system, priced at $60 million, containing 9,600 Nvidia A100 GPUs and 4,200 AMD EPYC CPUs — and it ships free.

The news first sparked discussion on Reddit, where one user joked, "Anyone willing to chip in $60 million for me?" Jokes aside, the signals behind this deal are more intriguing than the price itself.

What Caliber of Hardware Is This?

First, the numbers: 9,600 A100s, 4,200 EPYC CPUs. The A100 is Nvidia's previous-generation flagship data center GPU. Although the H100 and H200 have taken over the baton, the A100 remains a workhorse for inference and mid-scale training tasks. At 9,600 units, this approaches the scale of a large AI cluster — not something a few servers could hold.

The accompanying EPYC CPU count is equally substantial. The GPU-to-CPU ratio is roughly 2:1, reflecting a design aimed at high-density parallel computing scenarios. The entire batch is bundled as "300 racks," indicating these are decommissioned assets pulled from a data center wholesale — racks and all.

Generational gap between A100 and H100: Nvidia's A100 is based on the Ampere architecture (released 2020), built on a 7nm process, equipped with 40GB or 80GB HBM2e memory, supporting third-generation NVLink and PCIe 4.0, with roughly 312 TFLOPS of FP16 dense compute. The H100 is based on the Hopper architecture (released 2022), built on a 4nm process, introducing the Transformer Engine and FP8 precision support, with HBM3 memory bandwidth boosted to about 3.35 TB/s and roughly 989 TFLOPS of FP16 dense compute. The efficiency gap between the two when training large models can reach several times — which is the core driver behind data centers accelerating A100 retirement. However, the A100's mature CUDA ecosystem and still-sufficient memory capacity for most inference scenarios mean it retains relatively high residual value on the secondary market.

The role of EPYC in AI clusters: AMD EPYC is a data center-oriented x86 server CPU series, primarily handling "host" responsibilities such as data preprocessing, task scheduling, I/O management, and GPU communication coordination. In GPU clusters, the CPU-to-GPU ratio directly affects overall data throughput efficiency. A GPU/CPU ratio of about 2:1 is a typical high-density AI training node configuration, meaning roughly four A100s are supported by every two CPUs, requiring sufficient memory channels and PCIe lanes to avoid bottlenecks. EPYC's high core counts (e.g., 64–96 cores in the Milan or Genoa generations) give it an advantage in multi-GPU parallel scenarios.

What Does a Decommissioned Supercomputer on eBay Signal?

The most direct interpretation is that hardware iteration is accelerating. Previous-generation flagship accelerators are going from "top-tier assets" to "secondhand goods" faster than many expected. As H100s roll out and newer architectures continue to ship, data centers holding old cards face a practical problem: the cost-effectiveness of keeping them powered on is declining.

Another angle is supply-side overflow. The scale of 300 racks far exceeds what the ordinary secondhand market can absorb. By listing it on eBay, the seller is essentially publicly seeking buyers with the capacity to take it on — possibly tier-2 cloud providers, research institutions, or startups needing low-cost compute.

The free shipping detail is rather darkly humorous. On a $60 million shipment, logistics costs are probably truly negligible, but if a deal actually closes, hidden costs like transportation, rack installation, power infrastructure retrofits, and operations adaptation could represent another substantial investment.

Depreciation and disposal logic for data center assets: AI accelerators typically have an accounting depreciation period of 3–5 years in data centers, but technological generational turnover often outpaces book depreciation. When a new generation of GPUs significantly outperforms old cards in energy efficiency (performance per watt), the electricity cost and rack-space opportunity cost of continuing to run old cards exceed their output value. Moreover, power and cooling capacity in hyperscale data centers are scarce resources — freeing up 300 racks of space for deploying H100/H200 clusters yields far higher economic returns than retaining A100s. Therefore, decommissioning does not mean broken; it means the asset has been "squeezed out" along the energy-efficiency curve. Such wholesale clearances typically go through asset disposal firms or ITAD (IT Asset Disposition) channels; listing on eBay is a relatively rare public retail-style attempt.

Who Will Low-Priced Secondhand A100s Disrupt?

If decommissioned clusters like this actually flow into the market, the first to feel pressure will likely be compute platforms whose business model centers on renting out old cards. Increased supply usually means falling rental prices — good news for budget-constrained AI developers.

But don't celebrate too soon. The whole-bundle sales format means an extremely high buyer threshold — ordinary teams can't participate in splitting it up. If this batch is ultimately acquired wholesale by a single institution, the actual circulation volume in the secondary market may not increase significantly.

Judging from the Reddit discussion atmosphere, most people treat it as conversation fodder. A $60 million price tag paired with the vibe of a clearance sale does create a sense of contrast. But the depreciation curve of the hardware world never slows down just because there are more zeros on the price tag.

Final Thoughts

This eBay listing is more like a thermometer for the industry: AI hardware refresh cycles are turning "supercomputer-class" equipment into tradeable secondhand assets. As for whether anyone will actually shell out $60 million — and enjoy free shipping to boot — only time will tell.

One-line Summary

An IT vendor is selling a decommissioned AI supercomputer with 9,600 A100s on eBay for $60 million, free shipping included.

Paragraph Summary

The article centers on a $60 million eBay listing for a decommissioned AI supercomputer containing 9,600 Nvidia A100 GPUs, 4,200 AMD EPYC CPUs, and 300 racks, with free shipping. The author first breaks down the hardware specs, noting that although the A100 has been succeeded by newer GPUs like the H100, it remains a workhorse for inference and mid-scale training, and this scale approaches a large AI cluster. The article then analyzes the signals behind it: AI hardware iteration is accelerating, data centers face the reality that keeping old cards powered on is increasingly cost-ineffective, and listing the whole system on eBay is a public search for buyers with the capacity to take it on. It also discusses how low-priced secondhand A100s could pressure compute rental platforms' pricing, but the whole-bundle sales format raises the buyer threshold, so actual secondary-market circulation may not increase significantly. Ultimately, it views this listing as a thermometer for AI hardware generational turnover.