Seller Asks $59.7 Million for 9,600 Used Nvidia A100s

Core 4 Solutions is asking $59.7 million on eBay for 300 used racks containing 9,600 Nvidia A100 processors, chips used to train and run AI. The offer allows negotiation. It records a seller’s asking price, with no completed transaction shown. Bulk listing

The advertised configuration is four compute nodes per rack, each holding eight A100 40GB SXM4 graphics processors (GPUs). That makes 32 per rack and 9,600 across the lot. Management servers, networking equipment and other components are part of the racks, so dividing the price by chip count would include more than processors.

The bulk price matches 300 individual rack offers

The same seller separately advertises one rack with the same four-node, eight-processor configuration for $199,000, also used and open to offers. Multiplying that asking price by 300 gives exactly $59.7 million. The two eBay advertisements show no apparent discount for taking the full lot at their stated prices. Single-rack listing

Core 4’s own website prices a refurbished 32-processor rack at $199,999 and advertises a one-year advanced replacement warranty. Its description says the equipment previously operated in production. Those are separate advertised terms, rather than a confirmed transaction. Seller’s product page

The eBay lot description also says remote management uses a command-line console, with no browser-based management interface or graphical remote screen. Buyers would need staff able to operate that setup.

Power and performance still need to be priced

Nvidia’s datasheet rates the 40GB SXM version at a maximum 400 watts. Across 9,600 processors, that adds up to 3.84 megawatts of GPU ratings. It excludes CPUs, networking, cooling and power-conversion losses, and is not a measurement of this cluster’s electricity consumption. Manufacturer’s specifications

Nvidia announced A100 production in May 2020. Older hardware can retain paying work: CoreWeave said in its September 30 Vera Rubin announcement that Nvidia’s 2017 Volta generation remained in commercial service on its cloud. A100 announcement CoreWeave statement

Core 4 reports that tuning one eight-A100 node increased Qwen3.8 27B chat output from 2,182 to 4,235 tokens per second, a measure of text-generation speed. Its table also reports an accuracy decline on GSM8K, a mathematics test, from 96.2% to 95.6%. These are seller-run results. Core 4 says multi-node figures are not published yet, leaving buyers of the 300-rack lot without a published whole-cluster comparison. Seller’s tests