An IT hardware reseller put a whole AI supercomputer on eBay this week. Minneapolis-based Core 4 Solutions listed 300 decommissioned server racks holding 9,600 Nvidia A100 GPUs for $59.7 million, "or Best Offer," with free FedEx Ground shipping, Tom's Hardware reported. The eBay listing has since ended, marked by the seller as no longer available. As TweakTown and VideoCardz note, there is no confirmation that a sale took place.
The free shipping got the jokes. The price is the more useful part. It is a public, itemized asking price for a large block of six-year-old AI hardware, at a time when investors are arguing about how fast GPUs lose their value.
What was in the listing
The seller notes on the eBay listing describe each rack this way:
- 4 compute nodes per rack. Each node has 8 Nvidia A100 40GB SXM4 GPUs with NVSwitch, 8 Nvidia ConnectX-6 200Gb/s network cards, 2 AMD EPYC Rome 64-core CPUs, 1 FPGA and 1TB of DDR4-2933 memory.
- 3 1U management nodes per rack, each with 2 more AMD EPYC Rome 64-core CPUs and 192GB of memory.
Tom's Hardware adds it up to 9,600 A100s, 4,200 EPYC CPUs, 268,800 CPU cores and about 1.4 petabytes of system memory. VideoCardz puts GPU memory at 384TB and says the hardware came from a decommissioned hyperscaler environment. The listing does not name the original owner.
The price per GPU
Divide $59.7 million by 9,600 and you get about $6,219 per A100, as TweakTown points out. That figure also pays for every CPU, network card, power supply and rack, so the GPU share is lower.
The bulk price was not a bulk discount. SnowRock found the same seller lists a single 32-GPU rack on eBay for $199,000, and 300 times $199,000 is exactly $59.7 million. Core 4's own product page prices a refurbished rack of the same layout at $199,999, shows 363 units in stock, and offers a one-year advanced replacement warranty.
Tom's Hardware made a rough parts estimate: about $24 million for the A100s, $3 million for the EPYC CPUs and $1.1 million for the memory, or about $28.1 million in core components before racks, networking and power. On that estimate, the asking price is roughly twice the value of the main chips.
For scale, Tom's Hardware says current Nvidia GB300 NVL72 racks cost up to $6.5 million each and Vera Rubin VR200 NVL72 racks up to $7.8 million. At those figures, $59.7 million is about the price of nine GB300 racks.
Why a six-year-old GPU still has a price
Nvidia announced the A100 in 2020. On paper, a chip that old should be close to worthless in a field that now gets a new Nvidia architecture about every year. In practice it still earns money, and that is the core of the GPU depreciation debate.
In August, CoreWeave CEO Mike Intrator said the company had signed an A100 contract that runs into 2029, Tom's Hardware reported. He said "pricing for prior generation SKUs is at or above where it was years ago." The same report says Nvidia CFO Colette Kress has said A100s sold six years ago still run at full utilization. On the other side, investor Michael Burry has accused hyperscalers of understating depreciation by $176 billion between 2026 and 2028 by stretching GPU useful-life assumptions to five or six years.
Power is a big part of the answer. Per Tom's Hardware, an air-cooled DGX A100 system draws 6.5kW at maximum load and fits in older data halls built for about 20kW per rack. Nvidia's GB200 and GB300 NVL72 racks draw 120kW to 140kW and need direct-to-chip liquid cooling. New chips cannot simply move into those older halls, so A100s keep them busy.
The eBay cluster is not small by that measure. Nvidia's A100 datasheet rates the 40GB SXM part at a maximum 400W. Across 9,600 GPUs that is 3.84 megawatts of GPU ratings alone, as SnowRock notes, before CPUs, networking and cooling.
What a buyer would get
Core 4 sells the racks for training and inference, but its published numbers are about serving models. On its product page, the company says its own tuning of one 8-GPU node raised Qwen3.8 27B chat output from 2,182 to 4,235 tokens per second, while GSM8K math accuracy went from 96.2% to 95.6%. These are seller-run tests on a single node. SnowRock points out that no whole-cluster results are published, and that remote management in the lot uses a command-line console with no browser-based interface.
Memory is one limit. Core 4's own list marks MiniMax-M3 428B as "memory-bound on 40GB" cards.
Why it matters now
Used-GPU prices feed into the money behind AI infrastructure. Per Tom's Hardware, concerns that faster GPU obsolescence could unravel AI financing go back years. That question is live this week. CNBC reports that CoreWeave shares have fallen 30% in October amid scrutiny of neocloud debt, and Reuters says tightening financing terms have tested the chip trade over the past month.
One eBay asking price does not settle that debate. It is not a sale, and Core 4 sets its own prices. But it is a public number for 2020 hardware, and it is about twice Tom's Hardware's estimate of the core parts. If a buyer pays near that, it is one more data point that old AI chips hold value longer than a strict yearly upgrade cycle suggests.
Sources
- eBay listing 880041409723 (ended)
- Tom's Hardware: IT vendor lists $60 million AI supercomputer on eBay
- TweakTown: An entire Nvidia A100 AI cluster hit eBay
- VideoCardz: Decommissioned Nvidia supercomputer hits eBay
- SnowRock: Seller asks $59.7 million for 9,600 used Nvidia A100s
- Core 4 Solutions: A100 40GB GPU rack product page
- Nvidia A100 datasheet (PDF)
- Tom's Hardware: CoreWeave signs A100 contract into 2029
- CNBC: Stock market today, October 29, 2026
- Reuters via Investing.com: Amazon burns cash, Intel flags weak PC demand
