eBay lists a $59.7M used AI supercomputer: 9,600 A100 GPUs, 300 racks, free shipping
An eBay seller is asking $59.7M for a decommissioned AI supercomputer: 300 racks, 9,600 A100 40GB GPUs, 4,200 EPYC CPUs — about double the parts value.
A used AI supercomputer is for sale on eBay at a $59.7 million asking price: the seller, Core 4 Solutions, is an IT hardware reseller based in Minneapolis, and the lot is a complete 300-rack cluster with 9,600 Nvidia A100 40GB SXM4 accelerators and 4,200 AMD EPYC Rome CPUs — with free FedEx Ground shipping. Tom’s Hardware reported the listing (item 880041409723) on October 28, a day after a tweet put it in circulation; VideoCardz, TweakTown and TechRadar have since checked the same bill of materials.
Key points
- Price and shipping: $59.7 million or best offer; local pickup or FedEx Ground is free, two-day delivery costs $25 and overnight $50.
- The hardware: 300 racks with 9,600 A100 40GB SXM4 GPUs and 4,200 64-core EPYC Rome CPUs — 268,800 CPU cores, roughly 1.4PB of DDR4-2933 and 384TB of GPU memory in total.
- The premium: by Tom’s Hardware’s estimate the key components are worth about $28.1 million (roughly $24M GPUs, $3M CPUs, $1.1M memory), so the ask sits at about twice the parts value.
Background
The A100 is Nvidia’s Ampere-generation data center GPU from 2020, the card that carried most large-model training until Hopper and then Blackwell displaced it. Per VideoCardz and TweakTown, this fleet came out of a decommissioned hyperscaler environment — a whole data center floor being cleared at once, not a pallet of loose cards. The listing surfaced in a tweet on October 27 and became the hardware world’s talking point after Tom’s Hardware picked it up the next day; the outlet also noted that eBay’s precedent for big-ticket items is a $168 million gigayacht that once sold on the platform.
The facts
The eBay supercomputer listing’s spec sheet, as transcribed and cross-checked by four outlets between October 28 and November 4:
| Item | Detail |
|---|---|
| Asking price | $59.7M, best offer accepted |
| GPUs | 9,600 A100 40GB SXM4, 384TB of GPU memory |
| CPUs | 4,200 64-core EPYC Rome (Zen 2), 268,800 cores |
| Memory | about 1,401.6TB of DDR4-2933 |
| Per rack | 4 compute nodes plus 3 management nodes; each compute node has 8 A100s on NVSwitch, 2 64-core EPYCs, 1TB DDR4, 8 ConnectX-6 200Gb/s NICs and an FPGA |
| Single-rack option | $199,000 per rack (seen by VideoCardz) |
| Seller | Core 4 Solutions, IT hardware reseller, Minneapolis |
| Origin | decommissioned hyperscaler environment; original owner unnamed |
| Logistics | free pickup or FedEx Ground; $25 two-day, $50 overnight |
Two figures should not be conflated: $59.7 million is the seller’s ask, and the $28.1 million parts estimate is Tom’s Hardware’s back-of-envelope math excluding racks, power, cooling and networking gear. Neither is a transaction price.
What others say
Tom’s Hardware puts the price tag against the current market: new GB300 NVL72 racks reportedly reach about $6.5 million each and VR200 NVL72 racks about $7.8 million, which makes a complete A100 cluster look like a budget alternative. VideoCardz runs the per-unit math instead: $59.7M spread over 9,600 cards is about $6,219 per A100 — and that price also covers every CPU, NIC, power supply and rack — while the same money buys roughly 29,865 RTX 5090s at MSRP. TechRadar warns anyone serious to verify condition, power draw, software compatibility and maintenance support, noting that receiving 300 racks is a logistics project of its own. The outlets disagree on one point — whether you can still buy it: VideoCardz and TweakTown saw the listing marked “no longer available” on October 29, while TechRadar’s November 4 write-up counted 11 watchers and one shopper with the lot in a cart. No sale has been confirmed.
Our take
The audience for this listing is not really a buyer; it is anyone budgeting for AI compute, because it hangs a public anchor on what decommissioned Ampere hardware is worth. The Blackwell transition is pushing whole A100 fleets onto the secondary market, and the cards still have years of useful work in inference, fine-tuning and scientific computing. Demand is real on the other side too — A100s still move through smuggling routes — which says the secondary market pays well above scrap value. Memory tells the same story: DDR4 contract pricing has been climbing through the shortage, so those 1.4PB of DDR4-2933 are appreciating assets. The two-times premium over parts value buys a wired, integrated, rack-level install — a real saving for a small team that cannot assemble one. The costs land after delivery: data center power, maintenance contracts, spares for out-of-production 40GB HBM2, and a CUDA stack now two generations behind. Whether that trade makes sense depends on how many years you plan to run the system.
What to watch
- Whether it sells, and at what price after Best Offer bargaining — the actual transaction is the real residual-value signal for used A100s; none confirmed as of publication.
- Listing status: marked unavailable on October 29, active again in November 4 reporting — if the bulk deal stalls, the $199,000-per-rack channel is worth watching.
- The next clearance: as Blackwell ramps, more hyperscalers will retire Ampere fleets; the discount rate on the next full-cluster listing is the number to track.