loading
loading
IT vendor lists massive $60 million AI supercomputer on eBay with 9,600 Nvidia A100 GPUs and 4,200 AMD EPYC CPUs, plus free shipping — 300 racks of decommissioned GPUs on offer If you ever wanted to own your own supercomputer or data center while earning eBay cashback, Core 4 Solutions has the perfect opportunity. The company is selling a 300-rack AI supercomputer on eBay for $59.7 million, complete with $50 overnight shipping, or free shipping if you're willing to go with FedEx Ground. The cluster packs 9,600 Nvidia A100 accelerators and 4,200 AMD EPYC Rome 64-core processors, enough to jump-start your own sovereign AI nation. Nvidia and AMD have moved to newer architectures like Blackwell, Vera Rubin, and EPYC Turin. Nonetheless, A100 (codenamed Ampere) accelerators and EPYC Rome processors remain highly relevant for many computing tasks, including AI, machine learning, and scientific research. With the modern Blackwell and Vera Rubin racks from Nvidia, like the GB300 NVL72 and VR200 NVL72, costing up to $6.5 million and $7.8 million per unit, respectively, the A100 has turned into a legitimate budget alternative that can also circumvent the long waitlists and supply chain delays on Nvidia's newer AI hardware. The cluster's sheer specifications are truly jaw-dropping. Each rack has four compute nodes and three management nodes. Each compute node has two AMD EPYC Rome 64-core processors, eight Nvidia A100 40GB SXM4 accelerators, and 1TB of DDR4-2933 memory. The management nodes are no less impressive. Each features two EPYC 64-core chips and 192GB of DDR4 memory. When you sum it all up, the cluster comprises up to 9,600 A100 accelerators and 4,200 AMD EPYC Rome processors. With 64 Zen 2 execution cores per chip, the entire cluster boasts 268,800 cores. Memory capacity is equally impressive, with up to 1,401.6 TB or 1.4 petabytes for the entire cluster. At first glance, $59.7 million sounds like a fortune, and it certainly is, but it seems like a reasonable valuation for the cluster. A rough estimate would be about $24 million for the A100S, $3 million for the EPYC Rome chips, and about $1.1 million for the DDR4 memory. That is about $28.1 million in crucial components, before factoring in other hardware like racks, cooling, power supplies, Nvidia Mellanox ConnectX-6 SmartNICs, etc. Listing a 300-rack supercomputing cluster on eBay sounds absurd, but given today's AI-driven world, it is not as surprising as it may seem. After all, a broker once sold a 405-foot luxury Gigayacht for $168 million on eBay, so there is no reason why Core 4 Solutions cannot move a $59.7 million cluster. As AI farms and data centers expand, some enterprise clients out there will recognize the value and opportunity in the listing. Shipping that many racks will be a logistics nightmare, and as funny as it sounds, Core 4 Solutions offers free local pickup or FedEx Ground delivery directly to your home. There is even an option to pay for a $25 two-day or $50 overnight delivery. Imagine FedEx casually dropping off 300 server racks at your front porch. Get Tom's Hardware's best news and in-depth reviews, straight to your inbox. The cluster seems to have already drawn attention, with 11 people reportedly watching the eBay listing and one prospective buyer already letting it sit in their cart. I guess the real question is not whether it will sell, but whether the buyer will choose $50 overnight shipping. Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds. Zhiye Liu is a news editor, memory reviewer, and SSD tester at Tom’s Hardware. Although he loves everything that’s hardware, he has a soft spot for CPUs, GPUs, and RAM. - pf100 In the future, for the plebs (we're the plebs here), we're going to have to form gamer coalitions where we pool our money to buy a datacenter and then strip the GPU cores and make our own gpu's out of the parts.Reply - usertests pf100 said:In the future, for the plebs (we're the plebs here), we're going to have to form gamer coalitions where we pool our money to buy a datacenter and then strip the GPU cores and make our own gpu's out of the parts.On the other hand, the future is bright for gaming on 11 year old integrated graphics. I7pSE2fLjsI OzmM9dFnJMs Reply - PEnns usertests said:On the other hand, the future is bright for gaming on 11 year old integrated graphics. I7pSE2fLjsI OzmM9dFnJMs The future (the past in this case) is so bright....I gotta wear shades...😊 OEH4jaQpOt4, list: RDOEH4jaQpOt4 Reply - bit_user pf100 said:In the future, for the plebs (we're the plebs here), we're going to have to form gamer coalitions where we pool our money to buy a datacenter and then strip the GPU cores and make our own gpu's out of the parts.You can usually find plenty of these parts being sold individually, on the used market. However, the sorts of GPUs they put on SXM cards, such as the A100's, can't be used for gaming. Everything after V100 has little or none of the hardware engines needed for interactive 3D graphics. Reply - Pierce2623 Yep the A100 might be technically “ampere” but it has very little in common with the gaming cards. As you mentioned, it has no actual render backend including ROPs etc. Aside from that it’s actually a different architecture from consumer Ampere and its also fabbed on TSMC N7 instead of that god-awful hot and inefficient Samsung 8nm node that was originally based in 12nm technology.Reply - bit_user Pierce2623 said:Yep the A100 might be technically “ampere” but it has very little in common with the gaming cards. As you mentioned, it has no actual render backend including ROPs etc. Aside from that it’s actually a different architecture from consumer Ampere and its also fabbed on TSMC N7 instead of that god-awful hot and inefficient Samsung 8nm node that was originally based in 12nm technology.I agree with everything you said, except that I think Anandtech reported that one of the SMs actually had the rendering fixed-function units and it might also have a display controller. So, it could technically be used on a card with a display output, but you'd get like iGPU-level performance in games. I never heard anything about H100, Bx00, or Rubin having any of that. So, A100 might've been the end of the line for anything you could put on a proper PCIe graphics card. Reply - DS426 bit_user said:You can usually find plenty of these parts being sold individually, on the used market. However, the sorts of GPUs they put on SXM cards, such as the A100's, can't be used for gaming. Everything after V100 has little or none of the hardware engines needed for interactive 3D graphics.Ah yes, so they aren't really GPU's anymore. Seems like Compute Accelerator Unit would be a more fitting name. One of those heavy inertia things I suppose. Reply - alan.campbell99 DS426 said:Ah yes, so they aren't really GPU's anymore. Seems like Compute Accelerator Unit would be a more fitting name. One of those heavy inertia things I suppose.Indeed, I really do think they shouldn't be called GPUs, your suggestion looks like a better descriptor. One of the first questions in my head when reading this was "Hmm, so how many of the compute units are functional, how badly have they been thrashed?" Some units that are in service are surely failing, we just aren't privy to that information afaik and related piffling details like how many of these sold Nvidia chips are actually in service and energised rather than gathering dust in a warehouse or unpowered AI datacentre. Reply - bit_user DS426 said:Ah yes, so they aren't really GPU's anymore. Seems like Compute Accelerator Unit would be a more fitting name. One of those heavy inertia things I suppose.Yeah, I sometimes try to call them compute accelerators, but it just depends on the context. Too many people know them as GPUs. Even Jensen tends to refer to them like that. Then again, when we're talking about chips like those from Tenstorrent, Google's TPUs, or OpenAI's Jalapeno, those have no GPU heritage and nothing about them is suited towards graphics. So, it really makes zero sense to call them GPUs and people generally don't. Reply - bit_user alan.campbell99 said:One of the first questions in my head when reading this was "Hmm, so how many of the compute units are functional, how badly have they been thrashed?" Some units that are in service are surely failing, we just aren't privy to that informationGood points. They might be holding up just fine, but they could indeed be significantly degraded. They're not necessarily that old, since H100 shortages meant new A100s kept getting sold until as recently as like two years ago. The warranty on server hardware is usually like 5 years, but I don't know about A100's. And even if they're substantially newer than that, you can pretty much bet they've been getting run hard. On the bright side, they're probably getting dumped not because they're unstable, but just to make room for newer stuff. So, there's a good chance the vast majority of the hardware is still in good, working order. If I'm buying used datacenter gear for pennies on the dollar, that's just part of the game. But, if I'm paying like $5k per A100 card, I'd want to know that it was properly functioning when pulled. Reply
document fetched 2026-10-29 03:00 · canonical https://tomshardware.com/desktops/servers/it-vendor-lists-massive-usd60-million-ai-cluster-on-ebay-300x-racks-of-decommissioned-gpus-includes-9-600-nvidia-a100-gpus-and-4-200-amd-epyc-cpus-plus-free-shipping