How to Sell Decommissioned Servers: Where the Value Actually Hides

By the SellUsedRAM bench · Reviewed July 28, 2026 · About SellUsedRAM

A decommissioned server is worth roughly the sum of four components — CPUs, memory, drives, and network cards — and very little else. The chassis, rails, fans, and power supplies are heavy, cheap, and freight-sensitive; the value you are trying to recover is concentrated in parts that fit in a flat-rate box. This guide takes a retired node apart in the order the money actually sits, so you can look at a rack you are about to retire and estimate its recovery value before anyone quotes you. Every figure here comes live from the Payout Index — the prices we pay, published in full and updated weekly, so a number in this paragraph and a number in a quote cannot disagree.

The chassis is not the asset

The instinct on a decommissioning project is to price the server: a 2U box with a service tag, a line in the asset register, a purchase price everyone remembers. The used market does not think that way. Complete-system demand exists, but it is narrow, it is generation-sensitive, and it competes with the cost of moving 60 pounds of steel across the country. Components are the opposite: dense, light, individually identifiable, and shopped for by refurbishers, integrators, and homelab builders who know exactly which part number they need.

So the first move on any retirement project is to stop asking "what is this server worth" and start asking "what is in this server worth." The gap between those two questions is usually the difference between a recycling invoice and a check.

Memory: dense, liquid, and mid-supercycle

The 2025–26 DRAM supercycle repriced everything. DDR4 is up roughly 60–80% as production winds down, and DDR5 has more than doubled — which means the modules in a four-year-old fleet may be worth more today than the quarter you bought them in. High-capacity registered modules are the most liquid used memory there is: a 64GB DDR5-4800 RDIMM pays $961–$1,174, a 32GB DDR4-3200 RDIMM pays $103–$120, and the Cascade Lake-era workhorse, a 64GB DDR4-2933 RDIMM, pays $165–$192.

Multiply by slot count and memory becomes the line that funds the project. It is also the line most likely to be thrown away by accident, because DIMMs look like nothing and weigh nothing. Across the nodes we quote, memory typically accounts for about 15–25% of recoverable value — a large share, but not the whole story, which is exactly the misconception the next two sections exist to correct.

Pulling and sorting the modules has its own process — BMC inventory first, sort by full part number, photograph the stacks — and it is covered step by step in how to sell RAM from decommissioned servers. This guide stays at the level of the whole node.

CPUs: the most bimodal line on the manifest

Processors are where estimates go wrong in both directions. Recent, high-core-count parts hold real money: an EPYC 7763 pays $656–$818 and frequently exceeds the value of a single stick of anything else in the same chassis. Meanwhile the mid-tier Skylake and Cascade Lake Xeons that filled the world's racks from 2017 to 2020 have collapsed under their own supply — a Xeon Gold 6338 pays $168–$219, and older Gold 6130-class parts are worth single dollars each in any quantity you care to send.

The practical rule: check the exact model before assuming either way. Two CPUs that look identical in a photo, in the same socket, from the same vendor, can differ by two orders of magnitude in resale. A pile of 6130s is a recycling line item. A tray of Milan or Ice Lake parts is a meaningful check. Read the model off the integrated heat spreader, or pull it from the BMC without opening anything.

Storage: usually the largest single line

Enterprise SSDs are frequently the biggest number on a decommissioning manifest, and the reason is endurance. Datacenter drives are rated in drive writes per day, and typical virtualization and storage workloads consume a fraction of that budget, so a drive that retired after four years of service still has most of its engineered life left. A Samsung PM9A3 3.84TB NVMe pays $352–$440, and even the SATA workhorses hold up — a PM893 1.92TB pays $138–$172. Multiply by 8, 12, or 24 bays and the storage line frequently clears everything else on the node combined.

Before you quote them, run smartctl -a and note the percentage used and the reallocated sector count; that one number decides the grade, and knowing it up front means the bench finds what your description already said. The full endurance explainer is in enterprise SSD endurance explained, and the grade thresholds themselves are published on how we grade rather than negotiated after arrival.

Spinning disks are a different conversation: high-capacity nearline drives retain modest value, small or old ones do not, and drives with SMART warnings are declined rather than quoted low. Every data-bearing device is handled the same way regardless of value — see the data section below.

Network cards: small, overlooked, occasionally excellent

NICs are the line most often left in the chassis on its way to the recycler, which is a shame at the top of the range and irrelevant at the bottom. A Mellanox ConnectX-5 25G dual-port pays $52–$65, and a modern Intel E810-XXVDA2 pays $66–$86. Higher-speed ConnectX-6 parts run well past that. At the other end, the ubiquitous Intel X710-DA2 is worth about as much as the postage, because every decommissioned rack in the country is shedding them at once.

One practical note: OCP mezzanine cards are vendor-keyed and sell into a much thinner market than standard PCIe cards. Pull both, but expect the PCIe cards to carry the line.

GPUs are quoted by hand

Accelerators do not have an index row, and that is deliberate. The market moves faster than a weekly reprice can honestly track, configurations vary in ways a slug cannot capture, and a wrong published number on a card worth thousands is worse than no number at all. Send the model, the memory configuration, and a photo of the label to the GPU desk and you get a hand-priced firm number in the same four business hours as everything else.

The 12-DIMM sled tell

You can read a node's generation off its memory layout before you power anything on, which matters because generation drives most of the value spread above. Twelve populated DIMM slots per socket is an EPYC Genoa signature — twelve memory channels per socket is that platform's defining feature, and nothing else in the modern server world is laid out that way. Eight per socket points at EPYC Rome or Milan, or at Xeon Scalable. Six per socket is older Xeon E5 territory and is usually a recycling conversation with a few salvageable drives.

The reason to care: a 12-DIMM sled is almost always populated with high-capacity DDR5 RDIMMs, which is the densest concentration of memory value in any chassis you will retire this decade. If a decommissioning project includes sleds like that, the memory line deserves its own count and its own photographs before anything is racked out.

What a buyer actually needs from you

A serious quote needs identification, not paperwork. Per line: what it is, how many, and one piece of evidence. That evidence can be any of these, in descending order of convenience for you:

• Part numbers — exact manufacturer part numbers for memory and drives, exact model numbers for CPUs and NICs. Every business server can export this from its BMC (iDRAC, iLO, XClarity, IPMI) without a screwdriver, and sudo dmidecode --type memory does the same from a shell.
• Label photos — genuinely fine, and often faster. A clear photo of a module label, a drive label, or a CPU heat spreader carries everything a part number does. Email them to quotes@sellusedram.com with the counts; a photo of a stack of identical sticks with the count written in the message is a complete line item.
• An asset export — a spreadsheet from your asset system, or a rack elevation. Imperfect is fine; reconcilable is what matters, because intake photographs every label and reconciles line by line against your list either way.

What nobody needs is a polished manifest before a conversation can start. Counts and part numbers are enough for a firm number; the formal document work belongs to the business desk, and it happens after you know what the lot is worth, not before.

Chain of custody and the data question, answered correctly

Memory is volatile. DRAM loses its contents within moments of power-off, NIST 800-88 treats powered-down volatile memory as sanitized, and there is no wipe step and no certificate for a stick of RAM — the physics, and the one exception for persistent memory modules, are in is it safe to sell used RAM. Your compliance attention belongs entirely to the drives.

Those get NIST 800-88 Purge sanitization on the bench, verified per device, with a serialized certificate emailed free for every drive — each carrying a public verification URL under /cert/ so an auditor can confirm it independently instead of taking a forwarded PDF on faith. Anything that fails to sanitize is physically destroyed rather than quietly passed along. Intake is photographed serial by serial, so the chain of custody your auditor asks about is a record that already exists rather than a promise. The full policy sits on the data security page.

How payment works

Send part numbers or photos and you get a firm quote within four business hours; a full manifest for a rack-scale project comes back within one business day. The quote is binding for seven days — 72 hours on the most volatile SKUs — and the lock only works in your favor: if the market drops after we quote, we pay the quote anyway.

Shipping is free on a prepaid label, and orders above $500 get a box and packing supplies sent out first, so nothing arrives in a cereal carton. Testing happens promptly on arrival, payment goes out within 48 hours of testing by Zelle, PayPal, or check, and the order minimum is $50. Volume moves the number: 8 or more identical units add 3%, 32 or more add 6%. If testing disagrees with the description you get photo and test evidence, a revised offer, and free return shipping of everything if you decline — never a silent adjustment.

Logistics scale with the project. Mail-in works in all 50 states. On-site decommissioning and pickup — de-installation, serialized capture at the rack, freight — run across New England, and are described on the data center decommissioning page.

Whole or parted: the actual decision

Part it out when the fleet is mixed-generation, when the memory is high-capacity DDR4 or DDR5, when the CPUs are recent, when the bays are full of enterprise SSDs, or when freight on complete chassis would eat the margin — which is most of the time, and increasingly so the older the sheet metal gets.

Sell whole when you have a run of identical, complete, current-generation nodes and a buyer who wants running systems; when the systems are still under a transferable support contract; or when your own labor is the binding constraint and the pull-and-sort hours cost more than the spread. That last case is real and worth saying out loud: a two-person IT team with a hard move-out date is often better off selling complete nodes at a lower unit price than spending three weekends with a screwdriver.

You do not have to guess which side you are on. Send the configuration and ask for both numbers — component-level and whole-unit — and pick the one that suits the constraint you actually have.

FAQ

What is a decommissioned server actually worth?

The sum of its components, which for a mid-life node means CPUs plus drives plus memory plus network cards, with memory usually landing around 15–25% of the total. Check three of your part numbers against the Payout Index before assuming anything is scrap; the spread between generations is far wider than most retirement projects assume.

Do I have to pull everything apart before getting a quote?

No. Quote first from the BMC inventory or photos, then pull once you know what the lot is worth and whether you are selling components or complete units. Disassembling a rack before you know the answer is how good parts end up in a scrap bin.

Do you buy the whole rack, or only the profitable parts?

The whole rack. The parts with resale value are bought and paid for; the remainder is recycled through documented downstream partners, with the recycling bill typically smaller than the check. Cherry-picking the good parts and leaving you with a disposal problem is not a service.

Should we wipe the drives before shipping them?

You are welcome to, and every drive is re-sanitized and certified to NIST 800-88 on our bench regardless — so a drive that arrives already wiped simply gets the same treatment twice. What matters is that no drive leaves your custody without a documented destination, which is what the serialized certificates are for.

How long does a rack-scale quote take?

Four business hours for a component list, one business day for a full manifest. If a project needs an on-site scoping visit first, that is scheduled through the business desk and does not delay pricing on the components you can already list.

Export the inventory, photograph the labels you cannot export, and paste the whole thing into the quote form — the rack gets a firm number before the end of the day.