The conventional account of used enterprise hardware is that it depreciates, and that the seller's task is to work out how far down the curve a given asset has slid since purchase. That account is wrong often enough to be worth abandoning.
A component does not read the depreciation schedule. It is worth what a buyer will pay for it this month, and that number can move in either direction, as anyone who has tried to price server memory through the last two years of supply disruption will confirm.
The practical consequence is that organizations pricing a decommissioned estate by applying a percentage haircut to the original invoice tend to be wrong twice. They overvalue the parts of the estate that the market has abandoned, and they undervalue the parts the market is short of. The correction is not a better haircut. It is a different procedure altogether.
What follows is that procedure, in sequence. It works for a single rack and it works for a facility. The steps are ordered deliberately, because several of them produce inputs that later steps consume, and running them out of order is the most common reason a valuation collapses under scrutiny.
Why Depreciation Is the Wrong Starting Point
Generation depreciation is the observed decline in resale value as newer server generations arrive and displace older ones in the buying market. It is real, and for complete systems it is the dominant force. But it describes a tendency across a class of assets, not the price of a specific one, and it says nothing at all about supply.
The clearest counterexample is the deliberate retirement of legacy product lines by the people who make them. In January 2026, TrendForce reported that Samsung and SK hynix were holding to plans to phase out DDR4, with DDR4 supply set to drop sharply through 2026 and per-gigabit prices pushed to record highs as a result. Older parts became more expensive, not less, because the people who could make more of them stopped.
So the starting question stops being how old the asset is and becomes three others. What exactly is this, what condition is it in, and who is short of it right now? The sequence below answers those three in that order.
Step 1: Build a Serialized Inventory Before You Estimate Anything
No valuation survives a vague inventory. Before any number is attached to anything, produce a line-by-line list in which each row is a physical object with a serial number, a model number, and a location.
At minimum, each row should carry:
- Manufacturer and full model or part number, taken from the label, not from the asset register.
- Serial number, which is what makes chain of custody and any later certificate meaningful.
- Configuration as built, meaning installed processors, memory modules by part number and count, drives by type and capacity, and any accelerators or network cards.
- Rail kits, caddies, bezels, and power supplies present or absent.
- Whether the unit powers on, and whether anyone has tested it beyond that.
The asset register will disagree with the floor. It always does. Trust the floor. A common outcome of this step alone is the discovery that memory or drives were pulled during earlier maintenance and never recorded, which changes the valuation materially before a single comp is consulted.
Data-bearing media should also be identified during inventory, before a price has been agreed. Drives and tape store data persistently. Processors, graphics cards, and memory modules do not retain user data in any persistent form, so the sanitization obligation attaches to a narrower set of rows than most inventories assume.
The current federal reference is NIST Special Publication 800-88 Revision 2, published on 26 September 2025, which superseded the 2014 Revision 1 that much of the disposition industry still cites by habit.
Step 2: Establish a Configuration Baseline Rather Than a Purchase Price
Configuration value is the principle that a used server is priced as the sum of what a buyer can do with its parts and its whole, not as a fraction of what the original purchaser paid for it. Two identically labelled chassis with different processor and memory populations are, for pricing purposes, two different products.
The inputs that move the number most, in rough order of influence:
- Processor generation and core count, usually the single largest determinant for a complete system, because it sets the workload envelope and the software licensing arithmetic for the next owner.
- Installed memory, by type, module density, and total capacity. In a tight memory market this line can dominate.
- Drive population, with NVMe and SAS solid state ahead of spinning disk in most current purchases.
- Accelerators. A populated GPU tray is frequently worth more than the host it sits in.
- Completeness. Missing rails, caddies, or a second power supply reduce the offer, because the buyer has to source them.
OEM premium describes the tendency of certain manufacturer lines, notably current-generation Dell PowerEdge, HPE ProLiant, and Supermicro, to hold value better than less liquid lines. The mechanism is the depth of the secondary buying pool and the availability of spares and firmware, not brand affection.
Step 3: Say Plainly What Works, What Is Untested, and What Is Broken
Condition grading is the assignment of each asset to a defined state that a buyer will recognize and price. The scale itself matters less than the fact that it is written down and applied consistently.
A workable three-tier scale:
- Working and testing. The unit has been powered, posted, and functionally checked, and the check is recorded against the serial number.
- Untested. The unit is physically complete and appears undamaged, but nobody has verified that it works. Expect a meaningful discount against the tested price, because the buyer is absorbing the risk.
- For parts. The unit is known to be faulty, incomplete, or damaged, and is priced on the recoverable components, not as a system.
The cosmetic condition sits alongside this scale rather than inside it. Rack rash and dusty bezels rarely change a wholesale number much. A bent chassis or a corroded backplane changes it a great deal, because both suggest handling or environmental history that a buyer cannot verify.
Grade honestly in either case: an inflated grade does not survive the buyer's own inspection, and the revised offer that follows almost always lands below what an honest grade would have produced, because the buyer now discounts the rest of the list too.
Step 4: Value the Memory Separately From the Chassis
Memory deserves its own valuation pass, and in the current market it deserves the first one. This is the step most internal estimates get wrong, because memory is treated as a specification line on a server instead of a separately traded commodity with its own supply dynamics.
In July 2026, TrendForce forecast a further increase in server DRAM contract prices through the third quarter of that year, and it noted that a server DRAM shortage is already anticipated for 2027 because RDIMM bit supply growth of 15 to 20 percent year over year is running behind server CPU shipments. No percentage haircut applied to an original invoice can represent movement of that kind.
Three technical distinctions carry most of the value difference.
What separates RDIMM, LRDIMM, and UDIMM
Registered memory places a register between the memory controller and the DRAM chips, which allows far greater capacity per channel than unbuffered memory. The module class is formally defined by JEDEC, whose DDR5 RDIMM common standard specifies the electrical and mechanical requirements for 288-pin, 1.1-volt registered modules intended for server, workstation, and database use.
Load-reduced modules go further by buffering the data lines as well, which supports the highest capacity configurations. Unbuffered modules, common in desktops and small workstations, are a different and generally less valuable product in an enterprise lot. Sorting a mixed pallet by module class before quoting is what separates a defensible number from an approximation that collapses on inspection.
Why generation is not the same as age here
DDR4 and DDR5 hold value differently from DDR3, and as the supplier exit described earlier works through the channel, some DDR4 densities have behaved less like obsolete stock and more like a constrained part. Density matters as much as generation: a single 64 gigabyte module is generally more liquid than four 16 gigabyte modules of the same generation, because it fits a wider range of target configurations.
How lots get priced
Module counts run into the thousands, and the inputs move weekly, so a memory lot is rarely worth pricing line by line in a spreadsheet that will be stale by Friday. Used RAM prices per lot, not per module, and the sorted lot goes to a specialist buyer against a written quote.
That is why the buyback program at Big Data Supply publishes no rate card at all, and prices each lot instead against its own configuration and against the market on the day it is submitted. Read that refusal correctly. It is not evasion; it is an admission that a published number would be wrong within a fortnight, and it sets the standard your own internal estimate has to meet: module class, density, count, and a date.
Any valuation that treats memory as a fixed percentage of a server's book value is not a valuation at all.
Step 5: Benchmark Against Transacted Prices, Not Asking Prices
Sold comp benchmarking is the practice of anchoring an estimate to prices at which comparable assets actually changed hands, as distinct from prices at which sellers merely hoped they would.
Public listings are a record of seller optimism. Completed transactions are a record of buyer behaviour. Anchoring to the former produces an estimate that nobody will meet, and an internal expectation that then makes every real offer look insulting.
Sources that reflect transacted value, in descending order of reliability for enterprise lots:
- Written quotes from more than one wholesale buyer for your actual serialized list. These are the strongest signals available, because they are priced against your specific configuration.
- Completed sales on public marketplaces, filtered to sold status and to matching configuration. Useful for single units, noisier for lots.
- Practitioner communities where operators discuss what they actually paid. Treat these as directional and check the date on every figure, since a price point from eighteen months ago in a market that has moved this fast is close to useless.
Two disciplines apply to all three. Date every figure you record, and record the configuration it applied to. An undated comp with no configuration attached is not evidence.
Step 6: Price the Lot as a Lot, Then Convert It Into Route Adjusted Net
Lot-level pricing is the per-unit value of a group of assets sold together, which is normally lower than the sum of the units sold individually. Sellers often experience this as unfair. It is simply the buyer pricing the work they are taking on: consolidated freight, testing at volume, holding inventory, and absorbing the units in the lot that turn out to be worth nothing.
The comparison is worth writing out in full. The figures below are placeholders chosen to make the structure visible, not market quotations, and no reader should treat them as an expected price.
Suppose your comp research supports a per-unit value of 100 units to account for a tested configuration, and you have forty of them. Sold individually, gross proceeds are 4,000. Now subtract the cost of doing that: listing and photography time, individual packing and freight, payment and marketplace fees, returns, and the storage and insurance cost of holding the estate while it sells.
Assume that it consumes 25 percent and takes five months. The individual route therefore nets roughly 3,000 over five months. A bulk buyout on the same forty units, offered at 70 each, nets 2,800 in one week, with no residual inventory and no ongoing custody risk.
That gap, small in this illustration and larger or smaller in real cases, is the actual decision, and it is a question about time, staff hours, and risk tolerance, not about who quotes the bigger headline number. Where the estate contains a long tail of low-value units, the bulk route often wins outright, because the tail costs more to sell individually than it returns.
The second half of the same step converts that gross figure into what you will actually bank. Route adjusted net is the amount received after the costs of the chosen sales channel are deducted from gross value. It is the only figure that should ever be compared across routes.
The deductions differ by route:
- Direct buyback. Deduct little or nothing at the seller's end. Freight and pickup are frequently absorbed by the buyer on bulk volume, and settlement is fast. What you give up in exchange is the retail premium, since the number quoted is a wholesale one.
- Consignment or brokerage. Deduct the commission and the cost of time, since proceeds arrive as units sell. Suited to high-value, slow-moving assets where the spread justifies the wait.
- Public marketplace. Deduct platform fees, payment processing, packaging, outbound freight, returns, and the internal labour of running the listings. Realistic for single high-value units, punishing across a large estate.
- Scrap or recycling. Appropriate for assets with no functional resale value. Priced on materials rather than on function.
Two costs belong in this calculation regardless of route: the cost of sanitizing or destroying data-bearing media to a documented standard, and the cost of the documentation itself. Both are real, both are usually forgotten in the first estimate, and an estimate that is quoted without them has been quoted incorrectly.
Where the Method Usually Breaks Down
Three failures account for most bad outcomes.
The first is starting from the invoice. Once a purchase price is in the room, every subsequent number gets measured against it, and the exercise turns into a negotiation with the past rather than an assessment of the present.
The second is waiting. Hardware sitting in a storeroom is not being preserved. It is aging against a generation curve while its warranty support lapses, and in most product categories that direction is downward, whatever the exceptional memory market is currently doing.
The third is running the steps out of order. Comps gathered before the inventory is serialized are comps for an imagined estate. Route comparisons made before condition is graded compare numbers that were never the same kind of number.
Key Takeaways
- Value memory as its own commodity, sorted by module class and density, and expect it to be quoted per lot against a market that moves weekly.
- Anchor to transacted prices with a date and a configuration attached, never to asking prices.
- A serialized inventory taken from the physical hardware, not the asset register, is what makes every other step defensible, and configuration as built is what you price against once you have it.
- Compare routes only on net proceeds after fees, freight, labour, holding time, and the cost of documented data destruction. A seller who has worked those numbers through has already set a floor, and a floor derived from method is worth more at the table than any amount of negotiating skill applied to a number nobody can defend.
