CARDINALA useful used-GPU price reference compares the same accelerator configuration, seller-described condition and offer scope. It also identifies the currency, tax treatment and observation date. An asking-price range helps frame a sourcing question; it does not establish the price of a completed sale or your final offer.
Identify the hardware behind the number
Start with the exact variant. A family name can cover different memory capacities or physical configurations. Comparing prices before resolving that identity creates a misleading result, even when every displayed amount has been copied correctly.
NVIDIA publishes distinct configurations within the H100 family. A reference should therefore identify the relevant form factor and memory configuration before grouping observations. Use the manufacturer's document to check the seller's part description. Source: NVIDIA H100 specifications
Record the manufacturer part number when available. Keep OEM-specific identifiers alongside the underlying accelerator variant rather than discarding them. If an identifier cannot be mapped confidently, leave the observation out of the published cohort until someone resolves it.
A search result is only a candidate. GPU searches can include replacement housings, damaged boards and complete servers. Inspect the item description and the unit of sale. A low number attached to the right family name may price an entirely different item.
The A100 PCIe 80GB reference, H100 PCIe 80GB reference and L40S PCIe 48GB reference provide exact configurations to use in a sourcing brief. They do not establish the condition of a particular offered unit.
Keep asking prices separate from completed sales
An asking price is what a seller currently requests. A negotiated offer can differ. A completed-sale record needs evidence of the actual transaction amount and its scope. Do not relabel an asking-price dataset as sales data because some listings later disappear.
Auction bids and rental rates belong in separate datasets. A current bid can change before an auction ends. A rental rate prices access over time under its own conditions. Neither is automatically comparable with buying a physical accelerator.
eBay describes its Product Research tool as including actual sold prices, including accepted Best Offer amounts. That distinction matters: a listing's visible asking amount and a verified accepted amount are different evidence fields. Source: eBay Product Research description
Ask a reference publisher which type of evidence supports the chart. If the answer combines listings, estimated ranges and transactions without separating them, the resulting number is difficult to interpret. A precise label is more valuable than an unexplained claim of accuracy.
Preserve the seller's condition statement
Used, seller-refurbished and unused-surplus offers should remain separate unless the methodology explains a defensible comparison. The condition label is part of the observation. It is not proof that every seller follows the same process.
Request the basis of the seller's description and the available evidence. For a refurbished offer, ask what work was performed and which records apply to the actual units. Keep warranty-provider and coverage questions with the offer instead of assuming a universal refurbishment premium.
A public BrightStar H100 listing, for example, distinguishes seller-refurbished condition and seller-specific warranty labels. Those are attributes of that offer. They should not become claims about all H100 cards or Cardinal's own terms. Source: BrightStar H100 offer
The seller evidence guide explains how to connect diagnostic records to unit identifiers. Pricing methodology should preserve those distinctions, while technical review determines what the records actually establish.
Normalize the unit without inventing a component value
If a price clearly covers several equivalent accelerators without a mixed bundle, divide the hardware amount by the accelerator count. Keep the original amount and quantity so another reviewer can reproduce the calculation.
per_accelerator_asking_price = comparable_hardware_amount
÷ accelerator_quantityDo not divide the total price of a complete server by its GPU count and call the result a card price. That bundle also contains other components and may include services. Compare the complete system separately unless the seller itemizes a defensible accelerator amount.
Keep shipping outside a hardware-only reference and state that boundary clearly. Record the original currency and tax treatment. If those fields are unknown, retain the record for review rather than silently assuming they match the other observations.
Supplier policies can make tax treatment depend on the destination and applicable documentation. Server Supply's published terms illustrate why that field needs explicit review. A storefront amount alone may not explain the buyer's final tax position. Source: Server Supply sales policies
Record when the offer was actually checked
Distinguish the date shown on a page from the time someone verified the observation. A search engine may return an earlier version. A product page can remain online after availability changes. Neither a recent search result nor a visible price proves current supply.
During our source review, an eBay A100 page displayed an asking amount while also stating that the item was out of stock. That record cannot support an active asking-price cohort without a later verified change in status. Source: reviewed eBay A100 listing
Keep an observation's status history. If a newer record marks an offer inactive, it should retire the older active record. Simply discarding the inactive row can leave a stale price incorrectly eligible.
Failed retrievals should not renew a price's age. If a weekly check cannot confirm the source, keep the previous verification date and apply the normal expiry rule. A chart should become unavailable when its evidence becomes too old.
Understand how the range is calculated
Cardinal's proposed asking-price method groups observations by exact variant, condition, region, native currency and comparable bundle. It uses hardware prices per accelerator excluding tax. This describes the intended evidence rules; no live Cardinal price feed is claimed here.
The proposed publication floor is five distinct eligible observations from at least three independent sellers, checked within the preceding fourteen days. This is a minimum evidence rule, not proof that the sample represents the entire market.
First, calculate the median eligible asking price for each seller. Then calculate the median and interquartile interval across those seller medians. This gives each independent seller equal influence rather than letting the largest listing catalog dominate.
Scroll across the diagram to read every label
Download diagram| Illustrative stage | Count | Meaning |
|---|---|---|
| Candidate records | 9 | Records awaiting the comparison rules |
| Excluded records | 3 | One rental, one different variant and one expired record |
| Eligible observations | 6 | Records meeting the stated comparison rules |
| Seller A observations | 3 | Contribute one seller median |
| Seller B observations | 2 | Contribute one seller median |
| Seller C observations | 1 | Contribute one seller median |
The interval describes the middle part of the seller-median distribution. It is not the lowest and highest available offer, a confidence interval or a guaranteed negotiating range. Small samples make the selected statistic especially important to explain.
Read the limits beside the number
Look for the observation count, independent seller count and source dates. Ten listings from one seller offer different evidence from ten listings across several independent sellers. A familiar marketplace name does not make its individual sellers interchangeable.
Duplicate and syndicated listings also need review. A tracking link should not create a second observation. If the same offer appears on several channels, count the underlying offer once where its identity can be established.
A change in the weekly range may reflect a changed sample. One seller leaving the eligible cohort can move the statistic even without a like-for-like price change. Review cohort membership before calling every chart movement a market trend.
Public visibility does not automatically authorize a recurring data feed or derived publication. eBay's agreement restricts automated access without permission, and its API program has access requirements. A credible reference needs a permitted evidence path as well as correct arithmetic. Source: eBay access agreement; source: eBay Buy API requirements
Use the reference to ask a better question
Attach the exact configuration and intended quantity to your request. State the condition you will consider and the destination. Ask the supplier to identify differences between the offered hardware and the reference cohort.
Use the offer comparison worksheet to preserve the original offer terms. The reference can help you investigate a difference, but the complete offer remains the document you need to evaluate.
Cardinal sources hardware and checks seller-supplied documents. A current quote should specify the actual configuration, included items and applicable terms. Until sufficient permitted evidence exists, a clear request for a quote is more informative than an invented price range.
Common questions
Is an asking-price range the amount buyers actually paid?
No. It summarizes seller requests under the stated method. Completed-sale evidence needs the accepted transaction amount and its scope. Keep those datasets separate.
Why can two prices for the same GPU differ?
Check the exact variant, condition, included items, currency and date first. Then examine the seller's specific offer terms. A family-name match does not establish an equivalent comparison.
Why would a product have no numeric range?
The evidence may be insufficient, stale or too ambiguous to normalize responsibly. Publishing no range preserves that uncertainty. A current itemized quote can still answer the buyer's sourcing request.
Does a lower price mean an offer is better?
The price is one field in the offer. Compare the exact hardware and required evidence before drawing that conclusion. Keep unresolved differences visible rather than reducing the decision to one number.