An A100 80GB that has served four or five years — theoretically past the half-way point of depreciation — carries a used asking median of $18,998, above the publicly compiled launch-price range of $10,000–18,000. This is not a residual-value reading (the denominator is not recomputable, and the basis is asking, not transacted), but taking the anomaly apart yields exactly the kind of verifiable observation Michael Burry's depreciation debate is missing.
Update 2026-11-12: the "denominator does not exist" claim in section ③ is now backed by an exhaustive EDGAR full-text audit (all 823 "A100" hits classified); the negative-search fields live in api/v1/grvi/anchors.json.
Nail down the basis before discussing the anomaly
The GRVI collection pipeline captured 46 validated asking listings on eBay this week (A100 80GB: 32; H100 80GB: 14; as of 2026-11-10). Each is recorded with title, condition (new/used), form factor (SXM/PCIe), listing currency and the FX rate applied. Medians and quantiles are recomputable from the raw panel (see the machine-readable endpoints).
Two caveats that cannot be skipped. First, these are asking prices, not transaction prices — eBay serves sold data only to logged-in sessions, a platform limit, not laziness. Second, our publication threshold is n≥5 per cell: used A100 has 17 listings (the asking median may be published), used H100 only 3 (listed, no figure). The GRVI residual-value reading remains unanchored — the index page continues to show a dash.
Same model, same platform, same week
| Card · form factor · condition | n | Median | p25 | p75 | Note |
|---|---|---|---|---|---|
| A100 80GB · SXM · used | 10 | $13,994 | $13,994 | $23,549 | mostly part-outs |
| A100 80GB · PCIe · used | 6 | $18,998 | $18,998 | $21,868 | plug-and-play, scarcer, pricier |
| A100 80GB · SXM · new | 7 | $19,499 | $18,343 | $22,499 | — |
| A100 80GB · PCIe · new | 7 | $28,798 | $22,597 | $31,912 | — |
| H100 80GB · new (all) | 11 | $45,673 | $40,878 | $51,096 | vs SEC anchor $39,216 |
| H100 80GB · used (all) | 3 | $37,975 | $17,999 | $53,847 | below n≥5, listed only |
Isolate A100 used vs new on the same platform and week: the used median is only 16% below new — far shallower than the intuitive "electronics lose half their value in four years" curve.
The form-factor detail carries information too: used SXM ($13,994 median) is well below used PCIe ($18,998) — SXM modules cannot run outside their OEM baseboards, so part-out supply is heavy and prices are depressed; PCIe cards are the plug-and-play "hard currency" of smaller operators.
Denominator mismatch is the main one — but not all of it
First, denominator mismatch: $10k–18k was a volume price, not today's market price. The publicly compiled A100 launch range comes from three sources (gpuresource: PCIe ~$10,000 / SXM4 ~$15,000; industry digests: ~$14,500 guidance at launch, and a $15,000–18,000 variant). Those are 2020–21 volume contracts and list guidance for hyperscalers; today's secondary market is single cards, small lots, retail channels. Comparing retail asking to bulk-contract pricing overstates the ratio by construction — which is exactly why we publish a bracketed 1.06–1.90 instead of a reading: no buyer-disclosed A100 unit price exists in SEC filings, so the denominator is not recomputable.
Upgrading "does not exist" (2026-11-12): an exhaustive audit of 823 hits. "Not in SEC filings" is no longer an impression — it is a recomputable conclusion. We ran an exhaustive EDGAR full-text sweep (from 2019, all form types): all 823 "A100" hits were classified; the buyer-purchase phrase matrix ("purchased A100" / "A100 GPUs for $" and six more) returned zero hits — and as a sanity check, the same phrasing recalls all nine IREN filings behind the H100 anchors, so the queries do catch buyer disclosures when they exist. Every GPU-context issuer (CoreWeave, Recursion, TeraWulf, Applied Digital, Super Micro, VivoPower, Cyxtera, Boost Run, One Stop Systems, NetApp and eight listed miners) was read in the original and ruled out: Recursion disclosed the count of BioHive-1 — 40 DGX A100 nodes — but no amount; CoreWeave and TeraWulf carry deployment narratives only; Applied Digital is a hosting note; VivoPower states outright that it "does not purchase or provide GPUs". The remaining 700-odd hits fall into five false-positive categories (poly-A tails, an AAV vector name, a clinical endpoint, ABS tranche codes, EV part numbers). The denominator is not "not yet found" — it demonstrably does not exist; the queries, the sanity check, the per-issuer exclusions and the false-positive taxonomy are published field by field in the GRVI anchor register (machine-readable: the negative_search field of api/v1/grvi/anchors.json).
Second, supply doesn't come to market: holders aren't selling. The A100 is still a workhorse of inference. Cloud and colocation operators are running theirs — earning, not listing. What circulates is part-outs, corporate clear-outs and retail cards. Thin supply keeps asking prices firm.
Third, a demand vacuum: buyers who want new A100s cannot get them. The A100 is discontinued and next-generation capacity goes to the largest customers first; smaller operators topping up inference capacity go to the secondary market. Discontinued supply plus live demand — "used above launch" is not mysterious in that structure. Implication for the Burry debate: Michael Burry argues hyperscalers stretch GPU depreciation to five or six years, so book depreciation lags real economic decay. Yet at least on the A100 asking market, "a card four or five years in still commands ~84% of new asking, or a bracketed 1.06–1.90× the launch range" — the debate needs denominator-verifiable observations like these, not positions.
Scoring eBay asking against a real SEC purchase price
GRVI's H100 denominator is anchored by two IREN SEC filings at $39,216/card (816 cards for $32.0m combined — quantities and amounts disclosed by the buyer, fully recomputable). The same week, the eBay new-listing median for the H100 is $45,673 (n=11) — 16% above the SEC price. That is the size of the asking premium: asking sits systematically above real transaction levels.
Apply that haircut to the A100: $18,998 ÷ 1.16 ≈ $16,400. The bracketed ratio against the $10,000–18,000 launch range narrows from 1.06–1.90 to 0.91–1.63 — even under the most asking-inflated adjustment, "a used A100 trades around the top of its launch range" still holds. The anomaly is robust, not an artefact of inflated listings.
One corroborating datapoint: used H100 asking (n=3, below threshold, listed only) has a median of $37,975 — 97% of the SEC-anchored $39,216. Two fully independent sources land in the same magnitude, suggesting asking prices are not unusably far off.
· This issue publishes no residual-value reading. GRVI = secondary price ÷ launch price; the A100 denominator is non-recomputable — settled by the exhaustive audit of 2026-11-12 (all 823 EDGAR "A100" hits classified, zero buyer-purchase phrase hits; see section ③) — so 1.06–1.90 and 0.91–1.63 are bracketed values: any citation must carry the "range, not a reading" label, and the index page keeps showing a dash.
· The basis is the median asking price, not transactions and not a residual value; eBay serves sold data only to logged-in sessions. The 16% asking-premium calibration itself has a timing mismatch (SEC purchases predate the listings) and should be read as an order of magnitude only.
· The sample is one platform (eBay, cross-checked across US/UK/DE/AU), one snapshot (2026-11-10), n=17; used H100 is structurally scarce (trades as full systems and broker quotes) and we do not relax the criteria to pad the count.
· SpreadBench aggregates public data only. This is not a credit rating, not investment advice, and expresses no view on any issuer or hardware vendor.