GRVI · GPU Residual Value Index

First issue · listing price panel (2026-11-10)

46 validated active eBay listings, each carrying a source_url and capture time. The table below is the median asking price — not a transaction price, and not the index reading.

Basis note: eBay listings skew to single retail cards and SXM modules; batch rack pricing is a separate regime and must be labelled separately, never mixed. Cross-marketplace audit (2026-11-10): to rule out a capture that is simply too narrow, the same keyword and filters were run against eBay UK, Germany and Australia for used H100 80GB — the UK returned 4 matches but all were complete servers (DL380 Gen10 / DGX H100 / ESC8000A-4U), and Germany and Australia returned zero qualifying single cards. Including the US sample, the worldwide count of qualifying single-card listings is 3. Used H100s change hands mainly as whole systems and through brokers who quote privately, with no public verifiable price — the scarcity is structural, not a capture gap, so the threshold is not relaxed to manufacture five observations.

Anchoring the denominator: what buyers actually paid, per SEC filings

Register of P_launch anchors. Only buyer-side GPU purchase quantities and amounts disclosed in SEC filings are recorded, each with a source link and the original quote; unit price = amount / quantity and is fully recomputable. Issuers describe counts as approximate, so unit prices are approximate too and must be cited with 'approx'.

H100 80GB is anchored: 816 cards for US$32.0m combined, weighted $39,216 per card (IREN 6-K). Denominator resolved; the block is sample size: eBay US has 3 qualifying single cards, eBay UK's 4 hits were all complete servers, and Germany and Australia returned zero (cross-marketplace audit 2026-11-10). Used H100s trade as whole systems and via private broker quotes, so the scarcity is structural and the threshold is not relaxed to manufacture five observations.

A100 80GB remains unanchored: An exhaustive negative search is now complete (2026-11-12, EDGAR full-text, all form types from 2019-01-01): all 823 'A100' hits were classified and every GPU-context issuer read in the original (CoreWeave, Recursion, TeraWulf, Applied Digital, Super Micro, VivoPower, Cyxtera, Boost Run, One Stop Systems, NetApp, Amazon, Infosys and eight listed miners) — every genuine mention is a deployment narrative, a hosting note or seller marketing, none is a buyer-side quantity-times-amount disclosure; the remainder are false positives (poly-A tails, an AAV vector name, a clinical endpoint, ABS tranche codes, part numbers). The A100 denominator keeps waiting for verifiable disclosure and will not be downgraded to third-party compilations of ~$10,000 (PCIe) / ~$15,000 (SXM4)

Cleaning rules (reproduce this exact sample)

• Capture: 1126 raw records from eBay active listings (keywords A100/H100 80GB, Graphics/Video Cards category, condition limited to used or new).
• Model check: the title must contain A100 or H100 as a word, and must contain 80GB — this drops P100/K80 items eBay injects as related matches, and drops 40GB variants.
• Anomaly exclusion: accessories and shells (bracket / backplane / heatsink / waterblock / housing / cable / fan…), faulty units (ECC error / for parts / not working), engineering samples (qual sample), and complete systems or lots (DGX / HGX / server / lot of / 2x / 4x) never enter the base curve.
• Price band: A100 80GB is bounded to $5,000–40,000 and H100 80GB to $8,000–80,000; anything outside is a miscapture or a scam listing.
• De-duplication: one row per URL; listings sharing card, condition, normalised title and identical price are treated as the same offer re-listed and counted once.
• FX: listings priced in HKD are converted at the capture-date rate 0.127513 (source open.er-api.com/v6/latest/HKD); the rate is stored on every row.
• Attrition: 1126 raw → 46 validated. Main rejection reasons: model mismatch 419, capacity mismatch 140, duplicate URL 493, excluded term 22, re-listed duplicate 5, price out of band 1.

1. Why this index exists

Since November 2025 Michael Burry has argued that hyperscalers understate depreciation by extending useful lives (his estimate: about $176bn cumulatively over 2026–2028, with Oracle/Meta earnings potentially overstated by 27%/21%). The rebuttal cites A100s running at full load into 2029 and a batch of maturing H100s re-leased at 95% of original price. Amazon shortened some server lives from six years to five in 2025; Meta has extended three times in three years.

Accounting standards (ASC 250) say this hinges on the asset's true economic life — yet the market still has no independent, recomputable residual-value series to adjudicate it. Both sides cite their own samples; neither publishes a public transaction series.

Whitespace check

Whitespace check performed 8 November 2026. Sources listed at the foot of this page.

2. Definition and formula

Stratification

3. Source register

Upstream: rentgpu.link (31-platform rental capture) · DCPI (power cost) · CIDX (China supply cost)

Downstream consumer: NCSX-IS (residual value feeds credit analysis)

4. Outputs and machine endpoints

/api/v1/grvi/latest.json /api/v1/grvi/panel.json grvi-observations-2026-11-10.csv

Field shape is identical across the family: ticker / status / reading / observations / unit / cadence / methodology_version / next_publication. Parse once, cover all five.

5. Status and milestones

6. Red lines

7. How to cite

Free to cite for research and press. Always give the ticker, the version and the URL — dropping any of the three breaks machine traceability.

Other citation formats (APA / MLA / Chicago / BibTeX) · Data licence