Better GPU price tracking

RTX 4090 · RTX 5090 · RTX PRO 6000 Blackwell 96GB

Research checked October 6, 2026 · updated 2026-10-06T21:59:16Z UTC · eBay excluded

A useful price estimate needs comparable cards, fresh evidence and a visible sample size. Collecting more search cards alone does not establish more used GPUs or a better market price.

What the app already does

The dashboard now shows the actual screened sample count toward the 12-offer target beside each median, plus the middle 50% price range. The live evidence report separates screened asks from pending and quarantined candidates and preserves direct links and observation times. Facebook and OfferUp observations are manually collected. The API makes bounded rechecks of known Craigslist URLs: up to 20 candidates, an eight-second request timeout and a ten-minute cache. This does not discover new listings or scan every marketplace. Observations older than 24 hours leave the current price pool; a dashboard reload does not itself renew manually collected evidence.

Actual sample counts and the middle 50% range are implemented. Further collection, a completed-price ledger and approved feeds remain the roadmap below. Native alerts and licensed feeds have not been connected. No completed-sale median, exhaustive US inventory count or independent hardware authentication has been established.

Make every sample comparable

1. Identify the exact product and condition. Separate used, new, unused open box, refurbished and parts. Exclude complete PCs, wanted ads, trades and unclear bundle pricing. Compare PRO 6000 full Workstation, Max-Q and Server editions separately; older RTX 6000 Ada, Quadro and A6000 cards are different products. NVIDIA documents the Blackwell 96GB family and editions.
2. Count distinct units, not repeated appearances. Deduplicate listing IDs across overlapping city searches, then flag likely crossposts and relists using consistent unit photos, model and location evidence. Claimed multi-card quantity is not confirmed stock. Record one price per comparable unit and distinguish local pickup from shipped cost and known buyer fees.
3. Review evidence before including the ask. Require consistent listing text, used condition, model evidence and a working or test claim under the existing screening policy. Keep missing or conflicting evidence outside the median. Physical photos and seller diagnostics support screening; they do not prove authenticity. A low price is a review signal, not proof of fraud.
4. Show the spread and coverage. Add the middle 50% range alongside the median; NIST describes quartiles and the interquartile range. Use a stable city/source search schedule rather than selecting only cheap cards. More observations cannot by themselves remove source or city selection bias; AAPOR explains sample-design and coverage considerations. This marketplace application is our inference from those principles.

Discovery and refresh methods

Proposed implementation order

1. Grow the current estimate. The actual n / 12 count and middle 50% range now appear on the dashboard. Continue screening eligible leads across sources and regions and disclose comparable edition/condition. Keep unfilled targets honest.
2. Add a completed-price ledger. Store the explicit sold amount, source link, edition, condition, date and evidence type. Use a separate recent-sale median; do not turn a vanished ad, current bid or old ask into a realized price.
3. Connect approved inputs. Prioritize a permitted Reddit or seller opt-in inventory/transaction feed. Native marketplace alerts can help a reviewer find new leads. A server schedule can process authorized data; it cannot supply platform access permission.
4. Measure freshness and coverage. Record first seen, last checked, price changes and availability changes. Show each source’s last successful discovery and listing recheck separately. Preserve unreachable status as unknown; remove explicit Sold offers from active asks. Audit duplicates and screening decisions regularly.