Where prices come from.
How to widen the US view.

Direct price references for RTX 4090, RTX 5090 and RTX PRO 6000 Blackwell 96GB, with a practical plan for better national asking-price coverage.

Price references, with their evidence type

Use recent exact-card sale comparisons as context. Keep them separate from active asks, retail offers and community opinions.

Where the screened asks are concentrated

Baseline snapshot: 2026-10-06T22:00:22Z. These are historical published counts by the seller’s state, not current inventory. Zero below means no screened observations in this sample; it does not mean no GPUs are for sale. Check current evidence and freshness in the dashboard.

A practical protocol for wider US asking prices

The following is our collection recommendation, applying Census geography and AAPOR disclosure principles to listing research. It is a saved plan, not a claim that national discovery is already running.

1. Run the same exact and fallback queries in a fixed geography/source rotation. Inspect the actual seller state and card identity; query location and text matches are not identity evidence.
2. Start with the nine divisions and 38 listed hubs, then rotate uncovered states and rural Craigslist areas. Alaska and Hawaii require their own checks. A city radius is not complete state coverage.
3. Record source, query, city, supported radius, selected sort, pages/cards checked, UTC observation time, access failure and review disposition. Unknown/inaccessible results are not zero inventory.
4. Collect broad active results plus newest results when the platform visibly supports them; do not select only low prices or promising thumbnails. Review all relevant offers where feasible; otherwise use a prespecified systematic or random selection within each discovered region/source cell and disclose that selection.
5. Recheck known URLs daily and discover new offers daily using supported access. The current dashboard button only rechecks known offers; a new discovery service is not implemented by this report.
6. Apply the existing explicit evidence review before admitting any ask. Keep ambiguous identity, parts-only, sold, sealed, auction, whole-PC, conditional trade and conflicting evidence outside the used core.
7. Deduplicate stable source IDs, then review possible cross-posts using seller/location, actual-unit photos, description and distinguishing labels. Do not merge merely because model and price match.
8. Keep each seller offer once. Preserve per-card asking amount versus bundle total; quantities and cross-platform hardware identity remain unknown unless evidenced.
9. Report local pickup, shipped asking amount and known mandatory shipping/fees separately. Unknown shipping is not zero; sales tax depends on destination.
10. Publish the listing-weighted observed median, middle 50 percent, counts, region/source cells and evidence freshness. Use one-day active asks and a separate recent-new-listing series; never mix these with 12-month sold averages.
11. Compare source-only and region-only medians as sensitivity checks once cells have data. Do not manufacture population weights from general population or PC gamer shares; actual GPU listing shares are unknown.
12. A non-probability marketplace sample cannot claim a national margin of error merely from sample size. More samples improve observed coverage, not a guaranteed national census or authenticity.

What makes collection more reliable