My Chart Had Estimates in It, and I Got Caught
A price-history chart mixed transaction data with remembered guesses. The web-sourced 'correction' was worse than the guess. Only sold listings survived.
While deciding whether C$3,000 per mining card was a good price, I built a price-history chart from eBay’s completed listings: every 170HX sale the platform would show me, 81 transactions across the full 90-day window.
The sales data was genuinely interesting:
Prices nearly doubled in six weeks — median C$1,776 in late July, C$3,343 by late August. A regime change, presumably unlock-driven demand meeting thin supply. Asks ran about C$500 above the sales fit, sellers pricing ahead of clearing. And the apparent 8 GB/10 GB price gap dissolved on inspection into a timing artifact: the 10 GB sales just clustered in the cheap July window.
So far so good. Then I added a comparison rail: where do the alternatives sit — A100s, A6000s, the current-gen cards? And here I did the thing this post is a confession about: I plotted remembered price estimates on the same chart as verified transaction data, visually indistinguishable from the real points.
I got called on it, correctly and bluntly. The first fix attempt made it worse. First correction pass: look the prices up on the web — buying guides, price-tracker articles. Second pass, after being pushed again: sweep actual eBay sold listings for all eight comparison cards, 115 completed sales, medians only.
| Card | My estimate | Web “correction” | Actual sold median |
|---|---|---|---|
| A100 40 GB PCIe | C$6,500 | C$3,800 | C$6,059 (n=5) |
| A100 80 GB PCIe | C$12,000 | C$7,600 | C$19,233 (n=5) |
The middle column is the punchline. The web-sourced correction was further from reality than my original guess. Buying-guide prices for datacenter cards are aspirational, stale, or quoting a different market entirely. My gut, trained on months of adjacent browsing, beat the articles — and the sold-listing medians beat both by not being an opinion at all.
The full verified set, for anyone pricing this tier (eBay.ca sold medians, delivered, swept Sep 3–5, 2026): RTX A6000 C$5,634 (n=28) · A100 40 GB C$6,059 (n=5) · RTX 5090 C$6,668 (n=26) · A40 C$6,898 (n=6) · RTX 6000 Ada C$8,182 (n=14) · L40S C$10,979 (n=7) · A100 80 GB C$19,233 (n=5) · RTX PRO 6000 C$20,854 (n=24).
Two rules came out of this, now standing for everything I publish:
- A chart may not mix evidence grades without marking them. If a point is an estimate, it renders as an estimate or it doesn’t render.
- For hardware prices, only completed transactions count. Asks measure hope. Articles measure nothing. Sold medians with an n attached are the only numbers that survived contact with reality — and when n is 5, say so.
The rebuilt chart reversed my substitution analysis: at C$47/GB against a real C$117/GB floor, nothing threatens the mining card’s value proposition until it roughly doubles. The conclusion I nearly published — built twice on estimates — had said otherwise.