Methodology

About josh.menu

josh.menu tracks GPU prices — consumer cards and datacenter accelerators — and ranks every one by how good today’s price actually is. Not “is this cheaper than yesterday,” but “is this a good moment to buy or rent, judged against this card’s own observed history and its original launch price.”

Where the prices come from

Every price is a real, timestamped observation from a named partner via their official API, polled on a schedule:

• • eBay Browse API — current consumer marketplace listings (new and used), with live buy links.
• • Vast.ai API — cheapest on-demand cloud GPU rental rates, with referral-tagged rent links.

We do not scrape retailer websites. As affiliate and data partnerships approve, each new source widens the “where to buy” picture — still through official feeds and APIs, not page scraping.

The deal score, explained

Every GPU gets a score from 0 to 100. A 50 means the price is normal. Higher is a better moment to buy. The score blends two views; the blend weight depends on how much history we have:

• • Versus its own recent history. We collapse noisy half-hour checks into one near-best legitimate price per day — the cheapest print that isn’t a statistical outlier (with a 25th-percentile fallback on thin days) — then ask how unusual today is versus the prior 90 days (30 when history is shorter), using a median and median absolute deviation so one weird day can’t warp the baseline. Today’s print is left out of that baseline on purpose. A calm dip scores hotter than the same percent move in a chaotic market. That gap maps onto 0–100 (50 = typical).
• • Versus launch MSRP (when we know it). A card that drifted down from wildly overpriced to merely overpriced is not a deal, even though it’s “cheaper than usual.” Anchoring against the original launch price keeps the score honest in absolute terms. Datacenter rentals have no retail MSRP, so they’re scored on history alone.
• • Confidence-weighted blend. With only a handful of baseline days we lean on MSRP; as history fills in, we trust the card’s own pattern more. Weight grows with baseline days (about 30% history trust at 5 days, ~63% at two weeks, ~88% at a month).

Three honesty rules on top: we never score on thin data (fewer than 5 distinct baseline days shows “not enough data yet” instead of a made-up number); used and refurbished listings never mix into a new-condition score; and when a price is the lowest we’ve ever recorded, we say “all-time low” — with “lowest since {date}” for anything short of that.

Marketplace caution

A low price on a marketplace is not the same as a safe purchase. GPU listings are a frequent scam target — empty boxes, bait-and-switch titles, and brand-new seller accounts with no feedback. We filter obvious junk and skip sellers with zero feedback, but we can’t verify every listing in person. Before you buy: check the seller’s history, read the full description, prefer buyers who use eBay Money Back Guarantee (or equivalent), and walk away if anything feels off. When in doubt, pay a bit more from a seller with a long track record.

How this site makes money

Some Buy and Rent links (for example eBay affiliate links, Vast.ai referrals) may earn josh.menu a commission at no extra cost to you. Affiliate status never reorders rankings or changes the deal score: the score is computed from price data alone, and no retailer can buy a better one. There are no ads.

Briefings

When the market genuinely moves — an all-time low, a sharp drop, a sustained shift — the system proposes a short data-grounded briefing. A human reviews, writes, and signs off on every published piece. If nothing profound happened, nothing gets published. No filler, ever.

Questions or a correction? The site is run by Josh — [email protected].