113 providers publish prices. 19 will tell you a machine exists.
A GPU price answers one question: what would the hour cost. It does not answer the question the buyer actually has, which is whether an hour exists to buy. Those are different data, published by different machinery, on different clocks. A rate card is a document someone edits every few weeks. An availability signal is a feed that has to be true right now, because the deploy button is wired to it. Most of the market publishes the first and calls it pricing.
Here is what that looks like tonight, September 22, 2026. The panel tracks 2,192 listings across 113 providers and 340 GPU models. 711 of those rows - 32% - come from live APIs across 19 providers, where price and stock update together. The other 94 providers publish rates with no machine attached. Some of them have plenty of capacity. Some have none. Tonight, Latitude.sh's entire published GPU fleet - H100, RTX PRO 6000, B300 - reads out of stock on its own pricing page, while the rates sit there looking perfectly rentable. A buyer comparing headline numbers would never know.
The rentable market is smaller than the listed market
Filter the panel to single-GPU H100 machines with a live availability signal saying yes, right now: 86 machines across 31 providers, median $4.50 an hour, from $1.30 at the floor. That is the real market - the set of things you can actually rent before dinner. Everything outside it is a document. And the documents have a pattern: they are cheapest exactly where they are least checkable. The providers without a stock signal skew to the low end of the price table, because a price you cannot test is a price you can set low. We are not saying the low rates are fiction. We are saying the burden of proof sits with the buyer, and most buyers never pick it up.
| Live-API rows | Catalog and scrape rows | |
|---|---|---|
| Providers | 19 | 94 |
| Listings | 711 | 1,481 |
| Stock signal | Available or sold out, refreshed hourly | None - a claim, not a machine |
| When the number changes | When the market moves | When someone edits the page |
What this costs a buyer
The failure mode is always the same. A team shortlists on headline price, picks the cheapest plausible provider, and discovers at deploy time that the cheap machine does not exist - it is sold out, region-locked, or reserved for annual contracts. The shortlist was built on a number that was never theirs to have. The fix is not cynicism about low prices; Lium rents real H100s at $1.30 an hour tonight, verified against a live feed. The fix is sequence: check that the machine exists, then compare what it costs. A sold-out $1.68 is not cheaper than an available $4.50. It is just cheaper-looking.
One honest caveat on our own numbers. Three days ago the rentable H100 median read $3.50 across 45 machines; tonight it reads $4.50 across 86. That is not prices moving - it is the panel adding 25 providers in between, so the two medians are measuring different markets. When the cohort changes we say so. A median that silently mixes cohorts is how fake trends are manufactured, and this panel exists to kill them, not mint them.
- Before any price comparison, ask one question per provider: can you show me the machine? A live feed, a deploy button, an API that says available - any of these counts. A PDF does not.
- Weight providers with stock signals first. Their prices are testable, and a testable price is worth a premium over a hypothetical one.
- When a provider shows sold out, believe it and move on. The rate will still be there when stock returns, and so will this panel.
- Distrust any comparison - ours included - that mixes rentable machines with rate-card claims without marking which is which.