The state of the panel, September 2026

This is a snapshot of the Compute Cafe panel as it stood on September 17, 2026. It is not a state of the market. It is a state of what we can see, from sources we can show you, and the difference matters - so this post ends with an explicit list of what the panel does not cover.

The numbers

• 14 providers, 227 listings, more than 100 distinct GPU models across 38 locations.
• About 6 in 10 rows are AVAILABLE (an endpoint said you can rent this now); the rest are CATALOG (a published rate card, no stock claim). Over 160 rows come from live APIs.
• H100, single-GPU configurations: $1.99/hr to $4.29/hr across 8 providers. Same GPU family, more than a 2x spread - that gap is the reason this site exists.
• A100, single-GPU: $0.89/hr to $2.30/hr across 8 providers.
• 15 listings sit in India (E2E Networks, JarvisLabs), priced in rupees; Scaleway lists in euros - both converted at the rate observed at refresh time, with the native price shown on the row.

Every one of those numbers is computed from the live panel at the time of writing, and every row behind them carries its own receipt: the source it was read from and the time it was observed. If you are reading this after September 17, the panel will have moved - trust the panel, not the post.

What changed this week

Five providers joined in the first two days: Nebius (seven on-demand rates, including B300 and B200), Crusoe (six rates, including the first AMD in the panel, MI300X), Latitude.sh (on-demand bare metal, H100 from $1.68/hr), Scaleway (EU on-demand L4 family, priced in euros), and Salad, which publishes only 'from' floor rates. Floor rates are always labeled (from) - a floor is a floor, not a quote. DataCrunch rebranded to Verda; the panel shows the new name, and the listing IDs and price history survived the rename untouched.

What the panel does not cover

• The hyperscalers. AWS, GCP, and Azure price GPUs behind instance families, region matrices, and committed-use discounts that do not reduce to one honest on-demand number per GPU. We would rather not show them than show them badly.
• TensorDock live availability. Their pricing page is in the panel, but their API requires an account, and new signups are temporarily paused on their side. Their rows are carried over and flagged; the live connector ships the day signups reopen.
• CUDO Compute. They publish no public rate card, so there is nothing honest to scrape.
• Committed-use, reserved, and enterprise contract pricing. By nature invisible from the outside.
• Interruptibility fine print. Marketplace rates (Vast.ai, Salad) can be preemptible; the receipt says where the number came from, and the provider's own terms govern what happens to your job.

Fourteen providers is a start, not a claim. The panel grows when a source is honest enough to quote - never before.