Four providers in one evening, and what each had to prove

Yesterday morning the panel had 14 providers and 227 listings. Tonight it has 18 providers and 290. Four sources joined in a single evening, and each one joined for a different reason, because the rule has never changed: a provider enters the panel when its source is honest enough to quote, never before. Here is what that rule did today.

TensorPool: a rate card with a trap in it

TensorPool publishes a clean on-demand rate card - B200 at $4.99, H200 at $2.99, B300 at $5.49, H100 SXM at $1.99, L40S at $1.49 per GPU-hour. The trap: the same page carries comparison tables for Lambda and the hyperscalers. We quote TensorPool's own table and nothing else; quoting a provider's marketing comparison as market data is exactly how comparison sites rot. Five rows, catalog labels, region not published, so the rows say so.

Massed Compute: check the arithmetic before you trust it

Massed Compute publishes per-configuration prices, and multi-GPU pricing is ambiguous across the industry - sometimes it is per-GPU, sometimes per-node. Before quoting a single row we verified the semantics against the page itself: the H100 x4 card prices at exactly four times the x1 card, the L40S x8 at exactly eight times. The shown price is the total for the configuration listed, so that is what the rows carry. Thirty-nine rows: the x1 and largest-config card per model, A30 through B300.

Lium: the page grades itself

Lium's pricing page states its own rule at the top: a range is the spread of live asks across pods rentable right now, a single figure is a reference price for a model with no pod available. So the connector quotes only the 15 models with live pods - B300 from $8.00 down to RTX 3090 from $0.16 - with availability taken from the page's own pod counts. The 76 reference-price models are excluded rather than shown as rentable, because a reference price for nothing you can rent is not a price.

GMI Cloud: floors, labeled as floors

GMI Cloud publishes 'from' floor rates - GB200 from $8.00, H100 from $2.00, H200 from $2.60, B200 from $4.00 per GPU-hour. Floors are welcome in the panel as long as they say what they are: every row is labeled (from), and floors stay out of the headline statistics. The GB300 preorder card carries no price at all, so it is not in the panel.

Who did not make it, and why

• CoreWeave: no hourly rates anywhere fetchable - pricing is a sales conversation.
• Hyperbolic: marketing numbers on the landing page; the real marketplace sits behind login.
• QuickPod: a JavaScript shell with no fetchable rates.
• GPU Outlet: redirects to a landing page with no prices.
• SimplePod: datacenter GPUs are reservation-only; consumer RTX cards carry 'from' floors. Possible later, low value today.
• Together AI: GPU pricing is per-token serverless or enterprise quotes - no hourly card exists to quote.
• Neysa: publishes monthly commitment pricing only. Turning that into an hourly number would be our arithmetic on their price, not their price.

What the bigger panel says tonight

• 18 providers, 290 listings. 140 rows AVAILABLE, 150 CATALOG.
• H100, single GPU, rentable now: $1.25/hr (Lium) to $2.69/hr (RunPod) across 6 rows. Yesterday the cheapest answer was $1.99 - it changed because the panel grew, and the receipt on the new answer says where it came from.
• The new providers added real stock, not just rows: Lium alone contributes 15 models with pods rentable right now.

Every number above is computed from the live panel at the time of writing, and every row behind them carries its source and observation time. The panel refreshes every six hours; if you are reading this later, the panel has moved - trust the panel, not the post.