Score out of 100. Failures cost 25. Warnings cost 10.
Open H100 listings →Listing checker
Cloud GPU listing checker
Fail a row that is not actually on-demand. Waitlist, SKU mismatch, missing fabric, quote theatre. Before it hits a spreadsheet.
Fail waitlist, quote, SKU mismatch, or multi-GPU with no fabric. Then look at the dollar.
If the job cannot die, do not take spot. Un-checkpointable fine-tunes pay on-demand.
- VRAM matches the catalog SKU
H100 80 GB.
- Region is set
US-East
- Marked in stock
Still confirm on the provider page.
- On-demand billing
This is the product GPUBeacon ranks.
How to use it
Four steps, then confirm the live SKU
Fail waitlist, quote, SKU mismatch, or multi-GPU with no fabric. Then look at the dollar.
- 01
Copy the claimed GPU and VRAM
H100 80 GB. Not “Hopper.” Not “H100 equivalent.” If listed VRAM is off the catalog SKU, it is a different product.
- 02
Set GPU count and interconnect
One GPU, eight GPUs, NVLink, InfiniBand, Ethernet, PCIe. More than one GPU and “not stated” is a fail. You are not comparing that to an SXM node.
- 03
Name region, stock, billing
US-East cheap and EU-West legal are not the same hour. Waitlist and talk-to-sales are flags. Spot is a bid. Reserved is a contract.
- 04
Read fail, warn, or comparable
Fail means do not paste it next to an on-demand H100. Warn means read the footnote. Comparable means you can finally look at the dollar.
The four fields I check before the dollar
Price last. If two of the four fields differ, stop. You are not looking at the same rental. Model and VRAM. GPU count and interconnect. Region. Stock. Then billing. Then the dollar.
GPU Cloud News loves “H100 available.” The row either provisions or it does not. A waitlist is not a price. A quota form is not a price. This checker exists so you fail that row in public, not in a spreadsheet your manager believed.
- 80 GB HBM is the default H100. Confirm it on the row, not in the hero.
- SXM when you need NVLink. PCIe when one GPU is the whole product.
- In stock beats a prettier number you cannot start.
What a fail means
Score starts at 100. Each fail costs 25. Each warning costs 10. Comparable means zero fails. I still want you to confirm the live SKU. This is a structure check, not a purchase.
Spot without a checkpoint is a fail. Multi-GPU without a named fabric is a fail. Waitlist is a fail. Quote and reserved are fails. Those are not on-demand hours.
Marketplace, managed cloud, hyperscaler
Say the business model first. A Vast.ai host is not a Lambda SXM node. Mixing them in one table is how the cheapest cell wins and the job dies. This checker does not name the provider. It names whether the row is even a product you can compare.
Catalog from-prices are a snapshot. Estimates are envelopes, not quotes. Confirm the live SKU, region and stock with the provider before you rent.
FAQ
Cloud GPU listing checker questions
Why does a waitlist fail?+
Because you cannot start it. A waitlist with a Buy button is still a waitlist. Do not put it in an on-demand spreadsheet.
Why does missing interconnect fail on multi-GPU?+
An 8× H100 SXM node is a training box. A pile of PCIe cards on Ethernet is not. If the listing will not say NVLink or InfiniBand, assume the worst fabric the provider sells.
Is a limited listing a fail?+
No. It is a warning. Confirm the live row still provisions. Limited is not waitlist. It is also not a promise.
Can I check a spot listing?+
Yes. Spot is a warn if you can checkpoint, a fail if you cannot. Then use the spot expected-cost calculator. Do not mix the bid with on-demand H100s.