1 × 24 h × $1.73 · 24 GPU-hours
~$1,246 / month at 720 hOpen H100 listings →GPU hour cost calculator
GPU hour cost calculator
Catalog from-price times GPU count times hours. Compare H100 against H200 on the same job before you upgrade the badge.
bill = from-price × GPU count × hours. Monthly uses 720 hours.
$30.48 more than H100
~$2,160 / month at 720 hOpen H200 listings →How to use it
Four steps, then confirm the live SKU
bill = from-price × GPU count × hours. Monthly uses 720 hours.
- 01
Pick the SKU
Start with the GPU you will actually rent. The from-price is the cheapest published on-demand hour in the GPUBeacon sample, not a hyperscaler list price.
- 02
Set GPU count and hours
An 8× node is eight GPU-hours per wall-clock hour. Put the runtime you expect, not the one in a vendor blog with a different batch size.
- 03
Compare a second card
H200 only wins if the extra memory drops a GPU or stops a swap. Do that math here, then open the live listings.
- 04
Confirm the live row
This is a snapshot. Open the GPU page and confirm stock, region and the provider’s live rate before you paste a card.
What this calculator actually computes
I multiply the catalog from-price by GPU count by hours. That is the bill for a published on-demand hour in the GPUBeacon sample. Monthly is the same rate times 720 hours. I do not invent a discount. I do not mix spot into this number.
The from-price on an H100 card today is a marketplace on-demand row. Managed clouds sit higher. Hyperscalers sit wherever procurement already lives. If you treat the cheapest cell as a promise that your region has that host, you will provision the wrong thing.
How to read H100 vs H200 vs B200 on the same job
Same GPU count. Same hours. Then look at the delta. H200 is not “better.” It is 141 GB instead of 80 GB. If you cannot name the bytes that spill, the extra hour is vanity silicon.
B200 is a different conversation: FP4, a newer fabric, and a catalog that often says waitlist. A waitlist B200 at a pretty from-price is not cheaper than an in-stock H100 you can start tonight.
- Stay on H100 if 80 GB holds weights, optimizer, activations and the KV cache you serve.
- Move to H200 when the extra 61 GB removes a GPU from the job.
- Look at B200 when the stack actually uses FP4 and the row is in stock.
What this number is not
It is not egress. It is not disk. It is not the hour you lose debugging NCCL. It is not a reserved cluster. It is not spot.
If the listing is waitlist, quote, or reserved, do not use this calculator. Use the listing checker, then walk away.
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
GPU hour cost calculator questions
How much does an H100 cost per hour?+
On GPUBeacon’s on-demand sample, H100 starts around $1.73 per GPU per hour on a marketplace listing. Managed clouds and hyperscalers sit higher. Put the SKU in this calculator, then confirm the live row on the H100 page.
How do you estimate monthly GPU cost?+
I use 720 hours per month: 24 × 30. That is a convention, not a calendar. If you run 12 hours a day, cut it in half. If the node sits idle, you still pay the on-demand hour until you stop it.
Can I compare H100 and H200 here?+
Yes. Same GPU count, same hours. The delta is the only honest upgrade test. If H200 does not drop a card or stop a swap, stay on H100.
Is this a live quote?+
No. Catalog from-prices are a snapshot. Always confirm the provider’s live SKU, region and stock before you provision.