Rent NVIDIA H100
Hopper datacenter GPU with 80 GB of HBM3 on SXM5 boards. Server sizes from 1 to 8 GPUs, billed by the hour. Pay in crypto, no KYC, root over SSH.
Live price by server size
Loading live data…| GPUs | Per GPU-hour | Server, per hour | Availability | Action |
|---|
Interruptible (spot)
| GPUs | Per GPU-hour | Server, per hour | Action |
|---|
Price history
Use the left and right arrow keys to read values at each point in time.
| Time (UTC) | Best 1-GPU hourly price | Free cards |
|---|
What you can run on the H100
The NVIDIA H100 is a Hopper-generation datacenter GPU with 80 GB of HBM3 memory on SXM5 boards and hardware FP8 support. It suits training, fine-tuning and serving large language models, on one card or split across several. Rent it here by the hour, paid in crypto, with root access over SSH.
Typical tasks
- Serving large language models, on one card or split across several
- Full fine-tuning and training runs
- Long-context inference
- FP8 inference and training, where your framework supports it
- Batch jobs: embeddings, evaluation, data processing
What fits
A rough inference guide, in billions of parameters: weights only, about 20% left for context; real limits depend on context length and the serving engine. It assumes the weights take up to 80% of the total memory, at 2 bytes per parameter for BF16, 1 byte for FP8 and half a byte for 4-bit. With several cards, the serving engine splits the model across them.
| GPUs | BF16 | FP8 | 4-bit |
|---|---|---|---|
| 1 GPU | 32B | 64B | 128B |
| 2 GPUs | 64B | 128B | 256B |
| 4 GPUs | 128B | 256B | 512B |
| 8 GPUs | 256B | 512B | 1T |
Based on 80 GB per card.
How to rent the H100
- Sign inSign in with your e-mail. No KYC: you pay from your own wallet.
- QuoteWe confirm the machine and the price within 1 hour, 03:00–23:00 UTC.
- PayYou approve a quote before anything is charged. Pay in USDC on Ethereum today; more tokens are coming.
- KeysAfter payment we test the server and send you SSH access.
Billed by the hour, no minimum term. Reserved terms for a month, a quarter or a year are quoted by hand and cost less per hour.
Open the order formBuilding for clients? See the partner program. AI agent? llms.txt describes the public price file and the order link.
H100 FAQ
What size of model fits on one H100?
As a rough guide for inference, with the weights taking up to 80% of the 80 GB: about 32B parameters at BF16, 64B at FP8 and 128B at 4-bit on a single card. Real limits depend on the context length and the serving engine. The What fits table on this page also covers 2, 4 and 8 cards.
How much memory does the H100 have?
80 GB per card, with HBM3 memory on SXM5 boards. The Specs section on this page shows the memory the cards on offer report right now.
Does the H100 support FP8?
Yes. Hopper GPUs have hardware FP8 support, so 8-bit floating-point weights and activations can be used for inference and training where your framework supports it.
Can I rent more than one H100 in a server?
Yes. Servers come with 1, 2, 4 or 8 GPUs. The live table at the top of this page shows which sizes are offered right now; a size that is not offered can be asked for and is quoted by hand.
Do I need KYC?
No. You sign in with your e-mail and pay from your own wallet.