The H100 is the GPU to reach for when a model or a training run no longer fits comfortably on a 48 GB card: 80 GB of memory per GPU, FP8 support that the A100 does not have, and up to 8 GPUs in one VM. Every offer on this page is a VM you can launch right now with SSH access, billed by the hour and settled to the second when you stop.
- Offers now
- 10
- From
- $2.50/GPU-hr
- GPUs per VM
- 1x, 2x, 4x
- Providers
- 3
H100 offers you can launch now
The table lists every H100 VM in stock on QuantaCloud, Massed Compute, Verda and Hyperstack, cheapest first, and refreshes every 15 minutes. The Launch link opens that exact offer on the deploy page, where the full hourly breakdown is shown before anything is charged.
| GPU | GPUs | Provider | Region | vCPU / RAM / Disk | Interconnect | Per GPU-hr | Per hour | Launch |
|---|---|---|---|---|---|---|---|---|
| H100 PCIe 80 GB | 1x | Hyperstack | CA | 28 / 180 GB / 850 GB | - | $2.50 | $2.50 | Launch |
| H100 PCIe 80 GB | 1x | QuantaCloud | Midwest, US | 20 / 128 GB / 1250 GB | - | $2.59 | $2.59 | Launch |
| H100 PCIe 80 GB | 1x | Massed Compute | Iowa, US | 20 / 128 GB / 1250 GB | - | $2.73 | $2.73 | Launch |
| H100 NVL 94 GB | 1x | Massed Compute | Iowa, US | 18 / 128 GB / 750 GB | - | $3.11 | $3.11 | Launch |
| H100 PCIe 80 GB | 2x | Hyperstack | CA | 60 / 360 GB / 1600 GB | PCIe | $2.50 | $5.00 | Launch |
| H100 PCIe 80 GB | 2x | QuantaCloud | Midwest, US | 40 / 256 GB / 2500 GB | PCIe | $2.59 | $5.18 | Launch |
| H100 PCIe 80 GB | 2x | Massed Compute | Iowa, US | 40 / 256 GB / 2500 GB | PCIe | $2.73 | $5.46 | Launch |
| H100 NVL 94 GB | 2x | Massed Compute | Iowa, US | 36 / 256 GB / 1500 GB | NVLink bridge | $3.11 | $6.22 | Launch |
| H100 PCIe 80 GB | 4x | Hyperstack | CA | 124 / 720 GB / 3300 GB | PCIe | $2.50 | $10.00 | Launch |
| H100 NVL 94 GB | 4x | Massed Compute | Iowa, US | 72 / 512 GB / 3000 GB | NVLink bridge | $3.11 | $12.44 | Launch |
DeployGPU only lists providers it can launch machines on. For H100 prices across the whole market, compare all providers on GpuPerHour. For a reference price over time, see the H100 rental index (GPUIndexes).
SXM, PCIe or NVL
The variant matters more than most listings make it look. All three use the same Hopper chip, but memory speed and the link between GPUs are very different:
| Variant | Memory | Memory bandwidth | GPU-to-GPU link | Where it makes sense |
|---|---|---|---|---|
| H100 SXM | 80 GB HBM3 | 3.35 TB/s | NVLink, 900 GB/s | Multi-GPU training, tensor-parallel inference |
| H100 PCIe | 80 GB HBM2e | 2.0 TB/s | PCIe Gen5, 128 GB/s | Single-GPU inference, fine-tuning, notebooks |
| H100 NVL | 94 GB HBM3 | 3.9 TB/s | NVLink bridge, 600 GB/s | Inference that needs the extra 14 GB |
Specifications from NVIDIA's H100 datasheet.
The rule is simple: match the variant to how many GPUs the job really uses. For work that fits on one card, the PCIe H100 is usually the cheaper option and handles notebooks, fine-tuning and moderate inference well. Token generation in an LLM is mostly limited by memory bandwidth, so the SXM card's 3.35 TB/s against 2.0 TB/s on PCIe is the number that tends to show up in tokens per second. Once a model is split across GPUs with tensor parallelism, or a training run syncs gradients every step, the 900 GB/s NVLink on SXM nodes stops being a spec sheet line and starts deciding how long you pay for.
What fits in 80 GB
These are weights-only sizes of the published checkpoints, calculated rather than measured. The space left over is what vLLM has for the KV cache after its default 92% memory limit, and that cache grows with context length and the number of concurrent requests.
| Model | Format | Weights | On one H100 80 GB |
|---|---|---|---|
| gpt-oss-120b | MXFP4 | 65.2 GB | Fits, about 8 GB left for the KV cache |
| Gemma 4 31B | BF16 | 62.5 GB | Fits, about 11 GB left |
| Qwen3.8-27B | BF16 | 55.6 GB | Fits, about 18 GB left |
| Llama 3.3 70B | BF16 | 141.1 GB | Does not fit. Use 2x H100 with FP8 weights, or 4x in BF16 |
| gpt-oss-20b | MXFP4 | 13.8 GB | Fits easily, but a 48 GB card costs less |
With 8 GB of headroom, gpt-oss-120b runs on one H100 but not with long contexts for many users at once. If that is the plan, an H200 with 141 GB gives the cache room to grow.
How long until it is ready
A VM usually reaches Running in 1 to 3 minutes. App templates take about 6 to 8 minutes, because the platform waits until the app itself answers before it marks the deployment ready. One thing catches people out: on some providers the machine reports Running a little before SSH accepts connections, so a Connection refused in the first minute usually clears on its own. The Connection refused fix covers what to check if it does not.
Paid time starts when provisioning starts, but only a launch that reaches Running is charged. A launch that fails refunds the reserved hour automatically.
One GPU or eight
H100 VMs come in fixed sizes of 1, 2, 4 or 8 GPUs, and the GPUs column shows what is in stock right now. The Interconnect column says whether a multi-GPU VM has NVLink between its cards. One VM is the largest unit DeployGPU launches: there is no self-serve multi-node cluster. For more than 8 H100s, or a fixed term at a fixed price, request a cluster quote and a person replies by email. The multi-GPU page goes deeper on how GPUs inside one node talk to each other.
Software on the machine
In the dashboard, every H100 VM boots an Ubuntu image with NVIDIA drivers and Docker installed by the provider, and you connect over SSH as the user shown on the deployment. Three app templates set up a ready app on a private HTTPS link instead: JupyterLab with PyTorch, ComfyUI and Open WebUI with Ollama. For an OpenAI-compatible server on your own H100, follow the vLLM guide.
What you pay
The price on each row is the hourly rate you are billed, including any platform fee. Launching reserves the first hour from your credit balance, and when you stop, the unused seconds of that hour go back to the balance. Credits are bought by card with a 5% + $0.30 processing fee on each top-up. The pricing page has the full breakdown.
Stopping an H100 deletes its disk, so copy results off first. The guide to stopping a GPU without losing work shows how.
Frequently asked questions
How much does an H100 cost per hour on DeployGPU?
The table on this page lists the current hourly price of every H100 VM you can launch, cheapest first, refreshed every 15 minutes. The price shown is the rate you are billed, including any platform fee. The only other charge is the card processing fee on credit top-ups.
Can I rent a single H100?
Yes, whenever a 1x offer is in stock. The GPUs column shows how many H100s come with each VM, and you pay only for the time the VM runs.
Can I rent an 8x H100 node?
Yes, when one is in stock. Use the 8x link above to filter the marketplace. For more than 8 GPUs or a fixed term, request a cluster quote instead.
Should I pick H100 SXM or H100 PCIe?
Pick PCIe for single-GPU work where price matters most. Pick SXM when a job spans several GPUs or when token generation speed matters, since SXM has 3.35 TB/s of memory bandwidth against 2.0 TB/s and NVLink at 900 GB/s between GPUs.
How long until I can SSH into an H100?
Most VMs reach Running in 1 to 3 minutes. On some providers SSH starts accepting connections a little after the machine reports Running, so retry for a minute before assuming something is wrong.
What happens if no H100 is available?
The table says so and links to the marketplace. Stock on these providers changes by the minute. An H200 gives more memory and an A100 80GB costs less if you cannot wait.
Do I need a contract to rent an H100?
No. Every H100 here is on demand with no minimum term. You prepay credits, launch, and stop whenever you want.