Hopper · data center GPU
Rent a dedicated NVIDIA H100 SXM server for $1,249 a month
The H100 SXM is the GPU the current generation of language models was trained on: 80 GB of HBM3 at 3.35 TB/s, 989 TFLOPS of dense FP16 tensor compute, FP8 through the Transformer Engine and 900 GB/s of NVLink between GPUs. This is the SXM5 part — not the slower PCIe card — delivered as a dedicated bare-metal server with its own 24 vCPU, 256 GB of RAM and 3 TB of NVMe, at one flat monthly price.
- Per GPU, per month$1,249$1.71/h effective · fixed for your term
- Same GPU, hourly on-demand$2,446market median $3.35/h × 730 h
- You keep−49%$1,197 a month, every month
Specification
NVIDIA H100 SXM server specifications
One dedicated machine, sized around the GPU. Nothing is shared, nothing is metered, and every line below is included in the monthly price.
Why teams rent the H100 here
- SXM5, not PCIe. Full 3.35 TB/s HBM3 and a 700 W power budget, versus 2 TB/s of HBM2e and 350 W on the PCIe card that shares the name.
- FP8 Transformer Engine. Mixed FP8/FP16 training and FP8 inference with automatic scaling — the format every modern LLM stack targets first.
- NVLink 4 at 900 GB/s. Tensor parallelism across the GPUs of a node without the PCIe bottleneck; 8× H100 nodes with InfiniBand are on the clusters page.
- One flat invoice. $1,249 a month covers the GPU, the host, the NVMe, 20 TB of transfer and 24/7 support — the same number every month, fixed for your term.
Below about 373 hours of use a month, an hourly provider is the cheaper way to run this GPU. Above it — and a machine that trains, serves or renders is above it — the monthly rate wins, and the gap is the $1,197 shown above.
Pricing
H100 price per month, by term
Per GPU, in USD, excluding VAT. Longer terms cost less; every term includes the same dedicated host, network and support.
| Term | Per GPU, per month | Effective hourly | vs on-demand | Order |
|---|---|---|---|---|
| MonthlyRolling month-to-month. Cancel with 30 days' notice. | $1,249 | $1.71/h | −49% | Deploy |
| 3 months5% off the monthly rate for a 3-month term. | $1,187 | $1.63/h | −51% | Deploy |
| 6 months10% off the monthly rate for a 6-month term. | $1,124 | $1.54/h | −54% | Deploy |
| 12 months15% off the monthly rate for a 12-month term. | $1,062 | $1.45/h | −57% | Deploy |
| 8× H100 nodeNVLink between the GPUs · NVLink 4 · 3.2 Tb/s InfiniBand · 192 vCPU · 2 TB RAM · 30 TB NVMe | $7,990 | $999/GPU | −20% | Cluster details |
On-demand comparison: market-median published hourly rate for the same GPU ($3.35/h) over 730 hours. Block storage $25/TB/month and extra IPv4 addresses $4/month are optional add-ons.
Market comparison
H100 rental price compared
Every provider in our audit that publishes a price for the H100, read on 15 October 2026 from their own pricing page: hourly rates multiplied by 730 hours — what the same GPU costs kept for a month — and monthly prices as listed.
| Provider | Published rate, H100 | One month, 24/7 | vs our $1,249/mo |
|---|---|---|---|
 GPU Cloud HQDedicated, billed monthly |
$1,249 per month, flat | $1,249 | Our reference |
 Microsoft AzureOn-demand, hourly |
$98.32/h for ND96isr H100 v5 (8× H100, East US) — $12.29 per GPU-hour · Spot $2.27 | $8,972 | −86% |
 Google CloudOn-demand, hourly |
$88.49/h for a3-highgpu-8g (8× H100) — $11.06 per GPU-hour · Spot $6.60 | $8,074 | −85% |
 AWS (Amazon EC2)On-demand, hourly |
$6.88/h for p5.4xlarge (1× H100) · $55.04/h for eight | $5,022 | −75% |
 CoreWeaveOn-demand, hourly |
$49.24/h per 8-GPU HGX H100 node — $6.16 per GPU-hour | $4,493 | −72% |
 NebiusOn-demand, hourly |
$4.50/h on demand · spot from $0.79 | $3,285 | −62% |
 DigitalOcean (Paperspace)On-demand, hourly |
$4.41/GPU/h on demand · $3.26 with a 12-month commitment | $3,219 | −61% |
 Verda (DataCrunch)On-demand, hourly |
$4.17/h on demand · $2.08 spot | $3,044 | −59% |
 RunPodOn-demand, hourly |
$3.99/h Secure Cloud (SXM) · $2.69 Community | $2,913 | −57% |
 LambdaOn-demand, hourly |
$3.99–$4.29/h (SXM) depending on instance size | $2,913 | −57% |
 Together AIOn-demand, hourly |
$3.99/h on demand · $3.69 reserved 3–30 days · $1.99 preemptible | $2,913 | −57% |
 Crusoe CloudOn-demand, hourly |
$3.90/GPU-h on demand (HGX) | $2,847 | −56% |
 ScalewayOn-demand, hourly |
€6.62/h for H100-SXM-2-80G (2× H100 SXM) — €3.31 ($3.71) per GPU-hour · H100 PCIe €2.87/h | $2,708 | −54% |
 JarvislabsOn-demand, hourly |
$3.49/h on demand (SXM) · $1.54 spot | $2,548 | −51% |
 Novita AIOn-demand, hourly |
$3.39/h (SXM) · $1.70 spot | $2,475 | −50% |
 HyperstackOn-demand, hourly |
$3.20/h SXM on demand · from $2.72 reserved | $2,336 | −47% |
 Thunder ComputeOn-demand, hourly |
$3.20/h (SXM), 1–8 GPUs | $2,336 | −47% |
 Massed ComputeOn-demand, hourly |
$3.14/h (SXM5) · $25.12/h for eight | $2,292 | −46% |
 OVHcloudCloud, monthly billing |
$2,070/mo for h100-380 (1× H100 PCIe) · $2.99/h by the hour | $2,070 | −40% |
 HOSTKEYDedicated, monthly |
€1,590/mo ($1,774) — 1× H100 80 GB virtual server, pre-order | $1,774 | −30% |
 VultrOn-demand, hourly |
$2.30/GPU/h for an 8× HGX H100 instance | $1,679 | −26% |
 TensorDockOn-demand, hourly |
from $2.25/h (SXM5) | $1,643 | −24% |
 Vast.aiOn-demand, hourly |
median $2.20/h (14 Oct) · offers from $1.73 | $1,606 | −22% |
 LeasewebDedicated, monthly |
CA$2,078.74/mo ($1,463) — 1× H100 NVL 94 GB (PCIe), Montreal | $1,463 | −15% |
 Voltage ParkOn-demand, hourly |
from $1.99/h for a single H100, no contract | $1,453 | −14% |
| Market median30–40 providers, GetDeploying index | $3.35/h on demand | $2,446 | −49% |
Rates are each provider's standard on-demand tier for the H100 — not spot, preemptible or community hardware — or its listed monthly price where it publishes one, as printed on their pricing page on 15 October 2026; each row links to the full comparison with its source. Where a provider is cheaper than us for a full month, the row says so. Method and caveats: GPU cloud alternatives; every GPU at a glance: GPU cloud pricing.
Workloads
What people run on an H100
Training & fine-tuning — and the three jobs below are where a dedicated H100 at a flat monthly price earns its keep.
-
Training and fine-tuning
Full fine-tunes of 7B–13B models on two to four GPUs, 70B on an 8-GPU node with DeepSpeed and optimizer offloading, checkpoints on local NVMe.
-
Production inference
FP8 serving of models up to about 35B on one GPU with vLLM or TensorRT-LLM, or 70B across two, at a flat cost that does not change with traffic.
-
Diffusion and video generation
SDXL, FLUX and video models at batch, where HBM bandwidth — not CUDA cores — sets the frames per second.
Is the H100 the right GPU for you?
The H100 SXM is the default choice for training and fine-tuning anything up to 70B parameters and for FP8 inference of models that fit in 80 GB. Take the H200 if inference is bandwidth- or memory-bound, and the A100 if the workload has no use for FP8 and budget matters more than speed.
Recommended for model training — the sizing guide there shows what fits in 80 GB HBM3.
Compare the H100 side by side with the cards people weigh it against — specifications, price per GB and the model sizes that fit — or check what your own model needs with the GPU memory calculator.
- H100 vs A100
- H200 vs H100
- B200 vs H100
- L40S vs H100
- RTX 4090 vs H100
- RTX 5090 vs H100
- RTX PRO 6000 vs H100
At FP8 with an 8k context, one H100 runs these open models on its own — each page has the memory by context length and the fine-tuning needs:
H100 rental: questions answered
How much does it cost to rent an H100 server?
$1,249 a month, flat, for a dedicated NVIDIA H100 SXM server with 24 vCPU · 256 GB RAM · 3 TB NVMe and a 25 Gbps uplink — $1.71 per hour effective over 730 hours. A 3-, 6- or 12-month term takes 5%, 10% or 15% off: $1,187, $1,124 or $1,062 a month. The market-median on-demand rate for the same GPU is $3.35 per hour, about $2,446 for a full month.
Is the H100 in stock right now?
Yes. 14 of the current batch of 64 were unallocated at the last stock update (18 September 2026). Provisioning is automated: your server is imaged, secured with your SSH key and handed over in under 10 minutes after your payment confirms.
What is included with a H100 server?
Everything on the pricing table: the GPU, 24 vCPU · 256 GB RAM · 3 TB NVMe, a 25 Gbps uplink with 20 TB of outbound transfer, a dedicated public IPv4 and a /64 IPv6 block, DDoS protection, root access and Ubuntu 24.04 with the NVIDIA driver, CUDA 12 and Docker — or PyTorch, vLLM, Kubernetes images at checkout. Hardware replacement within 4 hours and a 99.9% uptime SLA are part of the contract.
Is this the H100 SXM or the H100 PCIe?
SXM5. The PCIe version carries HBM2e at 2 TB/s, a 350 W power limit and no NVLink in most builds; the SXM part has 3.35 TB/s of HBM3, 700 W and 900 GB/s of NVLink, and is what the published benchmarks refer to. Our price is for the SXM part, in a dedicated server.
Can I rent several H100s or a full 8-GPU node?
Yes: order 1, 2, 4 or 8 GPUs at once, subject to the units left in the batch, on the same monthly terms. The 8× H100 node — NVLink between the GPUs, 3.2 Tb/s InfiniBand, NVLink 4 · 3.2 Tb/s InfiniBand · 192 vCPU · 2 TB RAM · 30 TB NVMe — is priced separately at $7,990 a month on the clusters page.
Do I need KYC or a card to rent it?
No. An email address is all we ask for. You fund a prepaid balance in Bitcoin, Ether or USDT (ERC-20 or TRC-20); the first invoice settles from it and monthly renewals charge to it automatically. No ID document, no card, no bank transfer.
Other data center GPUs by the month
Same dedicated hosts, same monthly model, same 10-minute provisioning. All 10 GPUs are on the pricing page.
Deploy an H100 in under 10 minutes
$1,249 a month, fixed for your term. Fund your balance in BTC, ETH or USDT and the server is imaged, secured with your key and handed over automatically.