Petaflare

H100 rental with crypto: rent an NVIDIA H100 80GB by the hour

Whole H100 cards with 80 GB of memory, one to eight per machine, with root SSH and a PyTorch, CUDA or Docker image, paid only in crypto and metered by the minute.

Crypto-only payment, no card, no KYCUpdated 19 November 2026Prices checked 19 November 2026

Yes, you can rent an NVIDIA H100 here today: H100 rental costs $2.39 per GPU-hour, or $2.15 on the 800-hour Fleet block. The smallest block is 12 GPU-hours for $28.68. Each machine has 1, 2, 4 or 8 whole H100 80GB cards and root SSH; the card is either the SXM or the PCIe version, and your dashboard shows which one before you connect. You pay in crypto with no ID check, and the meter only runs while your instance does.

H100 specs: SXM and PCIe side by side

NVIDIA sells the H100 80GB in two forms. The SXM module sits on an HGX board with NVLink between cards; the PCIe card goes in a standard server slot with less bandwidth and power. Both have 80 GB.

Architecture
Hopper
Memory
80 GB HBM3 (SXM) or HBM2e (PCIe)
Memory bandwidth
3.35 TB/s (SXM) · 2.0 TB/s (PCIe)
BF16 Tensor (dense)
989 TFLOPS (SXM) · 756 TFLOPS (PCIe)
FP8 Tensor (dense)
1,979 TFLOPS (SXM) · 1,513 TFLOPS (PCIe)
Board power
up to 700 W (SXM) · 350 W (PCIe)
GPU link
NVLink 900 GB/s (SXM)
Form factor
SXM or PCIe, shown before you connect

NVIDIA figures from the official H100 page. Tensor figures are without sparsity.

H100 price per hour across clouds

The H100 rents for anywhere between $1.33 and $4.89 per GPU-hour, depending on the form factor and on whether the price is a fixed on-demand rate or a marketplace offer. Here is what each provider showed on its own pricing page on 19 November 2026, for one card unless stated.

Sources: RunPod, Lambda (plus sales tax), Massed Compute, Thunder Compute, Vast.ai public price feed (18 November median, cited with attribution), Silicon Data H100 index. Read on 19 November 2026.
ProviderPrice typeH100 versionPer GPU-hour
PetaflareFixed, on demandSXM or PCIe, shown before you connect$2.39 · $2.15 on Fleet
RunPod Secure CloudFixed, on demandPCIe · NVL · SXM$2.89 · $3.19 · $4.39
RunPod Community CloudThird-party hostsPCIe · SXM$1.99 · $2.69
LambdaFixed, on demandPCIe · SXM · 8 × SXM (HGX)$3.79 · $4.89 · $4.59
Massed ComputeFixed, on demandSXM5$3.14
Thunder ComputeFixed, on demandSXM$3.20
Vast.aiMarketplace, median of the daySXM · PCIe$1.33 · $2.00
Silicon Data indexComposite of neo-clouds and marketplacesH100$2.53

Sources: RunPod, Lambda (plus sales tax), Massed Compute, Thunder Compute, Vast.ai public price feed (18 November median, cited with attribution), Silicon Data H100 index. Read on 19 November 2026.

Reading those H100 prices

Our $2.39 is below every fixed on-demand H100 price in the table and close to the Silicon Data index. It is not the lowest price you can find. A Vast.ai host or RunPod's Community Cloud can be cheaper, because those cards belong to third parties who set their own terms; Vast.ai also bills storage and bandwidth on top. Both take crypto only through payment processors, and RunPod's billing documentation says any required identity check must be done before a first crypto payment.

If you need a specific form factor, note that we do not let you choose between SXM and PCIe. For multi-card training where NVLink matters, ask support before you buy which version the machine will have.

Paying for H100 hours in crypto

The smallest H100 block is 12 GPU-hours for $28.68; the 800-hour Fleet block costs $1,720, or $2.15 an hour. Pay in USDT on Tron, usually credited in about a minute, or in Bitcoin, Ether, USDC, Solana, Litecoin or Monero. Crypto is the only way to pay and no card or ID is asked: see how to pay with USDT or Bitcoin.

What fits in the H100's 80 GB

A model's weights take 2 bytes per parameter in FP16 or BF16, 1 byte in FP8 and about half a byte in 4-bit. Serving also needs room for the KV cache, which grows with batch size and context length. Full fine-tuning with the Adam optimizer needs roughly 16 bytes per parameter for weights, gradients and optimizer states.

Rules of thumb before activations and framework overhead; measure your own job before buying a large block. The fine-tuning memory guide and the LLM serving guide work through the sums.
JobMemory for weights or training stateOne H100 80GB?
Serve an 8B model in BF16About 16 GB, lots of room for the cacheYes
Serve a 70B model in FP8About 70 GB, little left for the cacheYes, with short contexts and small batches
Serve a 70B model in BF16About 140 GBNo: 2 H100s, or one H200 with 141 GB at a squeeze
QLoRA fine-tune of a 70B modelAbout 40 to 48 GB with 4-bit weightsYes
Full fine-tune of an 8B model with AdamAbout 128 GBNo: 2 H100s, or LoRA on one
Train a 1B model from scratchAbout 16 GB plus activationsYes

Rules of thumb before activations and framework overhead; measure your own job before buying a large block. The fine-tuning memory guide and the LLM serving guide work through the sums.

One to eight H100s in one machine

An instance uses one GPU-hour per card for each hour it runs. Eight cards at the Fleet rate cost less per hour than eight at the Burst rate, so plan big jobs on the Fleet block.

Hourly cost is the GPU-hour rate times the number of cards. More on 8-card machines on the GPU server page.
MachineTotal GPU memoryPer hour at $2.39Per hour on Fleet ($2.15)
1 × H10080 GB$2.39$2.15
2 × H100160 GB$4.78$4.30
4 × H100320 GB$9.56$8.60
8 × H100640 GB$19.12$17.20

Hourly cost is the GPU-hour rate times the number of cards. More on 8-card machines on the GPU server page.

SXM or PCIe: what changes for you

On one card the two versions run the same software and hold the same 80 GB, but the SXM module is faster: 3.35 TB/s of memory bandwidth against 2.0 TB/s, and about 30% more Tensor throughput, because it runs at up to 700 W against 350 W. Memory-bound jobs, like serving an LLM, feel the bandwidth gap most.

With several cards the gap widens. SXM cards talk over NVLink at 900 GB/s; PCIe cards go through the PCIe bus unless they have a bridge. Data-parallel training of a model that fits on one card loses little on PCIe; tensor-parallel training of a model split across cards loses more.

Your dashboard shows the form factor and the host's country as soon as the instance runs, before you have used a minute of real work.

Rent or buy an H100?

Take a purchase price and divide it by the hourly rate. If an H100 cost you $25,000, that money buys about 10,460 H100 hours here at $2.39, or 11,630 hours at the Fleet rate: 14 to 16 months of one card running day and night. A card you own also needs a server, power, cooling and someone to run it, and it loses value as newer cards arrive.

Renting wins for jobs that run in bursts, for trying a model before you commit, and when you need eight cards for a week rather than one card for two years. Buying wins when one card will be busy most hours of the next two or three years.

H100 or another card?

  • A100 80GB at $0.99: same memory, no FP8 and about a third of the BF16 throughput, at less than half the price. Good for BF16 training on a budget.
  • H200 141GB at $3.29: the same Hopper chip with 76% more memory and 43% more bandwidth. Pick it when 80 GB is the limit.
  • B200 180GB at $4.89: roughly twice the H100's FP8 throughput and more than twice its memory bandwidth.

All seven cards are on the price list. On AWS the H100 comes in 8-card p5 machines; AWS GPU pricing and Azure H100 prices put those rates per GPU-hour.

H100 rental questions

How much does it cost to rent an H100 per hour?

$2.39 per GPU-hour on the Burst and Session blocks, $2.27 on the 200-hour Project block and $2.15 on the 800-hour Fleet block.

Can I rent a single NVIDIA H100?

Yes. A machine can hold 1, 2, 4 or 8 H100s, and the smallest purchase is 12 GPU-hours for $28.68.

Do I get an H100 SXM or an H100 PCIe?

Either one: the machine you get has one or the other, and the dashboard shows which before you start working. If you need SXM for NVLink, ask support before buying.

What is the difference between the A100 and the H100?

Both have 80 GB. The H100 adds FP8 support and two and a half to three times the A100's BF16 Tensor throughput, with more memory bandwidth; the A100 costs less than half as much per hour here.

Which software comes on the machine?

Ubuntu with the NVIDIA driver and either PyTorch with CUDA or the CUDA toolkit alone. You can also start from your own Docker image.

Sources