Petaflare

GPU server rental with crypto: whole machines with 1 to 8 GPUs

When one card is not enough, rent a machine with two, four or eight whole NVIDIA GPUs of the same model, by the hour and paid only in crypto, for distributed training or for serving several models at once.

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

GPU server rental on Petaflare means a machine with 1, 2, 4 or 8 whole GPUs of one model, with root SSH, billed by the minute from the GPU-hours you buy. An 8-GPU server costs from $2.48 an hour with RTX 4090s to $35.20 an hour with B200s at the Fleet rate; eight H100s come to $17.20 an hour. You pay in crypto with no ID check, Bitcoin included, and you can stop the machine whenever the job is done.

What an 8-GPU server costs

Hourly price of a machine with eight cards of each model, at the list rate and at the Fleet rate (10% off), and the cost of keeping it running for 30 days.

30 days = 720 hours × 8 GPUs = 5,760 GPU-hours, priced at the Fleet rate. Buying whole Fleet blocks rounds the hours up; what is left stays on your account. Prices of 19 November 2026, on the full price list.
8 × GPUTotal GPU memoryPer hour, list ratePer hour, Fleet rate30 days non-stop at Fleet rate
RTX 4090192 GB$2.72$2.48$1,785.60
RTX 5090256 GB$3.92$3.52$2,534.40
L40S384 GB$6.96$6.24$4,492.80
A100 80GB640 GB$7.92$7.12$5,126.40
H100 80GB640 GB$19.12$17.20$12,384.00
H200 141GB1,128 GB$26.32$23.68$17,049.60
B200 180GB1,440 GB$39.12$35.20$25,344.00

30 days = 720 hours × 8 GPUs = 5,760 GPU-hours, priced at the Fleet rate. Buying whole Fleet blocks rounds the hours up; what is left stays on your account. Prices of 19 November 2026, on the full price list.

Paying for a multi-GPU server in crypto

An 8-GPU H100 machine uses 8 GPU-hours every hour, so the 800-hour Fleet block at $1,720 runs it for 100 hours. Blocks are paid in crypto only; at this size, the Bitcoin network fee is a small share, and USDC or USDT on Ethereum keeps the price fixed in dollars. See GPU server with Bitcoin.

What you get, and what we do not claim

  • Whole GPUs of one model. 1, 2, 4 or 8 cards, never a time-shared slice or a MIG partition, and never two different models in one machine.
  • Root SSH with Ubuntu and a PyTorch with CUDA image, a CUDA-only image, or your own Docker image.
  • A local disk that stays while the machine is stopped and is wiped when you delete it.
  • The host's country and the cards' form factor (SXM or PCIe) shown in your dashboard once the machine runs.

We do not promise bare metal, a given CPU model, a given amount of RAM or a disk size: those depend on the host our team prepares, and we have not published them yet. Check them as soon as you connect with nproc, free -h and df -h, and stop the machine at once if they do not suit you; you only pay for the minutes it ran.

Picking 1, 2, 4 or 8 cards

  • One card when the model and its working memory fit on it. It is the cheapest and simplest setup, with no communication between GPUs.
  • Two or four cards to split a model that does not fit on one card, such as a 70B model in BF16 on two H200s, or to serve several copies of a smaller model.
  • Eight cards for distributed training of large models, or to serve the biggest open models in FP8.

More cards finish a training run sooner but do not make it cheaper: eight GPUs for 10 hours use the same 80 GPU-hours as one GPU for 80 hours. They also add communication overhead, so scaling is never perfectly linear. Memory needs per method are in the fine-tuning guide.

Before you start a distributed job

How the cards are connected decides how fast they share data. SXM H100s, H200s, B200s and A100s talk over NVLink; PCIe cards such as the L40S and the GeForce cards go through the PCIe bus. Check what your machine has before launching a long run:

# How the GPUs are linked: NV# = NVLink, PIX/PXB/PHB/SYS = PCIe paths
nvidia-smi topo -m

# NVLink state on SXM machines
nvidia-smi nvlink --status

# Measure all-reduce bandwidth across 8 GPUs (github.com/NVIDIA/nccl-tests)
./build/all_reduce_perf -b 8M -e 1G -f 2 -g 8

# Start a PyTorch job on all 8 cards
torchrun --nproc_per_node=8 train.py
Run these right after connecting; on PCIe machines, expect lower all-reduce bandwidth and prefer data-parallel jobs.

Renting by the hour or a dedicated GPU server

Classic dedicated GPU servers are rented by the month, often with a setup fee and a minimum term. That suits a machine that will be busy around the clock for a year. Here you pay for GPU-hours: a training run that takes three days costs three days, and the machine can be stopped and started again from its disk. For work that comes in bursts, renting by the hour almost always costs less. How a cloud GPU differs from a physical one is covered in what a cloud GPU is. Single cards and every model we offer are on the main GPU rental page.

GPU server questions

How much does it cost to rent a GPU server?

It depends on the card and the number of cards: from $0.34 an hour for one RTX 4090 to $39.12 an hour for eight B200s at the list rate. The Fleet blocks take 10% off.

Can I rent a GPU cluster of several machines?

We rent machines of up to 8 GPUs. For more, you can run several machines, but we do not promise a fast network between them; ask support before planning multi-node training.

Are the GPUs shared with other customers?

No. Every GPU in your machine is yours alone for as long as it runs.

How long does a multi-GPU machine take to prepare?

A person on our team prepares each machine, so there is a wait we do not promise yet. The meter starts only when the SSH address appears; see how renting works.