Petaflare

Cloud GPU pricing per GPU-hour, paid in crypto

One rate per card, paid only in crypto, sold in blocks of GPU-hours that never expire. Every price below links straight to payment, and the calculator works out the cheapest way to cover a job.

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

Petaflare's cloud GPU pricing is a fixed rate per GPU-hour: RTX 4090 $0.34, RTX 5090 $0.49, L40S $0.87, A100 80GB $0.99, H100 $2.39, H200 $3.29 and B200 $4.89. You buy hours in four block sizes per card; the Project block is 5% cheaper per hour and the Fleet block 10% cheaper. Instances draw from those hours by the minute, multiplied by the number of GPUs, and only once their SSH details appear. The smallest payment is $25, and every block costs more than that.

Every GPU and every block, priced

Click a price to pay for that block. Each card name opens its page with specs, the market price next to ours and the jobs it fits.

Prices in US dollars, set on 19 November 2026. Hours are GPU-hours: one hour on four GPUs uses four. The Burst block of every card is the smallest purchase.

Price a job before you pay

Choose the card, how many GPUs go in the machine and how long it will run. The calculator compares every block, and multiples of each, and keeps the cheapest. The button opens payment for exactly that.

petaflare/ new instancemetered per minute
GPU
GPUs in the instance
Machine image

Estimate

Rate × GPUs
$2.39 × 1
Per hour
$2.39/hr
Needed
24 GPU-hours
You pay
$114.72

Best buy: Session (48 GPU-hours)

Rent H100 · $114.72

No charge until the SSH details of your instance appear. Unused GPU-hours stay on your account.

What real jobs cost

The same rule as the calculator, applied to jobs people often price before renting. The cheapest purchase is not always the biggest block: for 24 hours on one H100, two Burst blocks beat one Session block.

Worked out on 19 November 2026 from the grid above. Hours left over stay on your account for the next job on the same card.
JobGPUs × hoursGPU-hoursCheapest purchaseTotal
A day of fine-tuning on one H1001 × 24242 × Burst (12 h)$57.36
Distributed training on an 8-GPU H100 server8 × 1008001 × Fleet (800 h at $2.15)$1,720.00
A month of image generation on an RTX 40901 × 3003001 × Session (300 h)$102.00
An A100 80GB running non-stop for 30 days1 × 7207206 × Session (120 h)$712.80
A 40-hour render on an RTX 50901 × 40401 × Burst (60 h), 20 h left over$29.40
Serving a 70B model on two H200s for three days2 × 7214418 × Burst (8 h)$473.76
A week of video generation on an L40S1 × 1601606 × Burst (30 h), 20 h left over$156.60
A one-day sprint on eight B200s8 × 2419232 × Burst (6 h)$938.88

Worked out on 19 November 2026 from the grid above. Hours left over stay on your account for the next job on the same card.

Hourly GPU rental, counted by the minute

Renting a GPU by the hour here does not mean paying for whole hours. Each running instance takes from your GPU-hours of that model, minute by minute, at one GPU-minute per card in the machine.

  • Start: nothing is counted while our team prepares the machine. Counting starts the moment the SSH address appears in your dashboard.
  • Several GPUs: an instance with 4 GPUs uses 4 GPU-minutes every minute. Thirty minutes on 4 H100s uses 2 GPU-hours.
  • Stop: stopping the instance stops the count. Its disk is kept, so you can start it again later.
  • Running out: when a card's GPU-hours reach zero, its instances stop and keep their disks. Buy another block and start them again.
  • Delete: deleting an instance wipes its disk for good. Copy your checkpoints out first.

All instances of the same card draw from the same pool. Hours bought for H100s cannot run an RTX 4090, so buy blocks for the card you plan to use. What happens from payment to SSH shows each screen.

What the hourly price covers

Included in every block

  • Whole GPUs, never a shared slice or a partition of a card
  • 1, 2, 4 or 8 GPUs in one machine
  • Root SSH access
  • A PyTorch with CUDA image, a CUDA-only image, or your own Docker image
  • A local disk that is kept while the instance is stopped
  • GPU-hours that do not expire

Not part of the offer

  • No spot or interruptible pricing: nobody takes your GPU back while you have hours
  • No free tier: the Burst block is the smallest way to try a card
  • No reserved contracts or sales calls
  • No guaranteed delivery time: a person prepares each machine, and the count only starts when it is ready

Paying for GPU-hours in crypto

You pay each block in Bitcoin, USDT on Tron or Ethereum, USDC, Ether, Solana, Litecoin or Monero. No ID, card or email is asked. Send the exact amount shown at checkout; the hours are added once the network confirms the payment. Because every block is priced above the $25 minimum, the amount you send is the block price, with nothing left on the side. Networks and confirmations are on the crypto payment page, and the mistakes that cost money on how to pay with USDT or Bitcoin.

How these prices compare

On 19 November 2026 these rates were below RunPod's Secure Cloud on all seven cards and below Lambda wherever Lambda lists the card. They are not the lowest anywhere: marketplace hosts and community clouds rent some cards for less, with trade-offs in reliability, extra storage fees and payment checks. The provider-by-provider numbers are on the cheapest cloud GPU page, and multi-card machines are on GPU server rental. Ready to start? Rent a GPU from the home page, or price a job with the calculator above.

Pricing questions

How much is a GPU rental per hour?

From $0.34 per GPU-hour for an RTX 4090 to $4.89 for a B200. The Project and Fleet blocks of each card are 5% and 10% cheaper per hour.

Do I pay for a full hour if I stop after ten minutes?

No. GPU-hours are used by the minute, so ten minutes on one GPU uses one sixth of a GPU-hour.

Can I use H100 hours on an H200?

No. Hours belong to the card you bought them for, because each card has its own rate. Buy a block for each card you plan to run.

What if I buy more hours than I use?

They stay on your account with no expiry date. If an instance fails to start, the hours it used come back; see the refund policy.

Are there volume discounts?

Yes, built into the blocks: 5% off the hourly rate on the Project block and 10% off on the Fleet block, on every card. There are no other discounts or coupons.