On-demand
$1.85/hr
For most work. Billed by the second with no minimum. Stop whenever you like and resume on the same disk and the same IP.
Pricing · updated 1 November 2026
Cloud GPU prices that include the whole machine, billed by the second.
The hourly price of each GPU already covers its vCPUs, RAM, local NVMe, a public IPv4 address and all the traffic you send or receive. The meter starts when your machine passes its health check and stops the second you stop it.
Prices in US dollars, excluding VAT. Paid from a prepaid balance topped up in crypto, from $25.
Lowest price
$0.23/hr
RTX 4090 · 24 GB
H100 SXM
$1.85/hr
Our priceMedian $3.20
Median of 40 GPU clouds, on demand
Persistent disk
$0.03per GB-month
The only other line on a receipt
01 / All GPUs
Each GPU is sold as its exact variant, with the machine around it.
Pick a mode and a unit. The hourly on-demand price is the one that appears on your receipt.
| GPU | VRAM | vCPU | RAM | NVMe | On-demand | Spot | Reserved | Price scale | Regions | Launch |
|---|---|---|---|---|---|---|---|---|---|---|
 B200 SXMBlackwell |
180 GB | 28 | 283 GB | 2 TB | $4.15 $0.001153 $3,030 | $2.49 $0.000692 $1,818 | $3.32 $0.000922 $2,424 | Frankfurt, Ashburn | Launch | |
 H200 SXMHopper |
141 GB | 24 | 276 GB | 2 TB | $2.49 $0.000692 $1,818 | $1.49 $0.000414 $1,088 | $1.99 $0.000553 $1,453 | Paris, Ashburn, Singapore | Launch | |
 H100 SXMHopper |
80 GB | 20 | 125 GB | 2 TB | $1.85 $0.000514 $1,351 | $1.11 $0.000308 $810 | $1.48 $0.000411 $1,080 | Paris, Frankfurt, Ashburn, Singapore | Launch | |
 MI300XAMD CDNA 3 |
192 GB | 24 | 240 GB | 2 TB | $1.65 $0.000458 $1,205 | $0.99 $0.000275 $723 | $1.32 $0.000367 $964 | Ashburn | Launch | |
 H100 PCIeHopper |
80 GB | 16 | 188 GB | 1 TB | $1.35 $0.000375 $986 | $0.81 $0.000225 $591 | $1.08 $0.000300 $788 | Frankfurt, Ashburn | Launch | |
 RTX PRO 6000Blackwell |
96 GB | 16 | 188 GB | 2 TB | $1.15 $0.000319 $840 | $0.69 $0.000192 $504 | $0.92 $0.000256 $672 | Paris, Ashburn, Singapore | Launch | |
 A100 SXMAmpere |
80 GB | 16 | 125 GB | 1 TB | $0.95 $0.000264 $694 | $0.57 $0.000158 $416 | $0.76 $0.000211 $555 | Paris, Frankfurt, Ashburn, Singapore | Launch | |
 L40SAda Lovelace |
48 GB | 16 | 94 GB | 1 TB | $0.55 $0.000153 $402 | $0.33 $0.000092 $241 | $0.44 $0.000122 $321 | Paris, Frankfurt, Ashburn, Singapore | Launch | |
 RTX 5090Blackwell |
32 GB | 9 | 35 GB | 500 GB | $0.48 $0.000133 $350 | $0.29 $0.000081 $212 | $0.38 $0.000106 $277 | Paris, Ashburn, Singapore | Launch | |
 RTX 4090Ada Lovelace |
24 GB | 8 | 41 GB | 500 GB | $0.23 $0.000064 $168 | $0.14 $0.000039 $102 | $0.18 $0.000050 $131 | Paris, Frankfurt, Ashburn | Launch |
Stock moves every few minutes. The live board shows how many cards you can launch in each region right now.
Planning 8 GPUs or more for a week or longer? Open a ticket from the console and we send a written quote within one business day.
02 / Per GPU
The price of each GPU, from one card to a full node.
H100 SXMHopper
![]()
- Memory
- 80 GB HBM3
- Interconnect
- NVLink 900 GB/s
- Per GPU
- 20 vCPU · 125 GB
H100 SXM · 80 GB HBM3 · NVLink 900 GB/s
$1.85/hr
$0.000514 per second, all-in
Price by mode
On-demand$1.85/hr
Reserved20 % off$1.48/hr
Interruptible40 % off$1.11/hr
Machine size
1 × H100 SXM$1.85/hr
Regions
Paris
Frankfurt
Ashburn
Singapore
Eight GPUs or more, for a week or longerAsk for a written quoteThe cards, the region, the dates and one price, within one business day.
The same GPUs elsewhereRead on 63 GPU cloudsEvery price here next to the ones other clouds publish, with sources and dates.
Data-center cards and the RTX PRO 6000 come in machines of 1, 2, 4 or 8 GPUs. The RTX 5090 and RTX 4090 come in machines of 1, 2 or 4. A machine of n GPUs costs n times the price of one, with n times the vCPU, RAM and NVMe.
03 / What is included
What one fine-tuning run cost, line by line.
Fine-tune · H100 SXM · eu-paris
Started 2026-10-02 09:14:07 UTC, health check passed after 47.8 s
Running time2 h 14 min 03 s (8,043 s)
Price per second$0.000514
GPU, 20 vCPU, 125 GB RAMincluded
2 TB local NVMe, public IPv4included
Traffic in and outincluded
Compute$4.13
Persistent disk, 200 GB for 3 days$0.59
Launch and health check$0.00
Total$4.72
Every line marked included is part of the GPU price. A persistent disk is billed only while you keep it, at $0.03 per GB-month of 730 hours, counted to the second. Stopping the machine stops the compute meter, and deleting the disk ends the last charge.
Never on a GPUBank receipt
- egress or ingress traffic
- a public IPv4 address
- CPU, RAM or NVMe as separate lines
- a minimum billing period
- a launch or health check that failed
- compute while the machine is stopped
04 / Billing modes
On-demand, reserved or interruptible (spot).
Reserved 20 % off
$1.48/hr
For production and deadlines. A fixed price for one week or one month, cards set aside for you, prepaid. You can cancel before the start date; once it has started, a reservation is not refundable.
Interruptible (spot) 40 % off
$1.11/hr
For batch jobs that save checkpoints. Five minutes of notice, your checkpoint hook runs first, and the machine is stopped, never deleted. A missed notice credits the last 30 minutes.
Reserved orders are placed from the console, under Capacity › Reserve.
05 / Estimate
Price a job before you start it.
Estimated cost
$23.58
compute $22.20 + disk $1.38
With a $100 top-up, credited as $135, this runs for 72 hours.
Top up and run this job
06 / Payment
A prepaid balance in dollars, topped up in crypto.
-
1
Choose an amount
From $25 to $5,000 per top-up.
-
2
Pay in the coin you hold
The amount in your coin is fixed for 60 minutes, then the payment is checked on chain.
-
3
Get credited in USD
With a bonus on larger top-ups. Usage is drawn by the second.
Coins accepted
- Monero
- Bitcoin
- Ethereum
- USDT
- USDC
- Solana
- Litecoin
- + 9 more
USDT and USDC on Tron, Ethereum, BNB Chain, Solana and Polygon. The full list, with the network fees, is in the top-up page of the console.
Bonus ladder
You top up · your balance is credited
You top upBonus creditedHeights on a square-root scale
- The balance never expires and, like a prepaid card, is not refundable except where the law requires.
- Set a monthly spend cap for the account, and a separate cap for each team member.
- When the balance reaches $0, machines stop and disks are kept for 14 days.
- A payment we cannot credit (late, short or a different amount) is sent back within 72 hours to the refund address you gave, minus the network fee.
- Every top-up and every session has a receipt in the console, and a monthly invoice groups them with your company name and VAT number.
- Card payments are not accepted for now.
07 / Market
What 40 other GPU clouds charge for the same H100 SXM.
On-demand list prices per GPU-hour, read on each provider's pricing page on 29 September 2026. The median covers the 40 offers sold on demand. Hyperscaler prices are for one GPU of an 8-GPU instance.
Four providers list it below us. Lium, Clore.ai and Vast.ai are marketplaces, where each host sets its own price, terms and hardware, and storage or traffic can be billed apart. CUDO Compute sells it at $1.79, six cents under our price. Compare GPUBank with 63 GPU clouds, line by line.
09 / FAQ
Questions about the price, answered before your first top-up.
We divide the hourly price by 3,600. An H100 SXM at $1.85 an hour costs $0.000514 a second, so a session of 8,043 seconds costs $4.13. The meter starts after the health check passes.
Only its persistent disk, at $0.03 per GB-month, for as long as you keep it. Compute, IPv4 and local NVMe cost nothing while the machine is stopped.
There is no minimum billing period: a 90-second test costs 90 seconds. The only minimum is the first top-up, $25.
They cost about 40 % less. We can take the card back with five minutes of notice, sent through the metadata endpoint, a webhook and the console. Your checkpoint hook runs, the machine is stopped with its disk intact, and you can resume on another card.
Yes. Reserved capacity is 20 % below on-demand for a week or a month. For 8 GPUs or more over a month or longer, open a ticket from the console and we reply with a written quote within one business day.
A price cut applies to running sessions the moment it takes effect. An increase is announced 30 days ahead and never applies to a session that is already running. Reserved orders keep their price for the whole term.
Prices exclude VAT. EU businesses with a valid VAT number are invoiced under the reverse charge, EU consumers pay the VAT of their country, and customers outside the EU are invoiced without VAT.
Not for now. Top-ups are in crypto: Bitcoin, USDT, Ethereum, Monero and more than a dozen other coins. For a reserved order above $5,000, we split the payment into several top-ups with you.
Top up $25 and launch your first GPU.
Once your balance is credited, a machine is ready in about 48 seconds (median). A $100 top-up is credited as $135.
