Cheapest cloud GPU in 2026: the lowest price for each card, crypto clouds included
There is no single cheapest cloud GPU, whether you pay by card elsewhere or, as on Petaflare, only in crypto. The answer changes with the card, with whether the machine can be taken back, and with what is billed besides the GPU. Here are the lowest prices we found for seven cards, and how to read them.
The cheapest cloud GPU depends on the card and on whether you accept a machine that can be taken back. On 19 November 2026, marketplaces of independent hosts had the lowest prices for most cards: a Vast.ai H100 SXM at a $1.33 median, an A100 PCIe at $0.51. Among fixed, on-demand prices, Petaflare's were the lowest of the nine providers we checked on all seven cards we rent, from $0.34 an hour for an RTX 4090 to $4.89 for a B200. The big public clouds charged the most, $6.88 to $12.29 per H100-hour.
Lowest price per card on 19 November 2026
One GPU, per hour, from the official pricing pages of RunPod, Lambda, Massed Compute, Thunder Compute, Salad, Vast.ai, AWS, Azure and Google Cloud. Fixed prices and marketplace or interruptible prices are kept apart, because they are not the same product.
| GPU | Petaflare, fixed | Next-lowest fixed price | Lowest marketplace or interruptible price |
|---|---|---|---|
| RTX 4090 24GB | $0.34 | $0.89, RunPod Secure Cloud | $0.16, Salad lowest priority (interruptible) · $0.34, RunPod Community Cloud |
| RTX 5090 32GB | $0.49 | $1.39, RunPod Secure Cloud | $0.25, Salad lowest priority (interruptible) · $0.46, Vast.ai median |
| L40S 48GB | $0.87 | $0.97, Massed Compute | $0.53, Vast.ai median |
| A100 80GB | $0.99 | $1.09, Thunder Compute | $0.51, Vast.ai median (PCIe) · $0.76 (SXM) |
| H100 80GB | $2.39 | From $2.73, Massed Compute · $2.89, RunPod Secure Cloud (PCIe) | $1.33, Vast.ai median (SXM) |
| H200 141GB | $3.29 | $3.62, Massed Compute (H200 NVL) | $3.17, Vast.ai median (H200 NVL) · $3.50 (H200) |
| B200 180GB | $4.89 | $7.69, Lambda (per GPU, 8-GPU machine) | $4.95, Google Cloud spot (read 29 October) · $5.98, RunPod Community Cloud |
Petaflare at its list rate; the Fleet blocks take about 10% off. Vast.ai figures are the 18 November median from Vast.ai's public price feed, cited with attribution; each host sets its own price. Other sources: RunPod, Lambda, Massed Compute, Thunder Compute, Salad, Google Cloud.
Reading the table
Marketplace medians beat every fixed price on most cards, by about 40 to 50% on the H100, the A100 PCIe and the L40S. A median is the middle of many offers: half the hosts asked less, half asked more, and each offer comes with its own machine and its own terms.
There are two exceptions. On the RTX 4090, RunPod's Community Cloud matches our $0.34 and the Vast.ai median is $0.36, so a cloud RTX 4090 costs about the same everywhere except at the big clouds. On the B200, our $4.89 was the lowest price we found of any kind, spot included.
Among fixed prices, the next-lowest provider asked from 10% more than us on the A100 and H200 to more than two and a half times our price on the RTX 4090 and RTX 5090. That is the gap the cloud GPU pricing page is built on.
Three kinds of cheap GPU rental
- Fixed, on demand. A set price per hour from a provider that runs the machines, and nobody takes the GPU back while you pay. Our rates are in this group, with RunPod's Secure Cloud, Lambda, Massed Compute, Thunder Compute and the big clouds.
- Marketplaces and community clouds. Independent hosts list their own machines and prices: Vast.ai, RunPod's Community Cloud, Salad's network of GPUs in homes and small businesses. Prices are often the lowest, but the host, the network and the reliability change from one offer to the next.
- Interruptible. Spot, preemptible or low-priority capacity, sold cheaply because the provider can stop it for someone paying more. Salad's lowest priority rents an RTX 4090 for $0.16 an hour, and Azure's spot price for a one-GPU H100 NVL machine was $1.42. Good for jobs that save checkpoints and restart, painful for interactive work.
If a marketplace already suits you, our page on alternatives to Vast.ai weighs what changes when you move to a fixed price; if you use RunPod, see an alternative to RunPod.
Most of the cheap offers above are paid by card. RunPod and Vast.ai also take crypto, through payment processors; Petaflare takes nothing else, from $0.34 an hour for an RTX 4090, with no ID check. See renting a GPU with crypto.
Charges that change the real price
The hourly GPU price is rarely the whole bill. What each provider says about disks, data transfer and billing, on its own pages, read in November 2026.
| Provider | Disk | Data transfer | Billing |
|---|---|---|---|
| Vast.ai | Billed while the instance exists, even stopped | Every byte sent or received, at a price each host sets | Per second |
| RunPod | Container disk $0.10 per GB a month; volume disk $0.10 running, $0.20 idle | No data transfer fees | Per second |
| Thunder Compute | Billed per minute while running; snapshots billed separately | No egress fees | Per minute |
| Salad | No persistent disk: containers restart from their image | Not stated on its pricing page | Per second while running; stops if preempted |
| Lambda | Local SSD included in the instance | No egress fees | Hourly rate plus sales tax or VAT |
| Massed Compute | Local storage included in each configuration | No bandwidth charges | Hourly rate |
| AWS, Azure, Google Cloud | Block storage billed separately | Data sent out to the internet billed per GB | Varies by cloud and service |
| Petaflare | Local disk included, kept while the machine is stopped | No separate line on our price list | Per minute from prepaid GPU-hours; smallest block $25.50 to $29.70 |
Sources: Vast.ai billing, RunPod pricing and billing, Thunder Compute pricing, Salad pricing, Lambda pricing, Massed Compute pricing, AWS EC2 pricing, data transfer section. A stopped Petaflare machine uses no GPU-hours.
Cheapest per hour or cheapest per job
What you pay for a job is the hourly price times the hours it takes. A faster card costs more per hour but runs fewer hours, so the cheaper card per hour is not always the cheaper card per job. If an H100 finishes a fine-tuning run in 10 hours ($23.90) and an A100 needs 20 hours ($19.80), the A100 wins even though it is twice as slow. If the A100 needs 30 hours ($29.70), the H100 wins.
The table below gives the break-even point for pairs of cards at our prices: the pricier card is cheaper per job only if it finishes the job at least that many times faster.
Break-even speed for each upgrade
| Cheaper card | Faster card | Hourly prices here | Faster card must be at least |
|---|---|---|---|
| RTX 4090 | RTX 5090 | $0.34 → $0.49 | 1.44 × faster |
| L40S | A100 80GB | $0.87 → $0.99 | 1.14 × faster |
| A100 80GB | H100 80GB | $0.99 → $2.39 | 2.41 × faster |
| H100 80GB | H200 141GB | $2.39 → $3.29 | 1.38 × faster |
| H100 80GB | B200 180GB | $2.39 → $4.89 | 2.05 × faster |
Ratio of the list prices. To compare two cards, run the same job for 15 minutes on each, note the images, tokens or training steps per second, and divide each hourly price by that rate. A job that does not fit in a card's memory has no price on that card at all.
Cheap GPUs on the big clouds
AWS, Azure and Google Cloud charge the most for flagship cards, but they have cheaper corners.
- Spot capacity. On 19 November, Azure's spot price for the NC40ads H100 v5 in East US, one H100 NVL, was $1.42 an hour against $6.98 on demand. Its eight-GPU H100 machine came to about $2.27 per GPU as spot. Spot machines can be evicted whenever Azure needs the capacity.
- Smaller cards. AWS's g6.xlarge with one NVIDIA L4 costs $0.80 an hour, and Google Cloud's g2-standard-4 with an L4 cost $0.71 on 29 October. The L4 has 24 GB like the RTX 4090, but it is a 72 W inference card with 300 GB/s of memory bandwidth, against 1,008 GB/s for the RTX 4090.
- Whole servers. A cheap GPU cloud server is not something the big clouds sell: eight H100s cost $55.04 an hour on AWS (p5.48xlarge) and $98.32 on Azure (ND96isr H100 v5), with large CPUs, memory and networking included. The same eight cards cost $19.12 an hour here.
The big clouds also bill data leaving their network and often ask for a quota increase before you can launch GPU machines.
When the cheapest option is the wrong one
- Interactive work. A notebook or a ComfyUI session that gets interrupted loses everything that was only in memory.
- Long runs without checkpoints. A 30-hour job evicted in its 25th hour has cost 25 hours for nothing.
- Data you care about. On a community or marketplace host, your files sit on a machine run by someone you do not know, unless you filter for vetted data-center hosts.
- Very short tests. Our smallest block is $25.50 to $29.70 depending on the card. For one hour of testing, a provider that bills by the second costs less; with us, the hours you do not use stay on your account and never expire.
- Payment hurdles. Some low-cost providers want a card on file, or identity checks before they accept crypto. How each one handles it is on renting a GPU with crypto.
Cheap cloud GPU questions
Which cloud provider is the cheapest?
For the lowest price at any cost, marketplaces such as Vast.ai and interruptible tiers such as Salad's lowest priority. For a fixed price that nobody can take back, Petaflare had the lowest of the nine providers we checked on 19 November 2026, on all seven cards we rent. AWS, Azure and Google Cloud were the most expensive per GPU.
What is the cheapest GPU for AI?
The cheapest card your model fits on. For image generation and language models up to about 8B parameters in BF16, a 24 GB RTX 4090 at $0.34 an hour is usually enough. A 70B model in 4-bit needs 48 GB or more, such as an L40S at $0.87 or an A100 80GB at $0.99.
How much does a GPU server cost?
By the hour, from $2.72 for eight RTX 4090s to $39.12 for eight B200s here, at the list rate. Eight H100s cost $19.12 an hour here, $36.72 on Lambda ($4.59 per GPU) and $55.04 on AWS.
Is there any free cloud GPU?
Yes, with limits. Notebook services such as Google Colab and Kaggle give free time on older GPUs, with session and usage limits, and some providers give trial credits. They suit learning more than long jobs; see free GPU tiers and their limits.
Can I rent a cheap GPU without a credit card?
Yes. Petaflare only takes cryptocurrency, with no identity check, from $0.34 an hour. Vast.ai and RunPod also accept crypto through payment processors, each with its own conditions.
Sources
- Vast.ai public GPU price feedread 19 November 2026
- Vast.ai, billing referenceread 19 November 2026
- RunPod pricingread 19 November 2026
- Lambda pricingread 19 November 2026
- Massed Compute pricingread 19 November 2026
- Thunder Compute pricingread 19 November 2026
- Salad pricingread 19 November 2026
- Azure Retail Prices APIread 19 November 2026
- AWS EC2 on-demand pricingread 19 November 2026
- Google Cloud GPU pricingread 29 October 2026
- NVIDIA L4 Tensor Core GPUread 19 November 2026