GPU cloud providers compared: 11 options, one H100 price, and who takes crypto
Petaflare, a GPU cloud paid only in crypto, compares here who rents GPUs, what one H100 costs at each of them on the same day, how they bill, and what they ask before you can pay. Our own offer is in the list, with what it does not do.
GPU cloud providers come in three families. The big public clouds (AWS, Microsoft Azure, Google Cloud) charged $6.88 to $12.29 per H100-hour on 19 November 2026. GPU clouds such as Lambda, DigitalOcean (which now owns Paperspace), RunPod's Secure Cloud, Massed Compute, Thunder Compute and Petaflare charged $2.39 to $4.89. Marketplaces of independent hosts, such as Vast.ai and RunPod's Community Cloud, were lowest, with a $1.33 Vast.ai median. The right provider depends less on price than on what you need around the GPU: clusters, serverless endpoints, payment without a card, or the lowest price at any cost.
GPU cloud providers side by side
One NVIDIA H100 on demand, per GPU-hour, from each provider's own pricing and billing pages, read on 19 November 2026 unless stated.
| Provider | Family | One H100, per GPU-hour | Billing | Paying |
|---|---|---|---|---|
| Microsoft Azure | Public cloud | $12.29 on the 8-GPU VM · $6.98 for one H100 NVL | Varies by service | Card or invoice; no crypto |
| Google Cloud | Public cloud | $11.06 on the 8-GPU VM (read 29 October) | Varies by service | Card or invoice; no crypto |
| AWS | Public cloud | $6.88 | Varies by service | Card or invoice; no crypto |
| Lambda | GPU cloud | $4.89 (1 × SXM) · $4.59 on 8-GPU machines | Hourly rate, plus sales tax or VAT | Major credit card; prepaid and debit refused |
| DigitalOcean (Paperspace) | GPU cloud | $4.41 (HGX H100) | Per second, 5-minute minimum | Card, per Paperspace's pricing page |
| RunPod Secure Cloud | GPU cloud | $2.89 (PCIe) · $4.39 (SXM) | Per second | Card, or crypto after any required KYC |
| Thunder Compute | GPU cloud | $3.20 (SXM) | Per minute | Not stated on its pricing page |
| Massed Compute | GPU cloud | $3.14 (SXM5) · family from $2.73 | Hourly rate | Not stated on its pricing page |
| Petaflare | GPU cloud | $2.39 (SXM or PCIe) | Per minute, from prepaid blocks | Crypto only; no ID, no card |
| RunPod Community Cloud | Marketplace | $1.99 (PCIe) · $2.69 (SXM) | Per second | As RunPod |
| Vast.ai | Marketplace | $1.33 median (SXM) | Per second | Card; crypto through BitPay and Crypto.com |
| Salad | Community cloud | No H100; RTX 4090 from $0.16 | Per second while running | Not stated on its pricing page |
Public-cloud prices are whole machines divided by their GPUs, with CPUs, memory and local disk included. Payment rules for RunPod, Vast.ai and Lambda are from their billing documentation read on 13 November 2026. Sources are listed at the end of the page.
How we compared them
We took one card that almost every provider rents, the H100 80GB, and read its on-demand price on each provider's official page on the same day. Where a provider only sells 8-GPU machines, we divided the machine price by eight. Marketplaces have no single price, so Vast.ai is shown as the median of the day from its public price feed, cited with attribution. Spot and reserved prices are left out of the table and discussed below. Billing and payment come from each provider's own documentation, not from reviews. For the lowest price on each of seven cards rather than the H100 alone, see the cheapest cloud GPU comparison.
The big clouds: AWS, Azure and Google Cloud
The three hyperscalers rent every current NVIDIA data-center card, usually inside large machines that bundle many CPU cores, a lot of memory and fast networking, and they connect those machines to storage, databases and identity services. That is what you pay for: one H100 costs $6.88 to $12.29 an hour there, against $2.39 to $4.89 at the GPU clouds below.
- AWS sells H100s in P5 instances, H200s in P5en, B200s in P6, A100s in P4d and P4de, L40S cards in G6e and L4 cards in G6. Prices per instance and per GPU are on AWS GPU pricing.
- Azure offers ND H100 v5 machines with eight H100s, NC H100 v5 machines with one or two H100 NVL cards, and A100 machines in the ND and NC A100 v4 series. See Azure GPU pricing.
- Google Cloud attaches GPUs to A2, A3, A4 and G2 machine types: the L4, A100, H100, H200 and B200, plus older cards such as the T4. See Google Cloud GPU pricing.
All three sell spot capacity at large discounts, and all three usually ask for a quota increase before you can launch big GPU machines.
GPU clouds: Lambda, DigitalOcean, Massed Compute, Thunder Compute
GPU clouds sell GPU time first and price it per GPU-hour, with fewer services around it.
Lambda, long known as Lambda Labs, rents one H100 SXM for $4.89 an hour or $4.59 per GPU on eight-GPU machines, and B200s from $7.69 per GPU. It also sells 1-Click Clusters and reserved capacity, states that it charges no egress fees, and adds sales tax or VAT to its prices. Its billing documentation asks for a major credit card and refuses prepaid and debit cards.
Paperspace is now part of DigitalOcean, as its own site says. Its hourly machines are billed by the hour, storage is still billed while a machine is off, and payment is by card through Stripe. DigitalOcean's GPU Droplets, at prices effective 1 August 2026, rent an HGX H100 at $4.41 per GPU-hour, an H200 at $4.47, an L40S at $1.57 and AMD's MI300X at $2.59. They are billed per second with a five-minute minimum, and a powered-off Droplet is billed until it is destroyed. If you want a Paperspace alternative without a card, a fixed price per GPU-hour paid in crypto is what we offer.
Massed Compute lists one H100 SXM5 at $3.14 an hour with 14 vCPUs, 122 GB of memory and 350 GB of storage, and says it does not charge for bandwidth. Thunder Compute lists an H100 at $3.20, bills per minute, has no egress fees, and its A100 80GB at $1.09 was one of the lowest fixed A100 prices we found.
Marketplaces: Vast.ai, RunPod and Salad
Marketplaces let independent hosts list their own machines and set their own prices, which is why they are often the cheapest and always the most varied.
Vast.ai is an open marketplace: the median H100 SXM was $1.33 an hour, but each offer has its own host, reliability, storage price and bandwidth price, and storage is billed even while an instance is stopped. It takes crypto through BitPay and Crypto.com and does not refund credits bought with crypto. Our page on Vast.ai pricing and alternatives goes through its fees.
RunPod runs two clouds side by side: Secure Cloud in partner data centers at fixed prices, and Community Cloud on third-party hosts for less. It also sells serverless GPU endpoints. Its billing documentation says any required KYC must be completed before a first crypto payment, and its credits are not refundable. RunPod alternatives and RunPod vs Vast.ai compare it in detail.
Salad runs containers on consumer GPUs in homes and small businesses, such as the RTX 4090 and RTX 5090, from $0.16 an hour at its lowest priority. Lower priorities can be preempted, which suits retryable batch work.
RunPod's closest competitors are the providers on this page: Vast.ai and Salad on price, Lambda and DigitalOcean on fixed capacity, and the big clouds for everything around the GPU.
Best GPU cloud provider for each need
No provider is best at everything. Here is where each one is strongest, based on what it offers and charges, not on rankings.
| If you need | Look at | Why |
|---|---|---|
| The lowest price, whoever the host is | Vast.ai | Lowest median prices on most cards |
| Cheap batch jobs that can restart | Salad | Consumer GPUs from $0.16 an hour at lowest priority |
| A fixed price without a card or an ID check | Petaflare | Lowest fixed price we found on seven cards, paid in crypto |
| Serverless inference endpoints | RunPod | Sells serverless GPU endpoints next to its pods |
| Large training clusters | Lambda, or a big cloud | 1-Click Clusters and reserved capacity |
| AMD Instinct GPUs | DigitalOcean | MI300X at $2.59 and MI325X at $3.80 per GPU-hour |
| GPUs next to your existing cloud data | AWS, Azure or Google Cloud | No data leaves the cloud; spot capacity for savings |
Choosing a cloud GPU for deep learning and machine learning
Whichever provider you pick, the same checks decide whether a deep learning or machine learning job runs well and costs what you expect:
- Memory first. The model, its optimizer states and a batch of data must fit on the card or across the cards. A card that is too small has no price, however cheap.
- How the cards are linked. Multi-GPU training needs NVLink on SXM machines; PCIe cards share data more slowly. Across several machines, look for a fast network such as InfiniBand.
- Where the data lives. Copying a dataset in once is cheap; moving it out of a big cloud is billed per GB, and some providers bill disks even while the machine is stopped.
- The software stack. Check the CUDA and PyTorch versions of the ready images. Blackwell cards such as the B200 and RTX 5090 need CUDA 12.8 or newer.
- Billing and minimums. Per-second billing suits short experiments; prepaid blocks and monthly plans suit longer runs.
- Interruptions. Spot and low-priority machines need checkpoints. A fixed on-demand machine does not.
- What it takes to pay. Card on file, identity checks, quotas or crypto: this often decides which provider you can use today.
If you are new to renting, start with what a cloud GPU is.
Where Petaflare fits: GPUs paid in crypto, and what we do not do
We rent GPUs by the hour at fixed prices: whole NVIDIA cards, 1 to 8 per machine, from the RTX 4090 to the B200, with root SSH. Our H100 at $2.39 was the lowest fixed on-demand price in this comparison, and the full list of what a GPU-hour costs has every card. Payment is crypto only, with no identity check, which is why we cover GPU rental without KYC in detail.
What we do not offer, so you can rule us out quickly:
- no multi-machine clusters with a fast network between them;
- no serverless endpoints, managed notebooks or Windows desktops;
- no AMD cards and no cheap interruptible tier;
- machines are prepared by a person, so they do not start in seconds;
- the smallest block of hours costs $25.50 to $29.70.
Who makes the GPUs
Almost every GPU in this comparison is made by NVIDIA: the H100, H200 and B200 for large models, the A100 and L40S a generation or two back, and GeForce cards such as the RTX 4090 and RTX 5090 at the low end. AMD's Instinct MI300X and MI325X are the main alternative, rented here by DigitalOcean. The big clouds also design their own AI chips, such as Google's TPUs and AWS's Trainium and Inferentia, but those are not GPUs and need software written or adapted for them.
GPU cloud provider questions
Which cloud providers offer GPUs?
The big public clouds (AWS, Azure, Google Cloud), GPU clouds (Lambda, DigitalOcean, RunPod's Secure Cloud, Massed Compute, Thunder Compute, Petaflare and others) and marketplaces (Vast.ai, RunPod's Community Cloud, Salad). Oracle Cloud and CoreWeave also rent GPUs.
What does Paperspace do?
Paperspace rents cloud machines with GPUs and runs Gradient, a notebook service. It is now part of DigitalOcean, which also sells its own GPU Droplets, with an H100 at $4.41 per GPU-hour since 1 August 2026.
Who are RunPod's main competitors?
Vast.ai and Salad on price, Lambda and DigitalOcean on fixed GPU capacity, and AWS, Azure and Google Cloud for teams that need other cloud services. For a fixed price paid in crypto, Petaflare is another option.
What GPUs are available on Google Cloud?
Google Cloud attaches the L4, A100 (40 GB and 80 GB), H100, H200 and B200 to its G2, A2, A3 and A4 machine types, and still offers older cards such as the T4.
What is the best cloud GPU for deep learning?
The one your model fits on at the lowest cost per job. Large training runs use H100, H200 or B200 machines with NVLink; fine-tuning and smaller models often run well on an A100 80GB or an L40S; image models and small LLMs on an RTX 4090 or RTX 5090.
Sources
- Azure Retail Prices APIread 19 November 2026
- Google Cloud GPU pricingread 29 October 2026
- AWS EC2 on-demand pricingread 19 November 2026
- Lambda pricingread 19 November 2026
- Lambda, managing billingread 13 November 2026
- DigitalOcean GPU Droplet pricingread 19 November 2026
- Paperspace pricingread 19 November 2026
- RunPod pricingread 19 November 2026
- RunPod, billing documentationread 13 November 2026
- Thunder Compute pricingread 19 November 2026
- Massed Compute pricingread 19 November 2026
- Vast.ai public GPU price feedread 19 November 2026
- Vast.ai, billing referenceread 19 November 2026
- Salad pricingread 19 November 2026