How to read a GPU price
A published "dollars per GPU per hour" figure is only a starting point. Check, at least:
- The exact GPU and interconnect. An H100 PCIe card and an H100 SXM are not the same machine. Memory, vCPUs and disk also differ by provider.
- On-demand, interruptible or reserved. Interruptible or spot capacity is usually cheaper and can be reclaimed. Reserved or committed terms are usually cheaper per hour than walk-up on-demand, if you will actually use the hours.
- Billing increment, and whether a powered-off instance still costs money.
- Data transfer and storage. Some providers publish no egress fees; others bill them separately. Always read the current pricing page.
If you can cover the work with free or matched credits, start there. Our guide to free GPU and cloud credits explains the programmes, and the compute credits list shows what is open now. This article is about what you pay when you are buying hours.
Official on-demand prices, as of 18 November 2026
The figures below are taken from each provider's own pricing page on 18 November 2026. They are list or default published rates, not a promise of what you will be quoted, and they are not adjusted for reliability or for extra services.
| Provider | Published H100-class on-demand rate | How billing works | What the price is |
|---|---|---|---|
| Runpod (Secure Cloud) | H100 PCIe $2.89/hr; H100 SXM $4.39/hr; H100 NVL $3.19/hr [source] | Pods can be billed per hour or per second, as shown on the pricing page, which was updated 12 November 2026. | List rates for Runpod's Secure Cloud pod tier. Runpod also offers a Community Cloud tier; those rates were not the figures shown on the default table we captured, so check the live toggle on the same page. |
| DigitalOcean GPU Droplets | NVIDIA HGX H100 $4.41 per GPU per hour [source] | Billed per second, with a five-minute minimum. A 12-month reserved H100 plan is published at $3.26 per GPU per hour. Spot plans are listed for some other GPUs. Powered-off Droplets are still billed, because disk, CPU, RAM and the IP address stay reserved. | On-demand list price. DigitalOcean says new pricing took effect on 1 August 2026, and that prices are subject to change. |
| Lambda GPU cloud | 1x NVIDIA H100 SXM $4.89 per GPU per hour; 8x H100 $4.59; 1x H100 PCIe $3.79 [source] | Lambda says you pay by the minute, with no egress fees. Sales tax, VAT or GST may apply. | Self-serve on-demand list prices on Lambda's GPU cloud page. Multi-GPU instances have their own per-GPU rates. |
| Vast.ai | No single official list price [source] | Per-second billing. Hosts set prices. The official pricing page offers on-demand, interruptible (described as 50%+ cheaper and preemptible) and reserved terms of 1, 3 or 6 months. | A marketplace across 40+ data centres, so the rate for a given GPU moves with supply and demand. Check the live rates on Vast.ai's pricing page rather than a figure copied from elsewhere. |
On those official H100 list prices, Runpod's published Secure Cloud H100 PCIe rate is the lowest of the fixed-price providers above. That does not make it the cheapest choice for every job: the card, the cloud tier, the region and whether the instance can be interrupted all change the comparison. Vast.ai may be lower or higher than any of those list prices on a given day, because the price is set by the host.
Other GPUs, and when a cheaper card is enough
Not every research job needs an H100. Official prices for some other widely used cards, from the same pages and the same date:
- Runpod Secure Cloud: A100 PCIe $2.09/hr; A100 SXM $2.19/hr; L40S $1.09/hr; RTX 4090 $0.89/hr; RTX 5090 $1.39/hr [source].
- DigitalOcean: NVIDIA L40S $1.57/hr; RTX 6000 Ada $1.57/hr; RTX 4000 Ada $0.76/hr. DigitalOcean also lists AMD Instinct MI300X at $2.59/hr and MI325X at $3.80/hr on demand [source].
- Lambda (1x instances): A100 40 GB $1.99 per GPU per hour; A6000 $1.09; A10 $1.29; GH200 $2.59; B200 SXM6 $7.99 [source].
A 24 GB consumer card is often enough for smaller fine-tunes, inference experiments or student work. Large-model training or multi-node jobs are a different comparison, and may be cheaper on reserved or committed capacity. DigitalOcean and Lambda publish reserved or multi-GPU rates; Runpod tells customers to speak to its enterprise sales team for reserved capacity [source] [source] [source].
Credit programmes that can cut the bill
Several of the same providers, and the large public clouds, also run credit programmes. The published amounts on listings we currently track include:
- Runpod Startup Program: Starter Tier: $1,000 in credits; Growth Tier: $25,000 bonus credits when you commit $50,000.
- Vast.ai Startup Program (GPU credit matching): Dollar-for-dollar credit match on verified GPU compute spend, up to an approved amount.
- Lambda Research Grant: Up to $5,000 in Lambda cloud credits, for qualifying researchers.
- AWS Activate Credits and Google for Startups Cloud Program: AI track, which we compare in Google, AWS and Azure credits for AI startups.
Credits expire, usually cannot be withdrawn as cash, and often exclude some services. Read the current terms on the organiser's page. Other GPU and inference credits are covered in GPU and inference credits beyond the big three.
A practical way to choose
- See whether a credit programme covers the work, or part of it, before you pay list price.
- Name the GPU you actually need, including memory and whether you need SXM or NVLink, then compare like with like.
- If the job can be checkpointed, look at interruptible or spot rates and at Vast.ai's interruptible offers, and budget for restarts.
- If the job cannot be interrupted, compare on-demand or reserved rates and read the uptime terms.
- Re-check the provider's pricing page on the day you start, because these figures move.
