Cheapest GPU Cloud in 2026: H100 and A100 Hourly Prices Compared

For one GPU running 100 hours a month, Hyperstack is the cheapest on-demand H100 at $2.50 an hour, or $250 for the month. It is also the cheapest A100 80GB at $1.35 an hour, or $135. Both prices are on Hyperstack's own pricing page, read on October 10, 2026. RunPod is next for an H100 at $2.89 an hour, while AWS and Azure charge more than twice Hyperstack's rate per GPU.

The hourly rate is only part of the bill. Some providers rent only eight-GPU nodes, and some bill CPU and memory on top of the GPU. This page puts eight providers on one basis: the on-demand price per GPU, dated, with the smallest unit you can rent shown next to it.

How we compared them

Every price was verified on vendor pages, Oct 2026, read on each provider's own pricing page or price list on October 10, 2026, in US dollars. The ranking uses on-demand rates only; spot and reserved rates come further down because they carry conditions. "Smallest on-demand unit" means the fewest GPUs a provider lets you rent at its listed rate. When a provider sells only eight-GPU nodes, its per-GPU figure is the node price divided by eight, and the node price sits next to it. The scenario is one H100 80GB and one A100 80GB, each running 100 hours a month, which is the load of a fine-tuning run or a batch inference job.

Devshot data: Toolradar evaluated 40 GPU cloud tools for its October 2026 GPU cloud ranking, and 45% of them have a free plan. This page ranks the paid hourly rates instead, because that is the number a buyer compares when a job needs a GPU for a fixed number of hours.

The method matches our object storage pricing comparison: dated vendor prices, one scenario, and the arithmetic shown once. Toolradar also keeps profiles for RunPod, Lambda, DigitalOcean, CoreWeave, AWS and Azure.

The on-demand table

Prices are per GPU per hour, cheapest H100 first. Modal bills per second, so its hourly figure is the per-second rate times 3,600.

The AWS and Azure rows come from the providers' public price lists for US East and East US, Linux pay-as-you-go. RunPod's figures are the pod card prices on its pricing page.

Among the providers whose pages show an eight-GPU H100 price, Lambda's $3.99 per GPU is the lowest, ahead of DigitalOcean's $4.41, CoreWeave's $6.16, AWS's $6.88 and Azure's $12.29. Lambda's 80GB A100 is sold only as an eight-GPU instance on its self-serve page, so its $2.79 comes with that minimum.

What 100 GPU-hours costs

Cost is the hourly rate times the hours: $2.50 x 100 = $250. The table below applies that to each provider's smallest unit. Where that unit is eight GPUs, the figure is the whole node.

Cost is linear in hours, so the ranking of rates holds at any monthly hour count. With several GPUs, the per-GPU rate matters more than the minimum. AWS and Azure are the two most expensive per GPU on this page, at $6.88 and $12.29.

Where each provider fits

Short single-GPU jobs suit Modal's per-second billing. A 10-minute H100 SXM5 job costs about $0.66 of GPU time at $3.95 an hour, before CPU and memory charges.

Long single-GPU runs favour the lowest hourly rate, which on this page is Hyperstack for both the H100 and the A100 80GB. Hyperstack bills per minute, so a run that stops at 90 minutes pays for 90 minutes.

Eight-GPU training is quoted two ways. CoreWeave, AWS and Azure quote the whole node, so divide by eight to compare. Lambda's eight-GPU H100 SXM is $3.99 per GPU, and DigitalOcean's eight-GPU Droplet is $4.41 per GPU. This table prices GPU time only, not the networking or support around it.

Spot and reserved rates

Spot capacity costs less and can be reclaimed by the provider, so it suits jobs that save checkpoints. Hyperstack's spot H100 PCIe is $2.00 an hour and its spot A100 is $1.08. CoreWeave's spot 8-GPU H100 node is $19.71 an hour, about $2.46 per GPU, and its A100 node is $9.65, about $1.21 per GPU.

Hyperstack also lists reserved rates starting at $1.75 for an H100 and $0.95 for an A100. Reserved rates suit a steady load you can plan months ahead, while spot suits interruptible batch work.

What the hourly rate leaves out

• Modal bills CPU and memory on top. Its CPU rate is $0.0000131 per physical core per second, about $0.047 per core-hour. Memory is $0.00000222 per GiB per second, about $0.008 per GiB-hour. A job with 8 physical cores and 64 GiB adds about $0.89 an hour to the $3.95 GPU rate.
• Lambda's prices exclude tax. They are listed "plus applicable sales tax/VAT/GST", so the bill rises once tax applies.
• Node minimums. CoreWeave, AWS and Azure show their H100 and A100 80GB rates for eight-GPU nodes, so a one-GPU job pays for eight GPUs. Lambda's 80GB A100 works the same way.
• Storage and transfer. The table covers compute only, so check each provider's storage and transfer lines before you commit budget. For the same kind of comparison on general-purpose servers, see DigitalOcean vs Hetzner.

Five checks before you rent

Read the vendor's own page with these questions in mind, because a low headline rate can hide a larger bill.

• The billing unit. Confirm whether the rate is per GPU or per node. For a job that needs one GPU, a node price is eight times the GPU's share of the invoice.
• The minimum. Check whether the cheapest rate needs a minimum count of GPUs, as Lambda's 80GB A100 does at eight.
• The tax line. Check whether tax is added to the listed rate, as it is on Lambda's page.
• The side costs. Compute is one line of the bill. Add the storage you keep between runs and the data you move out of the provider.
• The reclaim rule. For spot, read what happens to a running job when capacity is taken back, then decide whether your job can resume from a saved checkpoint.

H100 or A100: is the cheaper card worth it?

At Hyperstack the A100 80GB costs $1.15 an hour less than the H100, a 46% saving at the same provider. The saving is real only if the A100 finishes your job at a speed you can accept, which depends on your model, framework and batch size. This page does not benchmark either card.

Frequently asked questions

What is the cheapest H100 GPU cloud in 2026?

Hyperstack is the cheapest on demand, as read on its pricing page on October 10, 2026. RunPod's H100 PCIe at $2.89 an hour is next, and Lambda's single-GPU PCIe is $3.29. Hyperstack's spot H100 PCIe is $2.00 an hour, but spot capacity can be reclaimed.

What is the cheapest A100 80GB cloud?

Hyperstack is the cheapest, at $1.35 an hour for its A100 80GB, on the same pricing page. RunPod's A100 PCIe at $1.79 an hour is next, and Modal's per-second A100 80GB rate is $0.000694 a second before CPU and memory charges.

How much does an H100 cost per hour on AWS and Azure?

AWS's p5.48xlarge is $55.04 an hour for eight H100s, or $6.88 per GPU. Azure's ND96isr H100 v5 is $98.32 an hour for eight, or $12.29 per GPU. Both are Linux on-demand rates for eight-GPU nodes in US East and East US.

Is an A100 cheaper than an H100?

Yes, at every provider in the table that lists both. At Hyperstack the A100 80GB runs $1.15 an hour below its H100. Whether that gap is a real saving depends on how much longer the A100 takes to finish your job.

Is there a free GPU cloud?

Toolradar's GPU cloud ranking shows that 45% of the 40 tools it evaluated have a free plan. This table ranks paid on-demand rates only, so check any free plan's limits before you rely on it for H100 time. For free tiers across the big clouds, see cloud free tiers compared. If you need a model rather than a GPU you manage, free LLM APIs covers that route.

Is spot pricing worth it for training?

Spot cuts the hourly rate, and the trade-off is interruption. Hyperstack's spot H100 PCIe is $2.00 an hour, and CoreWeave's spot 8-GPU H100 node is $19.71 an hour against $49.24 on demand. Save checkpoints often enough that an interrupted run resumes instead of restarting.

Cite this: Devshot, "Cheapest GPU Cloud in 2026: H100 and A100 Hourly Prices Compared", October 2026.

Rental rates change without notice, so check the vendor page before you commit a budget. Devshot's newsletter tracks the price moves that developers act on, at devshot.co.

Some offers on this page may be paid placements or contain affiliate links.