On October 30, 2026, Lambda listed an eight-GPU H100 SXM instance at $3.99 per GPU-hour and CoreWeave listed an eight-GPU H100 node at $49.24 per node-hour in North America. At 720 uninterrupted hours, those compute-only examples are $22,982.40 and $35,452.80 respectively. They have different hardware configurations and are not equivalent service quotes. Runpod lists an H100 NVL pod at $3.19/hour; check availability, GPU form factor, networking, storage and actual terms before comparing.
Eight cloud providers compared on published H100 pricing, capacity type and the charges beyond GPU time.
Some links on this page are affiliate links: if you sign up through one, Toolradar may earn a commission. Commissions never change our ranking or what we write. How we make money
A GPU-hour is one GPU running for one hour. Eight GPUs running for a 30-day month use 5,760 GPU-hours, assuming 24 hours every day. A node-hour can include several GPUs together, so comparing its hourly price with a single GPU-hour without converting the units is misleading.
Toolradar data: our GPU cloud category tracks 44 products in the wider market. The count does not mean every product sells H100 capacity. Our methodology compared eight providers on publicly described GPU products, billing units and the published rates available on October 30, 2026. We did not rent or benchmark the hardware, and paid placement did not determine the order.
For the compute illustration we use eight GPUs for 720 uninterrupted hours. Lambda's listed rate is per GPU-hour for its eight-GPU configuration; CoreWeave's rate is per node-hour for an eight-GPU HGX H100 configuration in North America. The prices below exclude tax, storage or transfer beyond any included allocation, support, interruptions and other charges. Systems differ in CPU, memory, local storage, network and availability.
Top Picks
Selected to compare published GPU units and different cloud operating models. How we rate
GPU instances and clusters with public per-GPU pricing.
Starting price
Check instance pricing
Best for
A team comparing a named eight-GPU H100 SXM instance.
GPU cloud with published node, spot and inference prices.
Starting price
H100 node $49.24/hour on demand
Best for
A team that can compare a full HGX node and its stated regional rate.
GPU clusters with on-demand, preemptible and reserved rates.
Starting price
Check cluster pricing
4.0 (6 reviews)
Best for
A workload comparing cluster pricing and term commitments.
| Tool | Starting price | Rating | Best for |
|---|---|---|---|
| Lambda | Check instance pricing | n/a | A team comparing a named eight-GPU H100 SXM instance. |
| CoreWeave | H100 node $49.24/hour on demand | n/a | A team that can compare a full HGX node and its stated regional rate. |
| Together AI | Check cluster pricing | 4.0 6 reviews | A workload comparing cluster pricing and term commitments. |
| Runpod Visit | H100 NVL pod $3.19/hour shown | 4.1 345 reviews | A developer comparing a single H100 NVL pod with cluster offers. |
| DigitalOcean | Check current GPU Droplet pricing | 4.6 914 reviews | A team already using DigitalOcean that wants to compare its GPU Droplets. |
| AWS | Check regional EC2 pricing | 4.4 17,730 reviews | A team with AWS requirements that can price the relevant P5 instance and region. |
| Google Cloud | Check regional machine pricing | 4.6 556 reviews | A team evaluating H100 workloads within Google Cloud. |
| Microsoft Azure | Check regional VM pricing | 3.8 2,109 reviews | A team comparing H100 capacity within Azure infrastructure. |
- Starting price
- Check instance pricing
- Free plan
- No
Best for: A team comparing a named eight-GPU H100 SXM instance.
- Pro: The pricing page states the selected GPU count and hourly unit.
- Pro: It lists host CPU, memory and storage with the configuration.
- Pro: The vendor describes self-serve instances and larger systems.
- Con: Availability is first come and must be confirmed.
- Con: Tax and other usage can add to the compute figure.
- Starting price
- H100 node $49.24/hour on demand
- Free plan
- No
Best for: A team that can compare a full HGX node and its stated regional rate.
- Pro: The pricing table identifies an eight-GPU node and host resources.
- Pro: The page shows an on-demand and a spot example.
- Pro: Its unit is explicit for the selected node.
- Con: Spot capacity can be interrupted and its price can vary.
- Con: Confirm region, networking and support terms for the job.
- Starting price
- Check cluster pricing
- Free plan
- No
- Rating
- 4.0 (6 reviews)
Best for: A workload comparing cluster pricing and term commitments.
- Pro: The cluster table labels prices per GPU per hour.
- Pro: It shows both on-demand and preemptible H100 rates.
- Pro: Reserved terms are listed separately.
- Con: Confirm available GPU count and configuration before using the illustration.
- Con: Preemptible capacity is not a guaranteed uninterrupted month.
Runpod
GPU pods, serverless inference and cluster options.
- 3.8 on Trustpilot (308 reviews)
- 4.5 on G2 (37 reviews)
- Starting price
- H100 NVL pod $3.19/hour shown
- Free plan
- No
- Rating
- 4.1 (345 reviews)
Best for: A developer comparing a single H100 NVL pod with cluster offers.
- Pro: The pod table lists GPU memory and host resources.
- Pro: The vendor describes per-second pricing options.
- Pro: Pods and clusters have separate pricing sections.
- Con: A single NVL pod is not an eight-GPU SXM node.
- Con: The H100 SXM cluster row directs buyers to sales.
- Starting price
- Check current GPU Droplet pricing
- Free plan
- No
- Rating
- 4.6 (914 reviews)
Best for: A team already using DigitalOcean that wants to compare its GPU Droplets.
- Pro: The vendor lists single-H100 and eight-H100 configurations.
- Pro: Its product page describes host memory and storage.
- Pro: GPU Droplets integrate with other DigitalOcean services.
- Con: Confirm the rate, region and capacity for the required configuration.
- Con: Other services and transfer can add charges.
- Starting price
- Check regional EC2 pricing
- Free plan
- No
- Rating
- 4.4 (17,730 reviews)
Best for: A team with AWS requirements that can price the relevant P5 instance and region.
- Pro: AWS documents H100-based P5 instances.
- Pro: Its ecosystem offers compute, storage and networking services.
- Pro: Different purchasing options can be evaluated.
- Con: Use the current regional rate for the exact instance.
- Con: Storage, networking and support may be separate costs.
- Starting price
- Check regional machine pricing
- Free plan
- No
- Rating
- 4.6 (556 reviews)
Best for: A team evaluating H100 workloads within Google Cloud.
- Pro: Google documents H100 A3 machine families.
- Pro: Different A3 options are described for training and serving.
- Pro: The platform publishes region and quota documentation.
- Con: Prices depend on the selected machine and region.
- Con: Confirm accelerator quota and availability before scheduling a run.
Microsoft Azure
Cloud virtual machines with H100 accelerator options.
- 4.6 on Capterra (1,979 reviews)
- 4.3 on G2 (73 reviews)
- 1.6 on Trustpilot (57 reviews)
- Starting price
- Check regional VM pricing
- Free plan
- No
- Rating
- 3.8 (2,109 reviews)
Best for: A team comparing H100 capacity within Azure infrastructure.
- Pro: Microsoft documents its ND H100 v5 series.
- Pro: The platform offers connected cloud services.
- Pro: Machine configuration details are available in the documentation.
- Con: A current regional VM price is needed for a comparison.
- Con: Quota, networking and data services need separate review.
Other GPU Cloud tools worth considering
More published tools from our GPU Cloud category, ordered by our category ranking. They are not part of the editorial picks above.
What a GPU cloud rental includes
A GPU cloud provider rents access to an accelerator in a virtual machine, a dedicated node or a larger cluster. The workload might train a model, fine-tune it or serve inference. Prices can be on-demand, interruptible spot, reserved for a term, or negotiated for a large deployment.
An H100 is not one universal configuration. SXM and PCIe or NVL systems can have different memory, interconnect and host resources. A training job needing eight tightly connected GPUs should not assume that eight independent one-GPU pods provide the same topology. Ask for a written configuration and run a representative workload.
The 720-hour compute calculation
Lambda's listed eight-GPU H100 SXM rate gives $22,982.40 for 5,760 GPU-hours for the assumed month. CoreWeave's North America on-demand eight-GPU node price gives 720 × $49.24 = $35,452.80. The $12,470.40 difference is arithmetic on published compute prices, not a claim that one setup is equivalent or will be available for the full period.
CoreWeave also shows a North America spot rate of $19.71 per node-hour for the same named GPU count. Multiplying by 720 gives $14,191.20 if that displayed rate and uninterrupted capacity held throughout the month. Spot pricing and availability can change, and an interrupted training job may need to restart. Budget for checkpoints and recovery rather than treating that illustration as a guaranteed bill.
Key Features to Look For
GPU-hour versus node-hour (Essential)
Multiply a per-GPU rate by the number of GPUs; a node rate may already include all accelerators in the node.
GPU form factor (Essential)
Record the exact H100 configuration, GPU memory and whether the accelerator is SXM, PCIe or NVL.
Interconnect (Essential)
For distributed training, verify GPU-to-GPU and node-to-node networking rather than assuming eight single-GPU pods form one node.
Billing duration (Essential)
Check whether a stopped instance, allocated disk or retained IP continues to incur charges.
Capacity type (Essential)
Separate on-demand, spot or preemptible, and reserved capacity. A lower spot rate may come with interruption risk.
Storage and egress (Important)
Price checkpoints, datasets, retained volumes and outbound transfer beyond any included allocation.
Host resources (Important)
Compare CPUs, system RAM, local disk and the GPU count as a single configuration.
Availability and support (Important)
Confirm region, quota, lead time, support response and whether the provider can supply the needed capacity.
Select the capacity for the job
-
For a short experiment, check the minimum billing period, GPU availability and how to release a machine after the run.
-
For eight-GPU training, request the interconnect topology and per-node configuration rather than extrapolating eight separate pods.
-
For interruption-tolerant work, compare spot pricing with the cost of checkpoint storage and potential restarts.
-
For a long run, ask about reserved terms, cancellation, start dates and capacity guarantees in writing.
-
For inference, compare an endpoint's usage model with a dedicated machine before assuming the training price applies.
Evaluation Checklist
Record the exact model, GPU count, host memory and network requirements.
Estimate actual running hours and any periods when the machine can be released.
Convert all offers to node-hours and GPU-hours without double counting.
Add persistent storage, transfer, checkpointing and required support.
Test restart from a checkpoint before relying on interruptible capacity.
Confirm the current region, quota, availability and contract with the provider.
Pricing Overview
On-demand instance
A workload that needs the listed configuration and can obtain on-demand capacity.
From $3.99/GPU-hour in the Lambda eight-GPU example
Spot or preemptible
A job that can recover from interruptions and accepts variable availability.
$19.71/node-hour in the CoreWeave example
Longer commitment
A team that can commit to a defined term and confirm reserved capacity in writing.
Terms vary
Pricing Comparison
| Provider and configuration | Published price | 720-hour compute illustration |
|---|---|---|
Lambda H100 SXM, eight-GPU instance | $3.99 per GPU-hour | $22,982.40 for eight GPUs |
CoreWeave HGX H100, eight-GPU node, North America | $49.24 per node-hour on demand | $35,452.80 for one node |
CoreWeave HGX H100, same GPU count, North America spot | $19.71 per node-hour shown | $14,191.20 if unchanged and uninterrupted; not guaranteed |
Together AI HGX H100 GPU Cluster | $3.99 per GPU-hour on demand | $22,982.40 for eight GPUs if available under those terms |
Together AI H100 cluster, preemptible | $1.99 per GPU-hour shown | $11,462.40 for eight GPUs if uninterrupted; not guaranteed |
Runpod H100 NVL pod | $3.19 per hour shown | $2,296.80 for one pod; do not treat eight pods as one node |
DigitalOcean GPU Droplets | H100 and H100×8 configurations listed | Check region, availability and current price |
AWS, Google Cloud and Azure | H100-capable instance families documented | Use the current regional calculator or quote |
Vendor pages checked October 30, 2026: Lambda pricing, CoreWeave pricing, Together GPU cluster pricing, Runpod pricing, DigitalOcean GPU Droplets, AWS P5 instances, Google GPU machine types and Azure ND H100 v5. Arithmetic assumes 720 hours and excludes other charges.
Mistakes to Avoid
- ×
Treating a node-hour as a single GPU-hour.
- ×
Comparing PCIe or NVL pods with an eight-GPU SXM node as if the topology were identical.
- ×
Assuming a spot price guarantees an uninterrupted month.
- ×
Leaving datasets, checkpoint storage and egress out of the budget.
- ×
Assuming an advertised price means capacity is available in the needed region.
Expert Tips
- →
Run a short representative workload and measure time to completion.
- →
Price the complete node and verify how it is billed when stopped.
- →
Test checkpoint recovery before moving an interruptible job to spot.
- →
Compare effective cost per completed job, not just cost per GPU-hour.
- →
Ask for quota and capacity confirmation in writing.
Red Flags to Watch For
-
A price without GPU count, form factor or billing unit.
-
A monthly estimate that assumes spot never interrupts.
-
A quote that omits networking or persistent storage requirements.
-
A claim that eight independent pods necessarily have the same interconnect as one HGX node.
-
A guaranteed completion time without a workload benchmark.
The Bottom Line
The Lambda and CoreWeave examples show why GPU-hour and node-hour prices must be normalized. Their 720-hour compute calculations are useful starting points, but the systems and commercial terms differ.
Use Together AI and Runpod to compare capacity types and named configurations, and request current regional figures from DigitalOcean, AWS, Google Cloud and Azure. For adjacent decisions see MLOps tools, AI model deployment and local LLM tools.
Frequently Asked Questions
Is there a free option for H100 rental?
These examples are paid list rates. Providers may offer limited trial credits, but eligibility, hardware and expiry vary; confirm whether a specific credit applies to the H100 configuration before relying on it.
How much does an H100 cost per hour?
Lambda's pricing page listed the per-GPU rate in the comparison table for its eight-GPU H100 SXM configuration on October 30, 2026. The rate depends on the configuration, provider, region and terms.
How many GPU-hours are in 30 days on eight GPUs?
At 24 hours per day, 30 days is 720 hours. Eight GPUs running continuously use 5,760 GPU-hours.
What does an eight-GPU H100 node cost for 720 hours?
At CoreWeave's displayed North America on-demand price of $49.24 per node-hour, the compute portion is $35,452.80 before other charges.
Is spot GPU capacity guaranteed for a month?
No. Spot or preemptible capacity may be interrupted, and prices or availability may change. Check the provider's terms and test checkpoint recovery.
Are eight single-GPU pods equivalent to one eight-GPU node?
Not necessarily. GPU interconnect, host memory, networking and topology can differ. Confirm the requirements of your distributed workload.
What costs besides GPU time matter?
Check retained storage, data transfer, networking, support, tax, checkpointing and the cost of restarting interrupted jobs.
Cite this page: Toolradar, "GPU Cloud Pricing Compared in 2026: H100 Cost per Hour", updated October 2026, https://toolradar.com/guides/gpu-cloud-pricing-compared
Spotted an outdated price or a wrong fact? Suggest a correction: we check it against the vendor's own page. Work for one of these vendors? Claim your listing to update it yourself.
Related Guides
Ready to Choose?
Compare features, read reviews, and find the right tool.
Some offers on this page may be paid placements or contain affiliate links.