NVIDIA H100 pricing: rental rates and total cost

NVIDIA H100 pricing depends on the GPU variant, the number of GPUs you rent, and what the service includes. On October 19, 2026, Lambda listed a single H100 PCIe instance at $3.29 per GPU-hour and a single H100 SXM instance at $4.29. Modal listed H100 SXM5 GPU time at $0.001097 per second, approximately $3.95 per hour, with other resources charged separately. The provider links and billing details are in the comparison below.

These are dated examples, not a market-wide minimum or a guarantee of available capacity. A useful comparison starts with the complete bill for a workload that meets your requirements. This guide explains the units, purchase options, and calculations to check before renting or buying H100 hardware.

What does an H100 price include?

An H100 GPU-hour measures one GPU allocated for one hour. Eight GPUs running for ten hours represent 80 GPU-hours, even though only ten hours have elapsed. If a provider quotes the whole instance, multiply that instance rate by time once. Multiplying it by the GPU count again would double-count the hardware.

A virtual machine may bundle GPU access with CPU cores, system memory, and local storage. A serverless platform can meter those resources separately. Neither format tells you the final cost without its billing rules: startup, idle running time, persistent storage, data transfer, and support may matter.

Confirm the H100 variant

H100 PCIe, SXM, and NVL are not interchangeable configurations. NVIDIA's H100 specifications list 80 GB of GPU memory for SXM and 94 GB per GPU for NVL, with different power and interconnect specifications. A provider's 80 GB PCIe listing is another configuration to identify explicitly. Ask whether an offer supplies a complete GPU, a partition, or a multi-GPU system.

Memory across several GPUs is not automatically one pool that every application can use. Model fit also depends on precision, activations, optimizer states, sequence length, and the software's parallelization. Our guide to AI accelerators explains the broader hardware choices. Establish those requirements before comparing hourly prices.

H100 cloud pricing examples

The prices listed below were checked on October 19, 2026 and are in USD. They illustrate different configurations and service models. They do not establish equivalent performance, availability, or a cheapest-provider ranking. Confirm taxes, region, account terms, and the launch price before committing.

Provider and configuration	Published rate	Billing unit	What to check
Lambda: 1× H100 PCIe, 80 GB	$3.29	Per GPU-hour; one-GPU instance	Bundled host resources, storage, tax, capacity
Lambda: 1× H100 SXM, 80 GB	$4.29	Per GPU-hour; one-GPU instance	Exact configuration and availability
Lambda: 8× H100 SXM, 80 GB each	$3.99 per GPU; $31.92 for eight	Per hour; full instance has eight GPUs	Budget for the whole instance
Runpod Pods: H100 PCIe, 80 GB	$2.89	Listed hourly rate per GPU	Deployment quote, region, host resources, storage, taxes
Runpod Pods: H100 SXM, 80 GB	$3.99	Listed hourly rate per GPU	Deployment quote and available configuration
Runpod Pods: H100 NVL, 94 GB	$3.19	Listed hourly rate per GPU	GPU count and interconnect; not a quote for a two-GPU pair
Modal: H100 SXM5 GPU tasks	$0.001097/second, about $3.95/hour	GPU component only	CPU, memory, storage, egress, plan and execution options
AWS EC2 P5	Price the selected instance and region	Instance, not a universal H100 rate	OS, purchase option, storage, transfer, capacity

Runpod's examples above are its public Pods listings, not a spot quote or a Serverless worker rate. Its Pod pricing documentation directs customers to check the deployment quote and explains that storage can remain billable on stopped Pods. Include those charges when comparing against an instance bundle.

Lambda describes minute-based billing and lists applicable taxes separately. Its eight-GPU rate should not be presented as the price of a single-GPU rental. Modal's hourly figure above is a conversion of its published per-second GPU rate, not an all-inclusive instance quote.

For AWS, the P5 specification table lists one H100 in p5.4xlarge and eight in p5.48xlarge. P5e and P5en use H200 GPUs. Check the exact instance name rather than grouping every P5-family configuration under H100 pricing.

Choose a pricing model around your usage

On-demand instances

On-demand access is useful for experiments, intermittent jobs, and measuring a workload before making a commitment. Record the minimum billable period and the event that stops compute charges. Finishing a script does not necessarily stop the instance. Also check whether storage remains billable after you stop or terminate it.

Commitments and reserved capacity

A lower hourly rate can be poor value if you pay for unused hours. Compare the total commitment with the on-demand cost of the hours you expect to use, allowing for seasonal demand and changes to your project. Discounted pricing is a contract-specific offer, not a fixed percentage that applies to every H100 deployment.

A billing discount and a capacity guarantee are separate things. The AWS Savings Plans FAQ explicitly says Savings Plans do not provide capacity reservations. For any provider, ask the sales team what capacity is held, where, for how long, and under which cancellation or interruption terms.

Spot and serverless services

Spot or other interruptible capacity can suit jobs with tested checkpoint and restart behavior. Include repeated work and recovery time in the budget; a lower advertised rate does not guarantee a lower completed-job cost.

Serverless describes an execution model, not one payment method. A service may bill GPU seconds, requests, or tokens. Modal's pricing page includes application loading, processing, and the configured container idle period in billable time. Region selection and non-preemptible execution can change its base rates. Check those options before comparing its GPU figure with a dedicated instance.

Region, networking, and restart availability

Compare the same configuration in regions your project can use. A lower GPU rate may be outweighed by data movement, a different CPU or memory bundle, or an interconnect that does not suit a multi-GPU job. Record where the data and compute will be placed, and check the complete regional quote. There is no reliable fixed regional discount to apply to every provider.

Also distinguish capacity available now from capacity available after a stop. Runpod's restart guidance explains that stopping a Pod releases its GPU slot, which another customer may occupy. Test your recovery procedure and account for storage access, migration time, and a fallback configuration before depending on a stop-and-restart schedule.

Calculate the complete H100 workload cost

For a rate expressed per GPU-hour, the starting calculation is:

Compute cost = GPU count × billable hours × rate per GPU-hour.

For example, eight H100 SXM GPUs at the table's $3.99 per-GPU rate for ten billable hours cost $319.20 before applicable extras and taxes. That is arithmetic using a dated listed rate, not a benchmark or a promised training budget. It says nothing about which model will finish in ten hours.

Add charges that are not already included. A practical estimate should cover:

• Host resources: separately metered CPU and RAM, if applicable.
• Storage: datasets, checkpoints, outputs, snapshots, and retention between runs.
• Networking: data transfer, public addresses, and any private networking services.
• Software and operations: licenses, orchestration, monitoring, support, and staff time.
• Commercial terms: taxes, currency conversion, minimum spend, and unused commitments.

Keep a dated estimate with the provider, region, instance name, GPU count, and assumptions. Set a budget alert, then confirm whether it only notifies you or actually limits spending. Review the first invoice against measured usage before scaling the deployment.

A simple commitment break-even example

Assume two otherwise identical offers: $4 per hour on demand, or $2.80 per hour with a commitment covering all 720 hours in a 30-day month. These are illustrative inputs, not provider quotes. The commitment costs $2,016. Dividing that by $4 gives 504 hours, or 70% of the month.

Under those assumptions, on demand costs less below 504 billable hours; the commitment costs less above that point. Change the rates, required capacity, payment terms, or utilization and the answer changes. An observed 70% utilization figure is not a universal recommendation to reserve H100s.

Buying an H100: request a complete system quote

There is no single purchase price that describes every H100 card, module, or server. Request a dated quote for the exact part number, quantity, condition, warranty, support, taxes, and delivery terms. Confirm whether the quote covers an accelerator alone or a working server. An SXM module requires a compatible platform; it is not a standard desktop graphics-card upgrade.

Compare ownership over a defined period. Include the host system, networking, storage, installation, power, cooling, maintenance, downtime, and staff time, then account for any defensible resale value. Divide by realistic useful capacity, not continuous theoretical operation. Our GPU computing guide provides context for matching the hardware to the task.

Renting can make sense when demand is uncertain or hardware requirements are changing. Buying can merit investigation when sustained workloads, facilities, and operational capacity support it. Neither choice has a universal eight-month payback or fixed savings percentage.

If you are evaluating a complete eight-GPU appliance, our NVIDIA DGX H100 guide separates system specifications and operating costs from the price of an individual accelerator.

Compare cost per useful result

For training, measure the cost of reaching an agreed quality target. For inference, compare cost at the required latency, output quality, and request mix. Keep the model, precision, software versions, and measurement method consistent. A faster accelerator can cost less per completed job, but its product name alone does not establish that result.

If a smaller GPU meets the workload's memory and feature requirements, include it in the comparison. Compute with Hivenet provides GPU and CPU instances for running your own stack. Evaluate a currently available configuration on the same workload; do not assume an RTX instance is equivalent to H100. You manage the operating system, frameworks, dependencies, and application workflow.

Questions to settle before renting

Are free credits a reliable pricing comparison?

Use credits to test a service, but compare ongoing usage at its normal applicable rate. Check eligibility, expiration, excluded services, and what happens when the balance runs out. A short promotion should not set the long-term budget.

Does managed Kubernetes come with H100 access?

Only if the selected service includes it. A GPU instance, an orchestrated cluster, and a managed inference endpoint have different responsibilities and charges. Ask what the provider operates and what your team must install, monitor, and maintain.

What should a provider confirm in writing?

Ask for the exact GPU configuration, billable unit, included resources, extra charges, accepted payment methods, billing cycle, capacity conditions, and support scope. For a commitment, include its full cost and exit terms. Keep that quote alongside a representative workload test so the purchase decision rests on both price and usable performance.