# H100 rental prices span 4.57×, even after matching the chip

**Summary:** Data Receipts finds a 4.57× H100 SXM rental-price spread. Compare full instance bills, networking, and availability before treating a lower GPU-hour rate as a saving.

- Canonical: https://markhuang.ai/news/h100-rental-price-gap
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-11-09
- Section: News
- Tags: NVIDIA H100, GPU Pricing, Cloud Computing, AI Infrastructure, Runpod, AWS, Azure
- Source: [Data Receipts on GitHub](https://github.com/datareceipts/data-receipts/tree/main/109-gpu-pricing)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![Matching GPU server modules arranged as a standalone machine and a cluster joined by blue networking cables](https://cdn.markhuang.ai/news/h100-rental-price-gap/hero.webp)

*Conceptual illustration: matching the accelerator is the first step. The surrounding infrastructure can make two GPU rentals useful for different jobs. Generated with OpenAI imagegen.*

The hourly rental price of an NVIDIA H100 SXM, an AI accelerator with 80GB of memory, varies widely by provider. In [Data Receipts' November 9, 2026 comparison](https://github.com/datareceipts/data-receipts/tree/main/109-gpu-pricing), Runpod Community comes to $2.69 per GPU-hour; Azure's ND instance comes to $12.29. That is a 4.57× spread after matching the GPU variant and memory capacity.

For someone planning an AI experiment, the comparison makes a useful choice visible: how much infrastructure does the job need around the chip? The price gap is worth investigating, but the table cannot establish that the cheaper rental will finish the same job at a quarter of the cost.

I like that the project shows its grouping rules and underlying records. That makes it possible to check what the prices mean before building a budget around them.

## First make sure "H100" means the same hardware

The repository separates H100 SXM 80GB, H100 PCIe 80GB, and H100 NVL 94GB into different groups. Those suffixes identify different hardware variants. Treating all of them as interchangeable would mix a hardware choice into what looks like a provider price comparison.

Within each group, the analysis keeps each vendor's cheapest recorded configuration. It also counts Runpod Secure and Community as two offerings from one company. Here is its headline group:

| H100 SXM 80GB offering  | Reported USD per GPU-hour |
| ----------------------- | ------------------------- |
| Runpod Community        | $2.69                     |
| Runpod Secure           | $3.99                     |
| AWS p5.48xlarge         | $6.88                     |
| Azure ND96isr\_H100\_v5 | $12.29                    |

These are the project's [November 9 snapshot values](https://github.com/datareceipts/data-receipts/blob/main/109-gpu-pricing/compare.json), not fresh checkout quotes. The analysis covers on-demand list prices, with AWS in us-east-1 and Azure in eastus. It excludes reserved, committed-use, and spot discounts. It also leaves Lambda's future-effective rates outside the headline spread.

## A comparison unit is different from a rental unit

A GPU-hour is useful for comparison, but check the size of the machine you must rent. [AWS lists eight H100s in p5.48xlarge](https://aws.amazon.com/ec2/instance-types/p5/), and [Microsoft lists eight in ND96isr\_H100\_v5](https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/ndh100v5-series). The source divides those machine prices by eight.

Its [underlying price records](https://github.com/datareceipts/data-receipts/blob/main/109-gpu-pricing/hyperscalers.json) put the full instances at $55.04 per hour for that AWS configuration and $98.32 for that Azure configuration. Someone who needs only one GPU cannot budget for these particular instances by setting aside $6.88 or $12.29 an hour. All eight GPUs belong in the bill, even if the job leaves seven idle.

For a budget, I would compare the smallest suitable configurations I can actually rent, then estimate the total cost of finishing the job. Dividing by the GPU count helps build a shortlist. The full machine determines the bill.

## The network premium depends on where the work runs

Microsoft's ND documentation describes a dedicated 400 Gbps InfiniBand connection for each GPU, supporting communication between virtual machines. It separately describes NVLink connections between GPUs inside a machine. The distinction matters: a job spread across servers has communication needs that a job contained in one server does not.

The source uses fine-tuning on one node, meaning one server, as an example of a workload that does not need the inter-server fabric. Large distributed training is a different purchase. Renting cheaper individual GPUs does not, by itself, reproduce a connected training cluster.

There is a qualification to the source's explanation of Azure's premium. It contrasts the ND rate with a $6.98 single-H100 NC offering. But [Microsoft identifies that NC family as H100 NVL with 94GB per GPU](https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/ncadsh100v5-series), while the headline ND comparison uses 80GB SXM hardware. That contrast changes more than networking. It cannot isolate how much of the price difference pays for InfiniBand.

Service differences also exist within Runpod. Its [Pod selection documentation](https://docs.runpod.io/pods/choose-a-pod) describes Secure Cloud as data centers with high redundancy and Community Cloud as peer-to-peer providers with variable reliability. Those are relevant buying criteria, although they do not measure how a particular job will perform.

## Price the work you can complete

The source explicitly leaves availability unpriced. In the [Hacker News discussion](https://news.ycombinator.com/item?id=50427592), commenter jdc0589 praises Runpod but describes difficulty obtaining a specific GPU when needed. That is one person's report, not an availability survey. It illustrates why a listed rate needs a capacity check before it becomes a plan.

[SemiAnalysis's March 2025 ClusterMAX evaluation guide](https://semianalysis.com/2025/03/26/the-gpu-cloud-clustermax-rating-system-how-to-rent-gpus/) offers a broader checklist. It assesses networking, storage, security, reliability, and support alongside price. Those criteria are useful here, though its historical provider ratings would not establish today's service quality.

My next step would be a small, representative job on a suitable configuration. Record its full rental cost and whether it produces the required result in time. For inference, that includes the application's response-time needs, the same concern behind [my earlier look at DeepInfer's infrastructure costs](https://markhuang.ai/news/deepinfer-missing-serving-system). For training, it means reaching the intended training result rather than merely keeping GPUs busy.

The 4.57× spread is a good reason to compare providers. Use it to find candidates, then ask each candidate to quote the configuration and capacity your job needs.
