Same H100 chip rents for $2.69 or $12.29 per hour across clouds and neoclouds

A GPU pricing analysis published in the data-receipts repository on GitHub and featured on the Hacker News front page on November 9 normalises on-demand rental rates across AWS, Azure and the smaller "neocloud" providers, expressing every price in dollars per GPU-hour. Its headline result: the identical H100 SXM 80GB chip rents for $2.69 an hour in one place and $12.29 in another, a 4.57x spread between the cheapest and most expensive option for the same silicon.

The price table

For the H100 SXM 80GB group, the analysis reports:

Provider and SKU	$ per GPU-hour	A month at 730 hours
RunPod Community	$2.69	$1,964
RunPod Secure	$3.99	$2,913
AWS p5.48xlarge	$6.88	$5,022
Azure ND96isr_H100_v5	$12.29	$8,972

Run continuously for a month, a single GPU therefore costs either $1,964 or $8,972 depending solely on where it is rented, according to the data-receipts figures.

Only identical chips are compared

The analysis groups strictly by GPU model and memory size, because an H100 is really three products: the SXM 80GB, the PCIe 80GB and the NVL 94GB variants, which are all marketed as H100s but do not cost the same. Pooling them into one chart, the author argues, would fabricate a price difference that does not exist.

The same discipline applies inside a single vendor. Azure lists the same A100 PCIe 80GB at one, two and four GPUs per virtual machine for $3.67, $4.78 and $4.78 per GPU-hour respectively, so the comparison takes each vendor's cheapest route to a given GPU and never counts an internal spread as a market gap. RunPod's Secure and Community tiers are treated as one company.

What the 4.57x gap is not

According to the analysis, the spread is not pure margin. Renting from RunPod Community instead of Azure's ND series means giving up several things:

• Azure's ND SKU carries 400 Gbps of Quantum-2 InfiniBand per GPU with GPU Direct RDMA, the interconnect that makes multi-node training practical, and that is why it is priced above Azure's own single-H100 NC SKU at $6.98.
• RunPod Community consists of third-party hosts renting out their own hardware, without a datacentre SLA.
• Enterprise compliance, committed capacity and the rest of the hyperscaler service surface are absent from the cheap rows.

The practical trade-off comes down to workload shape: someone fine-tuning a model on one node is paying for interconnect they will never use, while someone training across 64 nodes cannot consider the cheap row at all.

Method and limitations

The underlying scripts pull the AWS Price List and Azure Retail Prices APIs, which require no authentication, while neocloud prices were transcribed by hand with a source URL and the date they were read. The author lists several caveats:

• Only on-demand list prices are compared. Reserved, committed-use and spot pricing would all reduce the hyperscaler figures, and spot in particular would narrow the gap considerably.
• Each hyperscaler is measured in a single region: AWS us-east-1 and Azure eastus.
• GPU counts come from vendor documentation rather than price labels, because AWS reports only an aggregate "640 GB HBM3" figure for a p5 instance and Azure reports an accelerator-memory total.
• Lambda is excluded from the headline spread because its published rates take effect on November 16, 2026, and its current rates sit behind a login; its rows are shown separately and labelled.
• Availability is not priced, so a cheap H100 that cannot actually be reserved does not show up as a cost.

Why it matters

Anyone budgeting GPU compute faces a 4.5x difference on nominally identical silicon, and this dataset shows both where the floor sits and what the premium rows actually buy. The cheap neocloud options are only real options for single-node workloads that tolerate community-grade hosting, while multi-node training effectively requires the InfiniBand-equipped SKUs that cost multiples more. The analysis is also a reminder to check any GPU price comparison for SKU hygiene: charts that mix SXM, PCIe and NVL variants, ignore spot pricing or omit availability can materially misstate the true cost of a run.