The NVIDIA H100 has never had a single price, but this week someone finally measured the gap. A dataset built straight from the AWS Price List and the Azure Retail Prices API puts the same H100 SXM 80GB at $2.69 per GPU-hour on RunPod Community and $12.29 on Azure's ND96isr_H100_v5. That is 4.57x, or $7,008 per GPU per month at 730 hours of continuous runtime.
The comparison holds the model fixed. Same silicon, same 80GB of HBM3, same on-demand billing model, four published offers:
- RunPod Community: $2.69 per GPU-hour, $1,964 for a 730-hour month
- RunPod Secure: $3.99 per GPU-hour, $2,913 per month
- AWS p5.48xlarge: $6.88 per GPU-hour, $5,022 per month
- Azure ND96isr_H100_v5: $12.29 per GPU-hour, $8,972 per month
The narrowest gap is the one people argue about least. RunPod Community to RunPod Secure is a 1.48x step from third-party hosts to company-operated capacity, and it is the only jump on the list that costs nothing in compute terms.
AWS to Azure, both hyperscalers, both eight-GPU nodes, is a 1.79x spread. That difference is not silicon and it is not network fabric. It is SKU packaging, region, and whatever each provider thinks the market will bear for a node that arrives with enterprise support attached.
The important detail in the methodology is the grouping. An H100 is not one product. The SXM 80GB, the PCIe 80GB and the NVL 94GB all get sold as H100s, and pooling them in a single chart manufactures a spread that does not exist. Every comparison in the dataset therefore stays inside a single model-and-memory group, and every vendor is represented by its cheapest route to that specific part.
That rule matters more than it sounds. Within Azure alone, the same A100 PCIe 80GB is listed at $3.67, $4.78 and $4.78 per GPU-hour depending on whether the VM carries one, two or four accelerators. Compare a four-GPU VM against someone else's single-GPU offer and you have invented a market gap out of pricing-page layout.
What the 4.57x actually buys
The instinct is to read the spread as margin. It is partly that, but the expensive rows are paying for hardware the cheap rows do not ship. Going from an Azure ND-series node to a community H100 gives up:
- InfiniBand: 400 Gbps per GPU of Quantum-2 with GPUDirect RDMA, which is what makes multi-node training scale at all. It is also the reason the ND SKU prices above Azure's own single-H100 NC SKU at $6.98 per GPU-hour.
- A datacenter SLA: community listings are third-party hosts renting out their own machines, with availability and uptime guarantees to match.
- The compliance surface: enterprise agreements, committed capacity, audit paperwork, and support channels that answer at 3am.
So the buying decision is not "which H100 is cheaper". It is "which H100 matches my parallelism". A single-node LoRA fine-tune on one or two GPUs never touches an RDMA fabric, and paying for 400 Gbps of interconnect it cannot use is pure waste. A 64-node pre-training run cannot use the $2.69 row at all, because the cheap capacity is not wired for it.
The dataset is also honest about where its own number stops being useful:
- On-demand list prices only. Reserved, committed-use and spot pricing all cut the hyperscaler figures, and spot in particular would compress the gap sharply.
- One region per provider, us-east-1 for AWS and eastus for Azure, so regional price variation is invisible.
- Lambda is excluded from the headline spread because its listed rates take effect on 2026-11-16 and current rates sit behind a login.
- Availability is not priced. A cheap H100 that is not allocatable is not a price.
Read that way, the 4.57x is less a scandal than a map. The H-series market has stratified into three tiers: community metal at roughly $2.70, managed neocloud at roughly $4, and hyperscaler nodes at $7 to $12. The chip is identical in all three. What changes is the wiring around it, and the wiring is what most training bills are really buying.
For anyone budgeting H100 capacity in late 2026, the practical takeaway is to price the interconnect separately from the FLOPs. The FLOPs are a commodity at $2.69. The fabric, the SLA and the ability to allocate 512 GPUs next Tuesday are not.
Comments