According to an exhaustive infrastructure pricing dataset released by open-source analytics initiative Data Receipts on November 9, 2026, the cost of renting an identical Nvidia H100 SXM 80GB accelerator on demand varies by 4.57 times depending strictly on cloud provider selection.
To ensure rigorous parity, the audit strictly isolated identical hardware specifications, discarding comparisons that conflate the SXM 80GB module with lower-tier PCIe 80GB or high-capacity NVL 94GB variants. Prices were normalized directly to dollars per GPU per hour based on publicly published retail price lists in single primary regions (AWS `us-east-1` and Azure `eastus`).
The results document an immense financial dispersion for equivalent silicon. On commodity provider RunPod Community, an H100 SXM rents for $2.69 per hour, amounting to $1,964 over a standard 730-hour operating month. RunPod Secure commands $3.99 per hour ($2,913 monthly). On tier-one hyperscalers, AWS charges $6.88 per GPU-hour for its `p5.48xlarge` instances ($5,022 monthly), while Microsoft Azure lists its `ND96isr_H100_v5` instance at $12.29 per GPU-hour—costing $8,972 per month for the exact same physical chip.
The report emphasizes that the 4.57x spread does not represent pure corporate markup. Azure's ND series integrates 400 Gbps per GPU of dedicated Quantum-2 InfiniBand networking with GPUDirect RDMA, an architectural requirement for multi-node distributed model pretraining that cheaper community providers do not offer. Azure's own non-InfiniBand NC SKU rents for significantly less at $6.98 per hour.
However, for teams executing single-node fine-tuning, evaluation passes, or exploratory prototyping, paying hyperscaler premiums forces organizations to subsidize high-bandwidth interconnects they cannot utilize. The findings highlight accelerating market segmentation between bare-metal compute providers and full-service distributed training fabrics.