Azure ND H100 v5 series (Standard_ND96isr_H100_v5, 8x H100, 96 vCPUs, 1.9 TB RAM) is the most expensive at $98.32/hr on-demand ($12.29/GPU-hr normalized) — consistent across July 2026 pricing surveys (vantage.sh, spheron.network). 1-year reservation brings it to ~$63.40/hr ($7.93/GPU-hr); 3-year reservation to ~$43.79/hr ($5.47/GPU-hr). Azure's quota approval process for 8+ node requests can stretch to 3-4 weeks before capacity is confirmed — a structural disadvantage for time-sensitive training jobs.
Vast.ai and the marketplace floor
Vast.ai's marketplace delivers the lowest published rates in the category at the cost of interruptibility and host variance.
Verified H100 PCIe prices on Vast.ai: unverified community hosts at $0.90-$1.60/GPU-hr; datacenter-verified hosts at $1.50-$2.27/GPU-hr (spheron.network, July 2026). The Vast.ai 'on-demand' label does not mean what it means on AWS or GCP — it means the host is currently offering the machine at the listed price and will not actively evict you, but if the host takes the hardware offline for any reason, the machine stops. There is no platform-level guarantee of availability. For fault-tolerant training with checkpoint discipline, Vast.ai interruptible at $1.49/GPU-hr is the cheapest credible option in 2026.
Three rules for Vast.ai: (1) sort by DLPerf score for reliability; (2) only use datacenter-verified hosts for production jobs; (3) always checkpoint every 30 minutes or less because interruptible means reclaimable without notice. The platform does charge for storage while the instance is paused, which adds 5-10% to total cost for low-utilization workflows.
Reserved-capacity economics
For predictable training workloads, reserved capacity delivers 30-60% off list — but the headline rates are misleading.
Together AI's reserved cluster (91-180 day) starts at $3.09/GPU-hr for H100 SXM. Lambda 1-Click Cluster 16-GPU reservation drops to $6.16/GPU-hr on 2-week-1-year terms, $5.54/GPU-hr at 256-GPU scale. CoreWeave reserved capacity ceiling is 'up to 60% off on-demand' — target $2.46/GPU-hr reserved for H100 — but only achievable through sales negotiation. Hyperscaler 1-year reservations bring H100 to $7-$8/GPU-hr; 3-year reservations to $5-$6/GPU-hr (nOps FinOps, August 2026).
The break-even calculation: for a job that runs continuously for 6+ months, reserved beats on-demand even with the deposit. For a workload that might run for 2 weeks and then stop, on-demand wins because reserved capacity is wasted if unused. The reservation deposit on hyperscalers is typically 30-50% of the 1-year value, refundable only if you commit to a longer term. Most mid-market teams should avoid 3-year reservations unless they have locked-in demand.
Side-by-side H100 pricing comparison
| Provider | SKU / Shape | $/GPU-hr | Billing | Notes |
|---|
| Vast.ai (verified DC) | H100 PCIe interruptible | $1.49-$2.27 | Per-hour | Cheapest; host-dependent reliability |
| RunPod Community | H100 PCIe pod | $1.99 | Per-second | Third-party hosts |
| RunPod Community | H100 SXM pod | $2.69 | Per-second | Third-party hosts |
| RunPod Secure | H100 SXM pod | $3.29 | Per-second | RunPod datacenters |
| Lambda | 1x H100 PCIe VM | $3.29 | Per-hour (1-min increments) | Transparent card |
| Together AI reserved | H100 SXM 91-180 day | $3.09 | Reserved cluster | Time-bounded reservation |
| Lambda | H100 SXM 1x VM | $3.99 | Per-hour | NVLink interconnect |
| CoreWeave spot | HGX H100 8x node | $2.46 | Spot interruptible | $19.71/hr node |
| CoreWeave on-demand | HGX H100 8x node | $6.16 | On-demand | 8-GPU minimum |
| AWS Capacity Blocks | p5.4xlarge single H100 | $5.191 | Reserved window | Future-dated reservation |
| AWS on-demand | p5.48xlarge 8x H100 | $6.88 | On-demand | $55.04/hr node |
| GCP on-demand | a3-highgpu-8g H100 | $11.06 | On-demand | $88.49/hr node |
| GCP 3-yr CUD | a3-highgpu-8g H100 | $4.86 | 3-year reserved | 55% off list |
| Azure ND H100 v5 | 8x H100 v5 | $12.29 | On-demand | Most expensive tier-1 |
| Azure 3-yr reserved | 8x H100 v5 | $5.47 | 3-year reserved | Best discount |
How to choose the right H100 provider in 2026
The decision hinges on three variables: workload duration, reliability tolerance, and budget.
- For a 1-day fine-tuning experiment: Vast.ai verified datacenter at $1.49/GPU-hr. Accept interruptibility; checkpoint every 30 minutes.
- For a 1-2 week iterative fine-tune: RunPod Community Cloud H100 SXM at $2.69/GPU-hr or Lambda 1x H100 PCIe at $3.29/GPU-hr. Per-second billing lets you stop and restart cheaply.
- For a 1-3 month training run you can't afford to lose: Lambda Reserved Cloud 16-GPU cluster at $6.16/GPU-hr, or Together AI reserved at $3.09/GPU-hr with commitment.
- For a 200+ GPU distributed training job: CoreWeave HGX H100 with reserved negotiation. The 8-GPU node minimum and InfiniBand interconnect pay for themselves at scale; naive scaling on a cheaper cluster runs at 70% efficiency and costs more per useful FLOP.
- For production inference with compliance requirements: AWS Capacity Blocks at $5.191/GPU-hr, or AWS 3-year reserved at $5-$6/GPU-hr. The integration with SageMaker and enterprise support justify the premium.
What enterprise buyers should do next
Three actions for organizations evaluating H100 capacity in 2026.
- Run a 2-week pilot across 2-3 providers on the same workload. Same training script, same dataset, same hyperparameters. Compare cost, throughput, and reliability. The cheapest published rate is rarely the cheapest end-to-end cost once you factor in support responsiveness and reclaim frequency.
- Negotiate a reserved contract before committing to a single provider. CoreWeave, Lambda, and Together AI all negotiate reserved rates meaningfully off list. Even a 1-year commitment typically unlocks 25-40% off list. The published rate card is a starting point, not the final price.
- Model egress and storage costs separately. Hyperscalers charge $0.087-$0.12/GB egress; Lambda, CoreWeave, and Crusoe publish free egress. For a 10 TB dataset with monthly retraining, egress alone adds $870-$1,200/month on AWS — more than the savings from a lower H100 hourly rate.
What to watch next
Three near-term datapoints. First, the CoreWeave Q3 2026 earnings call — CoreWeave CFO stated the company is 'largely sold out' of 2026 capacity and raising list prices; expect rate-card increases by Q4 2026. Second, AWS's planned p6-b200.48xlarge at $98.84/hr effective — Blackwell-generation capacity will relieve H100 demand pressure and likely pull H100 list rates down 10-15% by mid-2027. Third, NVIDIA Vera Rubin full production at CoreWeave, Google, Azure, OCI, and Nebius — Rubin GPUs hit $40B/gigawatt revenue per NVIDIA's earnings, and as Rubin capacity comes online, H100 spot prices will fall faster than reserved rates (NVIDIA Q2 FY27 earnings, August 26, 2026).