The cheapest H100 NVL on the table right now is $0.400 per GPU-hour at Vast.ai (spot), out of 14 offers from 5 providers. The H100 NVL carries 94 GB of VRAM on NVIDIA's Hopper generation. That is peer/marketplace capacity — rented from other users' machines, and not directly comparable to a dedicated instance.
On-demand pricing for the same card spans $2.40 to $6.98 per GPU-hour — a 2.9× spread between the cheapest and the most expensive provider for identical silicon. That gap is the entire reason this table exists.
Cheapest by pricing model: on-demand $2.40 at Vast.ai · spot $0.400 at Vast.ai.
| Provider | Tier | $/GPU-hr | Pricing | Node | Commit | In stock | |
|---|---|---|---|---|---|---|---|
| Vast.ai | marketplace | $0.400 | spot | 1× GPU | — | yes | rent |
| Azure | dedicated | $1.29 | spot | 1, 2× GPU | — | — | rent |
| Vast.ai | marketplace | $2.40 | on-demand | 1× GPU | — | yes | rent |
| Runpod Community | marketplace | $2.59 | on-demand | 1× GPU | — | yes | rent |
| Runpod Community | marketplace | $2.59 | spot | 1× GPU | — | yes | rent |
| Massed Compute | dedicated | $3.11 | on-demand | 1–8× GPU | — | — | rent |
| Runpod | dedicated | $3.19 | on-demand | 1× GPU | — | yes | rent |
| Runpod | dedicated | $3.19 | spot | 1× GPU | — | yes | rent |
| Azure | dedicated | $6.98 | on-demand | 1, 2× GPU | — | — | rent |
How much does an H100 NVL cost per hour?
As of the latest scrape, $0.400 per GPU-hour at Vast.ai is the cheapest H100 NVL offer across 5 tracked cloud providers. The cheapest on-demand rate is $2.40 per GPU-hour at Vast.ai, and the most expensive on-demand rate is $6.98 at Azure.
Which cloud is cheapest for the H100 NVL?
Vast.ai at $0.400 per GPU-hour on spot capacity. Prices move constantly, so this page is regenerated from the live scrape every 15 minutes. Note that marketplace tiers are peer hardware and are not directly comparable to dedicated capacity.
Is spot H100 NVL capacity cheaper?
Yes — the cheapest spot H100 NVL is $0.400 per GPU-hour at Vast.ai, -83% versus the cheapest on-demand rate of $2.40. Spot instances can be reclaimed by the provider at any time, so they suit checkpointed training and batch inference rather than long-lived services.
How many GB of VRAM does the H100 NVL have?
94 GB, on the Hopper architecture. Multi-GPU instances multiply that: an 8× H100 NVL node exposes 752 GB of GPU memory in total.