Short answer
An NVIDIA H200 costs about $31,000 to $32,000 to buy as a single NVL card, and an 8-GPU H200 board costs around $308,000 to $315,000. To rent, the median on-demand price across the providers we track is about $4.92 per GPU-hour in November 2026, ranging from about $2.09 to $10.85. At providers that offer both, the H200 costs on average about 13% more per hour than the H100 but has 76% more memory. Check today's prices on the live GPU price tracker.
What is the H200?
The H200 is NVIDIA's upgraded Hopper GPU. It uses the same compute design as the H100 but replaces 80 GB of HBM3 with 141 GB of HBM3e and raises memory bandwidth to 4.8 TB/s. Because large language model inference is limited mainly by memory, the H200 runs big models noticeably faster and fits larger models or longer contexts on fewer GPUs. Our H100 vs H200 guide covers the differences in detail.
H200 rental prices by provider
On-demand prices per GPU-hour from our tracker in mid-November 2026:
| Provider | H200 price per GPU-hour | Notes |
|---|---|---|
| Beam | $2.09 | Serverless |
| Ori | $3.50 | Single GPU |
| Vast.ai | $3.85 | Marketplace median |
| Hyperbolic | $3.99 | |
| Crusoe | $4.29 | |
| DigitalOcean | $4.47 | |
| Modal | $4.54 | Serverless |
| RunPod | $5.29 | |
| Nebius | $5.40 | |
| CoreWeave | $6.31 | 8-GPU servers |
| AWS | $7.91 | p5en.48xlarge |
| Oracle Cloud | $10.00 | 8-GPU bare metal |
| Azure | $10.60 | ND H200 v5 |
| Google Cloud | $10.85 | Reservation only |
Prices move every week. The GPU cloud pricing tracker shows today's rates, the daily price history and a cost calculator.
How much does an H200 cost to buy?
| What you buy | Typical price (USD) |
|---|---|
| Single H200 NVL card | $31,000 to $32,000 |
| 4-GPU H200 SXM board | about $170,000 to $175,000 |
| 8-GPU H200 board | $308,000 to $315,000 |
| NVIDIA DGX H200 system | $400,000 to $500,000 |
Like other data center GPUs, the H200 has no official list price, and SXM modules are sold inside servers from NVIDIA's partners rather than on their own.
Is the H200 worth it over the H100?
For many AI workloads, yes:
- Memory per dollar: a small price premium buys 76% more memory (141 GB vs 80 GB).
- Fewer GPUs per model: a 70-billion-parameter model in FP16 needs about 140 GB for weights alone, which fits on one H200 but needs two H100s.
- Faster token generation: higher bandwidth means more tokens per second for memory-bound inference; see GPU memory bandwidth explained.
The H100 can still be the better deal for compute-heavy training of models that fit comfortably in 80 GB, or when a provider offers H100s at a big discount.
H200 vs other GPUs on price
| GPU | Median rental per hour | Memory | Best for |
|---|---|---|---|
| H100 | about $3.92 | 80 GB | General training and inference |
| H200 | about $4.92 | 141 GB | Large-model inference, long context |
| B200 | about $7.34 | 180 to 192 GB | Highest throughput, FP4 and FP8 |
See also the H100 price guide, the B200 price guide and cheapest cloud GPU providers.
Frequently asked questions
How much does an H200 cost per hour?
About $4.92 per GPU-hour at the median in November 2026, with on-demand listings from around $2 to almost $11 depending on the provider.
How much does an H200 cost per month?
At $4.92 per hour, one H200 running 24/7 costs about $3,590 per month.
Why is the H200 cheaper on some clouds than the H100?
Pricing depends on each provider's supply, contracts and demand. Some providers price the two close together; on average the H200 costs about 13% more.
Sources and further reading
- TensorCore.si, Live GPU cloud pricing tracker (updated daily)
- Thunder Compute, NVIDIA H200 price comparison (November 2026)
- IntuitionLabs, NVIDIA AI GPU pricing guide
- NVIDIA, H200 Tensor Core GPU




Leave a Reply