H100 vs H200
NVIDIA H100 and H200 compared: memory, bandwidth, compute and current cloud prices.
Published Oct 9, 2026
Same generation, more memory
H100 and H200 both belong to NVIDIA's Hopper generation. According to NVIDIA, the SXM versions have the same FP8 Tensor Core performance.
The difference is memory: the H200 SXM has 141 GB at 4.8 TB/s, the H100 SXM 80 GB at 3.35 TB/s.
| Model | Memory | Memory bandwidth | FP8 Tensor | Max. power |
|---|---|---|---|---|
| H100 SXM | 80 GB | 3.35 TB/s | 3,958 TFLOPS | 700 W |
| H200 SXM | 141 GB | 4.8 TB/s | 3,958 TFLOPS | 700 W |
Manufacturer figures. Tensor Core values with sparsity, H200 values preliminary according to NVIDIA.
What the difference means
More memory lets larger models and longer contexts run on fewer GPUs. Workloads limited by memory bandwidth, such as inference of large language models, benefit most.
When a model fits into 80 GB and compute is the limit, the H100 delivers comparable throughput.
Cloud prices
The median on-demand price is USD 3.75EUR 3.35CHF 3.09 per GPU-hour for the H100 and USD 5.64EUR 5.04CHF 4.65 for the H200, as of Oct 25, 2026.
AI Compute Prices lists 404 H100 offers and 124 H200 offers.
As of Oct 25, 2026, 21:05 UTCSource: public price lists and pricing APIs of the providersConverted at ECB reference rates of Oct 23, 2026Method