NVIDIA B200 vs H200
Specs below are from NVIDIA's own spec sheets; rental prices are from each cloud's own price list. 7 of the GPU clouds tracked here price at least one of the two, and 5 price both.
Specifications
Each card's flagship server variant as NVIDIA names it, figures as NVIDIA prints them. B200: Per-system figures for an eight-GPU NVIDIA HGX B200 board, from the specifications table on NVIDIA's HGX platform page, which prints no per-GPU figures except NVLink GPU-to-GPU bandwidth; where NVIDIA prints one figure for two precisions (FP8/FP6, FP16/BF16, FP64/FP64 Tensor Core) it is recorded under each. Figures marked ¹ are shown sparse | dense and figures marked ² are sparse, with dense half the figure shown, as NVIDIA's footnotes state. H200: Figures from the Specifications table on NVIDIA's H200 product page, which marks both columns as preliminary specifications that may be subject to change. Figures marked ² are with sparsity, as NVIDIA's footnote states.
Not like for like: NVIDIA publishes B200 figures only for the whole HGX B200 (8 GPUs) board, and H200 figures only per GPU. The HGX B200 (8 GPUs) column is the whole board; this page does not divide it into a per-GPU figure NVIDIA does not print. The rental prices further down are per GPU or per node as each cloud states, and are the like-for-like comparison here.
| Spec | HGX B200 (8 GPUs) | H200 SXM |
|---|---|---|
| GPU memory | 1.4 TB | 141GB |
| Memory bandwidth | Not stated | 4.8TB/s |
| FP4 Tensor Core | 144 PFLOPS | 72 PFLOPS¹ | Not stated |
| FP6 Tensor Core | 72 PFLOPS² | Not stated |
| FP8 Tensor Core | 72 PFLOPS² | 3,958 TFLOPS² |
| INT8 Tensor Core | 72 POPS² | 3,958 TFLOPS² |
| FP16 Tensor Core | 36 PFLOPS² | 1,979 TFLOPS² |
| BF16 Tensor Core | 36 PFLOPS² | 1,979 TFLOPS² |
| TF32 Tensor Core | 18 PFLOPS² | 989 TFLOPS² |
| FP32 | 600 TFLOPS | 67 TFLOPS |
| FP64 Tensor Core | 296 TFLOPS | 67 TFLOPS |
| FP64 | 296 TFLOPS | 34 TFLOPS |
| Max thermal design power (TDP) | Not stated | Up to 700W (configurable) |
| Interconnect | NVIDIA NVLink Fifth generation; NVLink GPU-to-GPU Bandwidth 1.8 TB/s; Total NVLink Bandwidth 14.4 TB/s | NVIDIA NVLink™: 900GB/s PCIe Gen5: 128GB/s |
| Form factor | 8x NVIDIA Blackwell SXM | SXM |
| Multi-Instance GPU (MIG) | Not stated | Up to 7 MIGs @18GB each |
What renting each costs
Hourly on-demand prices from each cloud's own price list, alphabetical by cloud and not ranked: some price per GPU and others a whole multi-GPU node, and the unit is on every figure. Clouds that rent only one of the two show a dash for the other.
| Cloud | B200 | H200 |
|---|---|---|
| CoreWeave | NVIDIA HGX B200: $68.80 for 8 GPUs 27 Sept | NVIDIA HGX H200: $50.44 for 8 GPUs 27 Sept |
| Crusoe Cloud | — | NVIDIA H200 141GB HGX: $4.29 per GPU 27 Sept |
| Hyperstack | NVIDIA B200: $6.00 per GPU 27 Sept | NVIDIA H200 SXM: $3.99 per GPU 27 Sept |
| Lambda | NVIDIA B200 SXM6: $6.99 per GPU 27 Sept | — |
| Nebius AI Cloud | NVIDIA HGX B200: $7.15 per GPU 27 Sept | NVIDIA HGX H200: $4.50 per GPU 27 Sept |
| RunPod | B200: $6.79 per GPU 27 Sept | H200: $4.59 per GPU 27 Sept |
| Vast.ai | B200: $6.25 per GPU 27 Sept | H200: $2.63 per GPU 27 Sept |