GPU comparison
B200 vs H200
B200 is a generation ahead: about 2.3× the dense FP8 of an H200, 180 vs 141 GB and 8 vs 4.8 TB/s. H200 costs much less per month and runs every Hopper recipe unchanged. Pick B200 for training throughput, H200 for cost-efficient fine-tuning and inference.
- Updated Nov 4, 2026
- Specs: vendor datasheets, dense Tensor figures
- Prices: live CoreGPU catalog
| Spec | NVIDIA HGX B200 | NVIDIA H200 SXM |
|---|---|---|
| Architecture | Blackwell | Hopper |
| GPU memory | 180 GB HBM3e | 141 GB HBM3e |
| Memory bandwidth | 8 TB/s | 4.8 TB/s |
| FP8 dense | 4.5 PF | 1.98 PF |
| FP4 dense | 9 PF | n/a |
| Interconnect | NVLink 5 · 1.8 TB/s | NVLink 4 · 900 GB/s |
| Power (TDP) | 1,000 W | 700 W |
| Price per GPU / month | $4,449 | $2,699 |
| Price per GB of memory | $24.72 | $19.14 |
| Price per FP8 PFLOPS | $989 | $1,363 |
| Availability | Back Nov 20 | 34 available |
Green marks the better value on each line. Prices are per GPU per month at CoreGPU, at least 25% below the market median.
- Pre-training and large fine-tunes on deadline
- FP4 inference (H200 has no FP4)
- Highest throughput per server
$4,449 /GPU/moJoin waitlist B200
- The best value for fine-tuning and FP8 inference
- Mature Hopper software and recipes
- Lower monthly budget
$2,699 /GPU/moConfigure H200