RentServersBy the RentServers team at Wells Software Consulting LLC · Updated
| H100 SXM | H200 | B200 | |
|---|---|---|---|
| Architecture | Hopper | Hopper | Blackwell |
| Memory | 80 GB HBM3 | 141 GB HBM3e | 180 GB HBM3e |
| Bandwidth | 3.35 TB/s | 4.8 TB/s | up to ~8 TB/s |
| FP8 dense | ~1.98 PFLOPS | ~1.98 PFLOPS | ~4.5 PFLOPS |
| FP4 | No | No | Yes |
| Median reserved | $3.34 | $3.84 | $5.62 |
| Median on-demand | $3.74 | $4.94 | $7.99 |
Specs from NVIDIA datasheets (H100, H200, B200). Prices per GPU-hour from GetDeploying GPU price tracker, October 16, 2026.
Price per unit of compute
At reserved medians, the B200 costs about 1.7× as much per hour as the H100 but offers roughly 2.3× the dense FP8 throughput. If your training code runs efficiently on Blackwell, that makes the B200 cheaper per unit of work. If it doesn't, you're paying a premium for performance you won't get.
When the H200 is the right answer
The H200 has the same compute as the H100, so it only pays off when memory is the constraint. The common case is serving a 70B-class model: in FP8 the weights take about 70 GB, leaving almost no room for KV cache on an 80 GB H100. On one H200 there's about 70 GB left for cache, so you can serve the model on half as many GPUs.
A quick decision guide
- Pre-training or full fine-tuning, cost-sensitive: H100.
- Pre-training where time-to-result matters and your stack supports Blackwell: B200.
- Serving 70B+ models or long-context inference: H200, or B200 with FP4/FP8 serving.
- LoRA on models up to ~34B: H100, or A100 80 GB to save money.