RentServersBy the RentServers team at Wells Software Consulting LLC · Updated
B200 specifications
| Architecture | Blackwell |
|---|---|
| Memory | 180 GB HBM3e |
| Memory bandwidth | Up to ~8 TB/s |
| FP8 Tensor (dense) | ~4.5 PFLOPS |
| FP4 Tensor (dense) | ~9 PFLOPS |
| NVLink | 1.8 TB/s (5th gen) |
| Max power | ~1,000 W |
Figures for the SXM/HGX version unless noted, from NVIDIA. Dense Tensor Core throughput; NVIDIA's headline numbers with sparsity are double.
Who the B200 is for
Teams training large models where wall-clock time matters, and inference fleets that can use FP4 or FP8 kernels. Software support has matured through 2025–26, but confirm your framework version supports Blackwell before committing to a long reservation.
Pricing and supply
The spread is wide. GetDeploying shows in-stock on-demand from $3.75, a reserved median of $5.62 and an on-demand median of $7.99, with hyperscaler list prices above $14. Lead times on large clusters are longer than for Hopper, so start sourcing early.
Per unit of FP8 compute, the B200 can work out cheaper than the H100 even at a higher hourly rate. Compare cost per training step on your own model rather than cost per hour.
What to check in a B200 quote
- Cooling and rack density. Some B200 systems are liquid-cooled, which limits which data centers can host them.
- InfiniBand or RoCE fabric bandwidth per GPU for multi-node jobs.
- Delivery date in the contract, and what happens if it slips.
- Driver and CUDA versions on the provided image.
B200 questions
How much does it cost to rent a B200?
On GetDeploying's tracker (October 16, 2026): median $5.62 per GPU-hour reserved and $7.99 on demand, with in-stock on-demand offers from $3.75.
B200 or H200 for inference?
If your serving stack supports FP4 or Blackwell-optimized FP8 kernels, the B200 usually delivers more tokens per dollar. If you're on an older stack, the H200 is the safer choice.