3 providers · per GPU-hour
NVIDIA B300 cloud pricing
3 offers from 3 providers, 3 with a published on-demand price. Per GPU-hour, US$. Prices checked 16 Sept 2026.
- Lowest on-demand
- $7.40/GPU-hr
- Lowest spot
- $4.30/GPU-hr
- Memory
- 270 GB HBM3e
Spec sheet: B300 (Blackwell Ultra, HGX B300) ↗
| # | Provider | Offer | On-demand | Spot | Committed | As published (original currency) · source | Actions |
|---|---|---|---|---|---|---|---|
| 1 | Hyperstack | B300 x1 1× B300 | $7.40/GPU-hr | — | — | ||
| 2 | Verda (formerly DataCrunch) | B300 x1 1× B300 | $7.65/GPU-hr | — | — | ||
| 3 | Nebius | B300 x1 1× B300 | $7.85/GPU-hr | $4.30/GPU-hr | — |
What is the B300 used for?
The B300 is a data-centre GPU positioned by its maker for large-scale model training, built on the Blackwell (Blackwell Ultra) architecture and released in 2025. Multi-GPU servers link cards over NVLink 5th gen (1.8 TB/s per GPU), which matters when a job is split across several GPUs.
How large a model fits on one B300?
With 270 GB of memory, one card holds the weights of a model up to about 135 billion parameters at 16-bit precision, or about 270 billion at 8-bit. That is weights only: inference also needs room for the KV cache, and training needs several times more for gradients and optimiser state, so larger jobs are split across multiple GPUs.
| Spec (as published) | B300 (Blackwell Ultra, HGX B300) |
|---|---|
| Architecture | Blackwell (Blackwell Ultra) |
| Form factor | SXM (HGX/DGX B300) |
| Memory | 270 GB HBM3e |
| Memory bandwidth | 7,700 GB/s |
| Interconnect | NVLink 5th gen (1.8 TB/s per GPU) |
| Max power (TDP) | 1100 W |
| Precisions listed | FP4, FP8, FP6 |
| Partitioning | Up to 7 MIG (datasheet) |
| Released | 2025 |
| Source | maker spec sheet ↗ |
Need B300 capacity for a longer run?
Committed and cluster pricing is often negotiated. Tell us the GPU count, duration and region.