1 providers · per GPU-hour
NVIDIA GB300 cloud pricing
- Lowest on-demand
- $8.80/GPU-hr
- Lowest spot
- $4.40/GPU-hr
- Memory
- 279 GB HBM3e
| # | Provider | Offer | On-demand | Spot | Committed | As published (original currency) · source | Actions |
|---|---|---|---|---|---|---|---|
| 1 | Verda (formerly DataCrunch) | GB300 x1 1× GB300 Grace Blackwell Ultra Superchip · 32 vCPU · 225 GB RAM | $8.80/GPU-hr | $4.40/GPU-hr | — |
What is the GB300 used for?
The GB300 is a data-centre GPU positioned by its maker for large-scale model training, built on the Blackwell (Blackwell Ultra) + Grace architecture and released in 2025. Multi-GPU servers link cards over NVLink 5th gen (1.8 TB/s per GPU; 130 TB/s NVL72), which matters when a job is split across several GPUs.
How large a model fits on one GB300?
With 279 GB of memory, one card holds the weights of a model up to about 139 billion parameters at 16-bit precision, or about 279 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) | GB300 Grace Blackwell Ultra Superchip (NVL72) |
|---|---|
| Architecture | Blackwell (Blackwell Ultra) + Grace |
| Form factor | superchip, NVL72 rack (liquid-cooled) |
| Memory | 279 GB HBM3e |
| Memory bandwidth | 8,000 GB/s |
| Interconnect | NVLink 5th gen (1.8 TB/s per GPU; 130 TB/s NVL72) |
| Max power (TDP) | 1400 W |
| Precisions listed | FP4, FP8, FP6, FP16, BF16, TF32, FP32, FP64 |
| Released | 2025 |
| Source | maker spec sheet ↗ |
Need GB300 capacity for a longer run?
Committed and cluster pricing is often negotiated. Tell us the GPU count, duration and region.