GPU Cloud Pricing · Blackwell Ultra
NVIDIA's flagship Blackwell Ultra accelerator: 288 GB of HBM3e per GPU, built for trillion-parameter training and long-context inference.
GPU.ai rates are the live fleet floor — the cheapest provider with stock, passed through at cost. Competitor rates are public on-demand list prices.
| Provider | $ / GPU / hr |
|---|---|
| GPU.ai — Secure Cloud | Out of stock |
| CoreWeave | Not available |
| AWS | $17.80 |
| Azure | Not available |
GPU.ai prices are read live from our aggregated fleet of 12+ GPU clouds and normalized to a per-GPU hourly rate (a multi-GPU node's price is divided by its GPU count). We charge the provider's raw rate with zero markup and bill per second. Competitor figures are each provider's public on-demand list rate for the closest equivalent SKU — never spot, reserved, or committed-use discounts. Stock counts reflect launchable inventory, not theoretical capacity. Last updated 2026-08-17; this page refreshes hourly.
| Architecture | Blackwell Ultra |
| VRAM | 288 GB HBM3e |
| Memory bandwidth | 8,000 GB/s |
| FP16 tensor (dense) | — |
| CUDA cores | — |
| Power (TDP) | — |
| Interconnect | NVLink 5 (1.8 TB/s) |
| Released | 2025 |
Minimum B300 count needed to hold popular open models in VRAM, at FP16 and 4-bit quantization.
| Model | Params | FP16 | 4-bit |
|---|---|---|---|
| Llama 3.1 8B | 8B | 1× B300 | 1× B300 |
| Mistral 7B | 7B | 1× B300 | 1× B300 |
| Gemma 2 27B | 27B | 1× B300 | 1× B300 |
| Qwen2.5 32B | 32B | 1× B300 | 1× B300 |
| Llama 3.1 70B | 70B | 1× B300 | 1× B300 |
| Qwen2.5 72B | 72B | 1× B300 | 1× B300 |
| Mixtral 8x22B | 141B MoE | 2× B300 | 1× B300 |
| Llama 3.1 405B | 405B | 4× B300 | 1× B300 |
| FLUX.1 dev | 12B | 1× B300 | 1× B300 |
| SDXL | 3.5B | 1× B300 | 1× B300 |
| Whisper large-v3 | 1.5B | 1× B300 | 1× B300 |
Weights-only estimates with a 20% margin for KV cache and activations at moderate context. Long contexts, large batches, and training need substantially more memory.
B300 capacity is temporarily out of stock on GPU.ai. When available, pricing is the provider's raw rate with zero markup, billed per second. Check this page again — stock and prices refresh continuously.
The NVIDIA B300 (Blackwell Ultra) has 288 GB of HBM3e memory with up to 8,000 GB/s of bandwidth. GPU.ai reports the minimum VRAM actually delivered across our fleet, so the figure you see at launch is the figure you get.
GPU.ai aggregates a dozen-plus GPU clouds and passes through each provider's raw price with no markup, so the same silicon is typically far cheaper than hyperscaler list rates. The comparison table on this page shows the current public on-demand list prices side by side.
Billing is per second, starting when your instance is SSH-ready and stopping the moment you terminate. There are no minimum commitments, no reservation fees, and no egress surprises — you pay the listed hourly rate pro-rated to the second.