Highest VRAM in the Blackwell lineup for training the largest foundation models.
Listings for the B300 reach $18.00/hr, often reflecting a premium for high availability. However, you might be able to find available instances from as low as $6.33/hr per GPU (on-demand). Spot instances start lower, at $2.99/hr per GPU.
USA 4 configs 1x-8x
USA 6 configs 2x-8x
UAE 2 configs 1x-8x
USA 1 config 8x
Singapore 3 configs 1x-8x
USA 4 configs 8x
Germany 8 configs 8x
Germany 9 configs 1x-8x
Finland 28 configs 1x-8x
USA 1 config 8x
UAE 5 configs 8x
USA 1 config 8x
USA 1 config 1x
USA 2 configs 1x
UK 1 config 1x
Heads up: We do our best to keep these specs & prices accurate. However, cloud costs may fluctuate based on region, usage, and other factors not listed here. These are estimates based on common setups and are for informational purposes only. Always verify current rates & exact specs with the provider before provisioning.
Comes with 288GB HBM3e - the highest VRAM in the Blackwell lineup. 1800 GB/s NVLink. Can serve 200B+ parameter models on a single GPU without quantization.
Extremely limited availability and the highest cost tier. For most workloads, the B200 or H100 offers sufficient memory at lower cost.
The median on-demand price across providers has risen about 58% since November 2025, from $5.00 to $7.89/hr per GPU. This reflects the market-wide median, which moves both when providers change prices and when lower or higher priced offerings enter the market.
With 288GB of VRAM, the B300 can usually run 70B-class models with headroom, and may handle much larger models in 4-bit quantized form depending on runtime overhead, KV cache, and context length.
The B300 has 288GB of VRAM. Multi-GPU setups increase total memory, but that memory is not automatically pooled across GPUs.
The B300 has 8,000 GB/s of memory bandwidth. Higher bandwidth helps with faster data transfer between GPU memory and compute cores.
The B300 supports 9 precision formats. Training: BF16, FP16, TF32, FP32. Inference: FP4, FP6, FP8, INT8. Scientific: FP64.
Yes. The B300 supports NVLink with 1800 GB/s of bidirectional bandwidth. This helps accelerate multi-GPU communication.
B300 pricing currently ranges from $2.99/hr to $18.00/hr per GPU, depending on the provider, instance type, and billing model.
At 720 hours per month, one B300 can cost between $2,152.80 to $12,960.00 per month, depending on the provider. Reserved and spot pricing can lower that further.
The B300 is available from 25 cloud providers, including Verda, Lyceum, Bentaus. Pricing and availability vary by region and billing model.
Yes. We currently track 93 B300 listings across 25 cloud providers:
| Billing type | Listings | Avg $/GPU/hr |
|---|---|---|
| On-demand | 32 | $8.99/hr |
| Reserved | 40 | $6.23/hr |
| Spot | 16 | $4.68/hr |
| Custom contract | 5 | $4.49/hr |
| Feature | Specification |
|---|---|
| GPU | NVIDIA Blackwell Ultra GPUs |
| Total GPU Memory | 2.3TB |
| Performance | 144 PFLOPS FP4 inference, 72 PFLOPS FP8 training |
| NVIDIA NVLink Switch System | 2x |
| NVIDIA NVLink Bandwidth | 14.4 TB/s aggregate bandwidth |
| Power Consumption | ~14kW |
| CPU | Intel® Xeon® 6776P Processors |
| Networking | 8x OSFP ports (800Gb/s InfiniBand/Ethernet), 2x dual-port QSFP112 BlueField-3 DPU (400Gb/s InfiniBand/Ethernet) |
| Management Network | 1GbE onboard NIC with RJ45, 1GbE RJ45 Host BMC |
| Storage | OS: 2x 1.9TB NVMe M.2, Internal: 8x 3.84TB NVMe E1.S |
| Software | NVIDIA DGX OS, Mission Control, Base Command Manager, AI Enterprise. Supports RHEL, Rocky, Ubuntu. |
| Rack Units | 10RU |
| Operating Temperature | 10°C - 35°C |
Source: official Nvidia B300 datasheet.
Grace + Blackwell Ultra rack-scale superchip. 4 providers, pricing on request.
Current generation Hopper GPU. From $0.92/hr per GPU across 36 providers.
Grace Hopper CPU+GPU superchip. From $6.50/hr per GPU across 3 providers.
Individual B200 GPU. From $3.35/hr per GPU across 27 providers.
Widely available Hopper generation GPU. From $0.60/hr per GPU across 52 providers.
Last updated
