Rent NVIDIA B300 GPUs

Rent a B300 on Akash

As of October 2026, on-demand B300s on Akash start at $6.00 per GPU-hour. Reserved B300 clusters are quoted on GPU count, term and workload.

For the cost per day, month and year, and whether to rent or buy, see B300 price and rent vs buy.

Two ways to get B300s

Not sure yet? Pilot your workload on-demand, then move steady workloads to reserved capacity.

Built for low-precision inference

The B300 is designed for low-precision AI workloads, including NVFP4 inference. Low-precision checkpoints cut the memory needed for model weights, leaving more room for KV cache, longer context and higher concurrency.

What's new in Blackwell Ultra

• More memory per GPU: 288 GB of HBM3e from 12-high stacks, 50% more than the B200.
• More low-precision compute: 13.5 PFLOPS dense NVFP4, 1.5x the B200.
• Faster attention: NVIDIA rates attention-layer throughput at about 2x the B200, which helps reasoning and long-context models.
• Faster networking: Up to 800 Gb/s per GPU with ConnectX-8, for multi-node training.
• Less FP64: About 1.25 TFLOPS, because silicon moved to AI precision formats.

For the latest benchmark results, see MLCommons MLPerf Inference.

Is the NVIDIA B300 the Right GPU for You?

• Choose the B300

  if you need up to 288 GB of HBM3e memory per GPU, are running memory-intensive LLM or MoE workloads, or need to serve large models and long-context inference workloads.

• Choose the B200

  if 180 GB of GPU memory per GPU is sufficient for your workload and you want Blackwell-generation GPU performance at a lower hourly rate.

• Choose the H200

  if your workload fits within 141 GB of HBM3e memory per GPU, or if your software and infrastructure are already optimized for the Hopper architecture.

For FP64-heavy scientific simulations, B200 or H200 can be a better fit than B300 because B300 prioritizes AI workloads and offers substantially lower FP64 throughput.

For the full side-by-side on specs, speed and cost, see B300 vs B200 vs H200.

The B300 is the GPU inside 8-GPU HGX and DGX B300 systems. GB300 NVL72 is a separate rack-scale platform that pairs 72 Blackwell Ultra GPUs with 36 Grace CPUs.

What is the NVIDIA B300 used for?

The NVIDIA B300 is used to serve and train the largest AI models, especially mixture-of-experts and reasoning models that need more memory than earlier GPUs provide. Its 288 GB per GPU, native 4-bit (NVFP4) compute and faster attention make it best suited to:

• Trillion-parameter model serving

  288 GB per GPU lets the largest open models run on a single 8-GPU server instead of two.

• Reasoning and agent workloads

  Up to 2× faster attention than B200 helps models that generate long chains of thought before answering.

• Long-context inference

  Extra memory holds the large KV cache that 100K–1M-token context windows need.

• 4-bit inference at scale

  Native NVFP4 runs 4-bit checkpoints at full speed, cutting memory per model by about 4× compared with 16-bit.

• Large-model fine-tuning and RL

  1.8 TB/s NVLink links all 8 GPUs inside a server, and InfiniBand links servers for multi-node training.

• Video and multimodal generation

  High-resolution image, video and speech models use the extra memory for longer sequences and larger batches.

Deploy a B300 in three steps

1. 1

   Add your app. Pick a template or paste a Docker image.

2. 2

   Choose B300. Turn on GPU and select B300.

3. 3

   Deploy. Accept a matching bid and start your workload.

NVIDIA B300 GPU specifications

NVIDIA B300 FAQ