One hour
$6.94
1 GPU × 1h · compute only
Blackwell Ultra cloud GPU
Rent an NVIDIA B300 GPU for frontier-scale AI.
Direct answer
NVIDIA B300 is a Blackwell Ultra data-center GPU with 288 GB of HBM3e memory per accelerator. GPURento lists B300 from $6.94 per GPU-hour in Paris and Frankfurt for high-memory training, reasoning and inference through the funded dashboard.
Reviewed by the GPURento infrastructure team · September 1, 2026
Cost scenarios
What one B300 costs at the listed rate.
One day
$166.56
1 GPU × 24h · compute only
Seven days
$1165.92
1 GPU × 168h · compute only
Average month · 730h
$5066.20
1 GPU × 730h · compute only
| Memory | 288 GB HBM3e | Per-GPU memory in NVIDIA’s HGX B300 reference architecture. |
|---|---|---|
| Architecture | Blackwell Ultra | A newer data-center generation than B200. |
| On-demand rate | $6.94 / GPU-hour | Current GPURento compute rate before storage and optional services. |
| Available regions | Paris · Frankfurt | Prepaid resources need a positive wallet balance. The $50 minimum applies to deposits, not to your remaining balance. |
Best for
- Frontier-model training and continued pre-training
- Very long-context or high-concurrency inference
- Memory-bound workloads that exceed B200 or H200
- Teams validating a Blackwell Ultra software stack
Not the best fit when
- Price-first experiments that fit on smaller GPUs
- Unprofiled software without Blackwell kernel support
- Jobs that cannot keep a high-end accelerator utilized
- Choosing from peak specifications without a pilot run
Decision method
What to verify before you deploy capacity.
The content below separates published specifications, GPURento catalog references and decisions that still require a workload benchmark.
01
Use B300 when memory changes the system design
The practical value of 288 GB per GPU is reducing unwanted sharding, supporting larger runtime caches or increasing useful batch and context sizes. Estimate the complete runtime state and confirm that the software can use the additional memory before paying the higher rate.
- Calculate weights, optimizer state, activations or KV cache
- Compare the GPU count required on B300 and the nearest alternative
- Keep model quality and service level fixed during benchmarks
02
Validate Blackwell Ultra compatibility first
Framework, driver, CUDA, collective and custom-kernel compatibility can determine time to result. Run a representative step or inference trace before allocating the full job, especially when migrating a stack tuned for Hopper or Ampere.
03
Price the complete topology
For multi-GPU B300 work, host memory, network fabric, GPU interconnect and checkpoint storage are part of the product. Compare cost per completed checkpoint or accepted request rather than multiplying a peak throughput claim.
Catalog shortlist
Relevant GPU options.
Blackwell UltraAvailable
NVIDIA B300
288 GB HBM3e · $6.94/hr
Frontier reasoning and very long contexts
EU West · ParisEU Central · Frankfurt
BlackwellAvailable
NVIDIA B200
180 GB HBM3e · $5.98/hr
Frontier training and inference
EU West · ParisEU Central · Frankfurt
HopperAvailable
NVIDIA H200 SXM
141 GB HBM3e · $3.59/hr
Large-model inference and HPC
EU West · ParisEU Central · Frankfurt
HopperAvailable
NVIDIA H100 SXM
80 GB HBM3 · $2.69/hr
Intensive training and FP8 inference
EU West · ParisEU Central · Frankfurt
Questions
Clear answers, including the limits.
How much memory does an NVIDIA B300 GPU have?+
The B300 configuration described in NVIDIA’s HGX reference architecture has 288 GB of HBM3e memory per GPU.
How much does it cost to rent a B300 GPU?+
The current GPURento catalog rate is $6.94 per B300 GPU-hour. Persistent storage and optional services are calculated separately.
Should I choose B300 instead of B200?+
Choose B300 when the larger memory or newer software path materially reduces sharding, improves the service level or shortens the run enough to offset the higher rate. Benchmark both when the workload fits either GPU.
Sources and scope
Last reviewed September 1, 2026. GPURento rates are current catalog references; provisioning state remains visible in the workspace, and manufacturer specifications do not substitute for workload benchmarks.
Continue the research
Deployment plan
Validate the B300 workload before scaling it.
Create a funded workspace and save the GPU count, image, region, runtime and storage for a representative pilot.