B200
180GB HBM3e per GPU
View GPU All GPUsB300
Blackwell Ultra architecture288GB HBM3e per GPU
GPU Catalog / NVIDIA
B300
GPU infrastructure.
A reserve-only Blackwell Ultra GPU for high-memory AI reasoning, inference, and training infrastructure.
- Architecture
- Blackwell Ultra
- GPU memory
- 288GB HBM3e per GPU
- Availability
- Reservation only
- Reservation rate
- Pricing on request
- Lead time
- 3-7 weeks
Reservation
Reserve capacity.
Estimated lead time is 3-7 weeks, subject to availability and shipping.
- Access
- Reservation only
- Rate
- Pricing on request
- Estimated lead time
- 3-7 weeks
- Qualification
- Subject to availability and shipping.
Hardware profile
Where B300 fits.
NVIDIA B300 is a Blackwell Ultra accelerator with 288GB of HBM3e memory per GPU. NVIDIA positions the platform for AI reasoning, large language model inference, and training workloads with large memory requirements.
Choose B300 when Blackwell Ultra memory capacity and reasoning-oriented AI infrastructure are central requirements and a planned reservation window fits the deployment schedule.
NVIDIA product information- 01
AI reasoning inference
Provision high-memory Blackwell Ultra infrastructure for reasoning models and test-time scaling workloads.
- 02
Large-model training
Plan multi-GPU training infrastructure for transformer models with substantial compute and memory requirements.
- 03
High-memory AI pipelines
Use 288GB of HBM3e per GPU for models and runtime state that exceed smaller accelerator memory envelopes.
Deployment sizing
Size before you provision.
Hardware fit depends on the model, precision, batch size, software stack, and total working set. Published use cases are not performance guarantees.
- 01 / Memory
Confirm the working set
Include weights, caches, activations, batches, and framework overhead.
- 02 / Software
Validate the stack
Check CUDA, driver, framework, precision, and container requirements.
- 03 / Capacity
Plan the lead time
Allow 3-7 weeks for availability and shipping before the planned deployment.