1.5X
GPU memory vs B200 (288GB)
GPU Compute · Blackwell Ultra
NVIDIA B300 — Blackwell Ultra for frontier-scale AI
288GB of HBM3e per GPU and a generational leap in FP4 compute — built for the largest models, reasoning workloads and real-time inference. Reserve dedicated B300 power, in limited supply.
Pricing starts at $2.93 per GPU/hr
Specifications
Built for the largest models yet
Blackwell Ultra brings even more memory and bandwidth than the B200 — with a dedicated decompression engine for data-heavy pipelines.
Architecture Blackwell Ultra
GPU memory 288GB HBM3e
Memory bandwidth 8 TB/s
NVLink 5th-gen · 1.8 TB/s per GPU
Server configuration
NVIDIA HGX B300 GPU Server
Everything a Blackwell Ultra node needs, delivered as one dedicated server — compute, memory, storage and networking pre-integrated and ready to run.
- GPUs
- 8x NVIDIA B300 SXM (HGX B300), 2,304 GB total HBM3e, NVLink-connected
- CPU
- 2x Intel Xeon 6972P — 192 cores total
- Memory
- 1.5 TB DDR5
- Local storage
- 8x 3.84 TB NVMe SSD (~30 TB)
- GPU fabric
- 8x NVIDIA ConnectX-8 InfiniBand, up to 800 Gb/s each (6.4 Tb/s aggregate) for multi-node scaling
- Data / storage network
- 2x 100 GbE
- Internet uplink
- Dual 100 Gbps — two independent 100GbE uplinks at the DC edge deliver redundant, high-throughput public connectivity to our upstream providers, with no single point of failure.
- Management
- Dedicated 10 GbE + IPMI/BMC
Performance
Beyond Blackwell
Compared against the B200 and the previous-generation Hopper, the B300 pushes both memory and compute further.
1.5X
FP4 compute vs B200
50X
AI factory output vs Hopper
Use cases
What the B300 powers
From frontier-model training and reasoning workloads to real-time production inference and HPC.
Frontier-scale training
288GB of HBM3e per GPU gives the biggest models and longest context windows room to train without partitioning.
Reasoning & agentic AI
Purpose-built for test-time scaling — high-throughput inference for reasoning models and multi-step agents.
Real-time inference
Serve the largest LLMs and multi-modal models at low latency, at production scale.
High-performance computing
Large-memory simulation, modeling and scientific workloads that need every bit of bandwidth the node can deliver.
Data analytics
An on-GPU decompression engine keeps data-heavy pipelines fed, so your GPUs spend their time computing instead of waiting on I/O.
Multi-node clusters
NVLink-connected nodes with high-bandwidth InfiniBand scale out to full training clusters.
Reserve your NVIDIA B300 access
Capacity is limited. Talk to our team to lock in dedicated B300 nodes for your frontier training and reasoning workloads.