NVIDIA has begun shipping Blackwell Ultra B300 accelerators to major hyperscalers, delivering the most significant performance leap in AI infrastructure. The new chips power next-generation data centers optimized for large language model training and inference.
Performance Specifications
The B300 Ultra features 208 billion transistors, 192GB of HBM3e memory, and delivers 10 exaflops of AI compute per chip. Compared to the H100, inference performance improves by 10x for large language models while energy efficiency improves by 4x.
The new NVLink Switch enables scaling to 10,000+ GPUs in a single training cluster. This enables training runs for models exceeding 1 trillion parameters with unprecedented efficiency.
Customer Deployments
Microsoft Azure, Amazon AWS, and Google Cloud have all announced B300 Ultra availability. Microsoft's deployment powers enhanced Azure OpenAI services, while AWS offers the chips through SageMaker and Bedrock.
Jensen Huang, NVIDIA CEO, stated: 'Blackwell Ultra is our greatest architectural leap ever. We're enabling AI systems that were impossible just a year ago.'
Supply and Demand
NVIDIA projects $100 billion in data center revenue for fiscal 2027, with Blackwell representing the majority of shipments by mid-year. Demand continues to exceed supply, with hyperscalers placing orders through 2027.
Taiwan Semiconductor's advanced packaging capacity remains the primary bottleneck. NVIDIA has secured priority allocation but lead times extend 6-9 months for new orders.
Key Takeaways
Blackwell Ultra delivers 10x inference improvement over H100 for LLMs. 208 billion transistors and 192GB HBM3e memory per chip. Major hyperscalers deploying for AI model training and serving. Demand continues to exceed supply through 2027.
Related: [AI Hub](/ai) • [AI Hardware Coverage](/topics/ai-hardware)