NVIDIA has announced the Blackwell Ultra B300 GPU at its annual GTC conference, delivering a tenfold improvement in AI training performance over the H100 and setting a new benchmark for data center computing. The chip features 288GB of HBM4 memory, a new NVLink 6.0 interconnect capable of 7.2 terabytes per second of bandwidth, and a 5nm manufacturing process from TSMC.
Technical Specifications
The B300 packs 208 billion transistors onto a single die — more than double the transistor count of the H100 — and achieves 20 petaflops of FP8 AI performance. The chip's new Transformer Engine 3.0 is specifically optimized for the attention mechanisms that underpin large language models, delivering up to 15x faster inference for models with more than 100 billion parameters.
"Every major AI lab in the world is training on NVIDIA hardware," said CEO Jensen Huang during the GTC keynote. "The Blackwell Ultra is not just an incremental improvement — it is a generational leap that will define the next era of AI development."
NVLink 6.0 and DGX B300 Systems
NVIDIA is also announcing the DGX B300 system, which connects eight B300 GPUs via NVLink 6.0 to create a unified 2.3 petabyte memory pool. This allows AI models with up to 10 trillion parameters to be trained on a single DGX system without the complex model parallelism strategies currently required for frontier model training.
The DGX B300 will be available as a standalone system for $400,000 and as a cloud instance through AWS, Google Cloud, and Microsoft Azure beginning in Q3 2025.
Market Impact
The announcement sent NVIDIA's stock up 8% in after-hours trading, pushing the company's market capitalization above $3.8 trillion. Analysts at Morgan Stanley estimate that the B300 launch will generate $120 billion in revenue for NVIDIA over the next 18 months, driven by demand from hyperscalers, AI startups, and national AI infrastructure programs.
AMD and Intel are expected to respond with competing announcements later this year, though industry observers note that NVIDIA's software ecosystem — particularly the CUDA platform — remains a significant competitive moat that hardware alone cannot easily overcome.
