NVIDIA Unveils Blackwell as Full-Scale AI Infrastructure Platform

Details

• NVIDIA is repositioning its Blackwell architecture from a standalone chip to a comprehensive AI infrastructure platform, aimed at powering massive AI factories capable of serving nearly a billion users each week.
• The GB200 NVL72 rack-scale system is the centerpiece, packing 72 Blackwell GPUs interconnected with 5,000 high-speed cables, delivering 130 TB/s bandwidth in a 1.5-ton rack built from over 600,000 components.
• The Grace Blackwell superchip integrates two Blackwell GPUs with a Grace CPU, leveraging NVLink interconnects for high-speed memory sharing and reduced latency, critical for demanding AI inference workloads.
• Production spans more than 200 technology partners and 150 factories globally, with NVIDIA reporting a current output of about 1,000 GB300 rack systems per week as of September 2025.
• NVIDIA Dynamo operates as the AI factory's OS, efficiently managing inference requests and resource allocation, while BlueField-3 DPUs support essential networking and security operations outside of AI processing.

Impact

NVIDIA's holistic approach positions it as a direct rival to major cloud hyperscalers building their own AI platforms, potentially accelerating mainstream enterprise AI adoption. By scaling production and integrating hardware, software, and manufacturing, NVIDIA raises the competitive bar for AI infrastructure. Its focus on optimizing inference aligns with industry shifts toward deploying, rather than just training, advanced AI models.