Nvidia has announced a significant breakthrough in data center efficiency with its new GB300 NVL72 system. The platform supports 61,400 concurrent AI agents per megawatt, representing a twenty-fold increase over the previous H200 generation. This advancement addresses critical energy constraints by fundamentally altering the economics of AI inference.
Built on the Blackwell Ultra architecture, the liquid-cooled rack integrates 72 GPUs and 36 Grace CPUs with 130 TB/s of NVLink bandwidth. The system delivers fifty times the AI factory output of older Hopper platforms and five times the throughput per watt. Software optimizations like fused Mixture of Experts techniques further maximize computation efficiency by activating only relevant model parameters.
Major cloud providers are already integrating the hardware. Microsoft Azure will deploy the first large-scale clusters to power OpenAI workloads starting in late 2025. CoreWeave has launched production instances, while Oracle Cloud Infrastructure is also preparing deployment. For investors, this efficiency gain offers a direct ROI pathway and strengthens Nvidia's position amid increasing scrutiny on AI energy consumption.