NVIDIA says Vera Rubin NVL72 delivers 30 times more performance per dollar on AI agents
NVIDIA claims its Vera Rubin NVL72 system offers up to 30 times more performance per dollar than GB300 NVL72 on AI agent workloads, with only 3% more HBM memory per GPU, according to a blog post signed by Amr Elmeleegy. The figure comes from the SemiAnalysis AgentX test with DeepSeek V4 Pro at 125 tokens per second per user. The company attributes the gain to the NVFP4 format, the Vera CPU and the Dynamo software with CMX storage.

















