# Nvidia Says Vera Rubin Reaches Up to 30 Times GB300’s Cost-Adjusted Inference Throughput

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/50498
Published: 2026-11-07T01:59:04.000Z
Updated: 2026-11-07T01:59:04.000Z
Section: Technology

> Vera Rubin NVL72’s per-GPU high-bandwidth memory (HBM) capacity is about 3% higher than GB300 NVL72’s. GB300 NVL72 reached up to 10 times H200’s cost-adjusted inference throughput on specific workloads.

Nvidia outlined a memory strategy for growing AI systems, with Vera Rubin NVL72 reaching up to 30 times GB300 NVL72’s cost-adjusted inference throughput.

The approach combines GPUs, CPUs, networking, storage and software to improve memory use as models grow, context windows lengthen and agent applications spread. Nvidia’s methods include compressing model weights and KV Cache with NVFP4, a low-precision format. The approach also moves inactive context to lower-cost storage and distributes work across multiple GPUs to free HBM.

In the same comparison, GB300 NVL72 delivered up to 10 times H200’s cost-adjusted inference throughput on specific workloads. Vera Rubin NVL72 reached up to 30 times GB300 NVL72’s cost-adjusted inference throughput, while its per-GPU high-bandwidth memory (HBM) capacity was about 3% higher than GB300 NVL72’s.