Google Cloud plans to scale Vera Rubin infrastructure to 960,000 GPUs

Google Cloud plans to stretch its NVIDIA Vera Rubin infrastructure to as many as 960,000 GPUs. The vehicle for that expansion is a new family of A5X instances built on NVIDIA’s Vera Rubin NVL72 systems. What Google actually built The A5X instances are bare-metal. Customers get direct access to the physical hardware, without a virtualization layer sitting between their software and the chips. Each instance is powered by NVIDIA’s Vera Rubin NVL72 rack. A single rack combines 72 Rubin GPUs and 36 Vera CPUs, all linked together with NVLink 6. Scaling beyond a single rack requires a different kind of plumbing. Google is leaning on its Virgo networking fabric and NVIDIA ConnectX-9 SuperNICs to link systems together across multiple sites. The performance claims The headline numbers are aggressive. The A5X instances reportedly deliver up to 10 times lower inference cost per token than the previous generation. They are also claimed to offer up to 10 times higher throughput per megawatt. The timeline so far NVIDIA launched the Vera Rubin platform on January 5, 2026. Google then announced the A5X instances at its Google Cloud Next event in April 2026. The first A5X instances went live on Octo...