Google Cloud A5X Vera Rubin Instances Launch Next Month With 10x Efficiency Claims

Google Cloud and NVIDIA will make A5X instances powered by NVIDIA Vera Rubin NVL72 generally available next month. The architecture combines Vera Rubin NVL72 GPUs, ConnectX-9 SuperNICs, Google Cloud's Virgo Network, and fourth-generation liquid cooling. Within a single site the instances scale to 80,000 Rubin GPUs; across multiple sites Google Cloud states the infrastructure can reach 960,000 Rubin GPUs. The systems deliver up to 10 times lower inference cost per token and up to 10 times higher token throughput per megawatt compared to the prior generation. The architecture is designed for reinforcement learning workflows involving repeated generation, scoring, and retraining. Ineffable Intelligence, founded by AlphaGo creator David Silver, was the first customer to access A5X instances and is using them for large-scale reinforcement learning research. WPP tested the instances and found that workloads normally requiring a multi-node cluster ran on a single A5X instance. Google Cloud also plans to offer NVIDIA Confidential Computing at rack scale on bare metal instances, including on GB300 NVL72 and RTX PRO 6000 Blackwell GPUs.