Google Cloud Plans to Deploy Up to 960,000 NVIDIA Rubin GPUs

en.Wedoany.com Reported - On October 21, Google Cloud and NVIDIA announced a new round of artificial intelligence infrastructure deployment plans at the GTC conference held in Berlin, Germany. The A5X cloud computing instances equipped with NVIDIA Vera Rubin NVL72 systems will officially open in November 2026, with plans to expand the service to Europe. Google Cloud intends to deploy up to 960,000 Rubin GPUs across multiple data center sites, providing computing power for large-scale artificial intelligence training, agentic inference, and reinforcement learning tasks.

The A5X adopts the Vera Rubin NVL72 rack-scale computing system, integrating NVIDIA ConnectX-9 high-speed network adapters, Google Cloud Virgo networking, and fourth-generation advanced liquid cooling technology, and connects to Google Cloud AI Hypercomputer infrastructure. According to the network expansion plan announced by both parties, a single site can support up to 80,000 Rubin GPUs, with a multi-site deployment cap of 960,000 units, and cross-site computing resources are scaled through the Virgo network architecture.

Ineffable Intelligence, an artificial intelligence company based in London, UK, has taken the lead in using A5X instances for large-scale reinforcement learning research. Founded by David Silver, a core developer of AlphaGo, the company is developing artificial intelligence systems capable of learning through continuous interaction and experiential feedback. The two parties previously signed a cloud computing cooperation agreement in June 2026, with Ineffable Intelligence selecting Google Cloud as its primary infrastructure partner.

WPP, a UK communications group, has also completed testing of a creative agent architecture on the A5X platform. The system uses NVIDIA Nemotron models and OpenShell technology and can execute tasks such as video generation, image quality enhancement, and asset reconstruction based on natural language instructions. WPP's preliminary tests show that some workloads that originally required multi-node clusters to run can now be executed on a single A5X instance.

Google Cloud also plans to extend NVIDIA confidential computing capabilities to rack-scale bare-metal instances and is studying the use of NVIDIA DSX MaxLPS technology to optimize the compute performance per unit of power and data center power capacity utilization of Rubin infrastructure.