Google Cloud readies NVIDIA Vera Rubin instances, with Ineffable and WPP as early users

Google Cloud and NVIDIA are preparing general availability next month for A5X instances built on NVIDIA's Vera Rubin NVL72 systems, with European expansion planned. The pitch is a single platform for training, inference and large-scale reinforcement learning, and Google Cloud says it could scale to as many as 960,000 Rubin GPUs across multiple sites.

The first customer to run on A5X is Ineffable Intelligence, the London lab founded by David Silver of AlphaGo fame, which is using it for continuous learning systems. Advertising and communications group WPP has also shown a creative agent setup on the same hardware. In early tests, WPP says work that normally spans a multi-node cluster ran on one A5X instance, which is the kind of consolidation that changes cloud bills if it holds up outside a demo.

Google Cloud is also making confidential computing generally available on G4 VMs, with customers including Vertebrae and Polaris Cloud running privacy-sensitive inference there. Energy efficiency is being explored through NVIDIA's DSX MaxLPS tools. None of the performance or cost figures are published yet, and the announcement comes from NVIDIA's own blog, so treat the single-instance result as a company-reported early test.

For rivals, the signal is about lead time. The companies that get Vera Rubin capacity first will set the price of serious agent and reinforcement learning work, and everyone else will be shopping for the leftovers.