Google Cloud, NVIDIA Launch Vera Rubin AI Infrastructure

Google Cloud and NVIDIA will make the new A5X instance generally available next month, with plans to expand into Europe shortly after. The instance pairs NVIDIA's Vera Rubin NVL72 system with ConnectX-9 SuperNICs, Google Cloud's Virgo Network and its fourth-generation liquid-cooling setup, built specifically for agentic inference and reinforcement learning at scale. Google Cloud says it intends to grow its Vera Rubin footprint to as many as 960,000 Rubin GPUs spread across multiple sites, folding the new hardware into its AI Hypercomputer as a single integrated system covering compute, networking and storage.

The first company given access to A5X is Ineffable Intelligence, a London startup founded by David Silver, the DeepMind researcher behind AlphaGo. Ineffable is using the infrastructure to train what it calls "superlearner" systems, models designed to improve continuously by generating experience, evaluating results and updating themselves in a tight loop rather than training on fixed datasets. That kind of learning puts unusual demands on infrastructure, since every step in the cycle has to run at scale without creating bottlenecks between generating experience and training on it. Silver described what that requires in his own words below.