Google Cloud will make its Nvidia Vera Rubin-based A5X instances generally available in November, the companies announced October 21 at GTC Berlin. The rollout will include new local regions serving Europe, the Middle East and Africa, aimed at customers with data-residency requirements.
The bare-metal instances combine Rubin NVL72 systems with Nvidia ConnectX-9 networking, Google’s Virgo network and its fourth-generation liquid cooling. They are intended for model training, inference and reinforcement learning. Access by early customers precedes the broader commercial rollout.
Nvidia named London-based Ineffable Intelligence, founded by AlphaGo creator David Silver, as the first customer to access A5X. The company is using the systems for reinforcement-learning research, which repeatedly generates experience, evaluates results and updates models.
WPP has also demonstrated a creative-agent system on A5X using Nvidia’s Nemotron models and OpenShell software. Its agents turn requests into workflows such as generating, upscaling and restyling video, with human approval before execution. OpenShell supplies controlled execution environments that restrict access to brand assets and creative tools.
Google says its network design will support as many as 80,000 Rubin GPUs at one site and up to 960,000 across multiple sites. Those figures describe supported scale and planned expansion, not an already installed fleet of that size.
The company also plans direct transfers from its Managed Lustre storage service into GPU memory, bypassing the CPU and host memory. A separate caching layer will hold and reuse model context beyond GPU high-bandwidth memory for longer conversations and multi-step agent tasks. Both features were described as forthcoming support.
Alongside the A5X announcement, Google made Confidential G4 instances generally available across 36 zones. Those machines use Nvidia RTX PRO 6000 Blackwell Server Edition GPUs, a different platform from Rubin. They run on Compute Engine virtual machines and Google Kubernetes Engine nodes.
Google says Confidential G4 combines memory encryption with hardware-based verification of the environment before sensitive data is released. The protections are designed to shield model parameters and customer inputs while they are being processed. Support for Confidential Space, which permits protected data collaboration among multiple parties, is rolling out in the coming weeks.
Google’s tests of Confidential G4 covered models including Llama 3.3 and SDXL. The company reported throughput degradation below 3.5 percent against standard G4 instances, while cautioning that results vary with software, hardware and workload configuration.
Cloud providers are also developing their own accelerators. Microsoft’s Maia 300 plans call for multi-gigawatt deployments beginning in 2027. Google has separately contracted for additional nuclear output from upgraded Constellation reactors, with the first increase expected in 2028.
Sources: Google Cloud, Nvidia, Google Cloud
–
By the Control Plane Editorial Team