Google Cloud and NVIDIA are preparing to make Vera Rubin-based A5X instances generally available in November, according to their October 21 announcements. The rollout is intended to support demanding AI training, inference and reinforcement-learning workloads.
TL;DR: A5X availability is planned for next month. The companies describe much larger future GPU deployments, but those figures are capacity plans rather than an inventory already operating.
Networking and storage accompany the new processors
Google’s technical overview describes A5X as bare-metal instances using NVIDIA Vera Rubin NVL72 systems. It also outlines expanded availability in regions serving European, Middle Eastern and African data-residency requirements.
Google says its Virgo networking architecture will support clusters of up to 80,000 Rubin GPUs at one site and 960,000 across multiple sites. Planned Managed Lustre support for GPUDirect Storage would allow training and inference data to reach GPU memory without passing through the CPU and system memory.
Those changes address how a large system moves information, alongside how quickly an individual processor computes. Adding processors can only help a workload to the extent that useful work and data can reach them. The announcements therefore describe an integrated infrastructure design, whose practical results will depend on how customers use it.
NVIDIA identifies Ineffable Intelligence, the London company founded by David Silver, as A5X’s first customer. It is using the infrastructure for reinforcement-learning research. NVIDIA also describes WPP demonstrations involving creative agents, with people approving actions before execution.
These are early examples, not a general performance guarantee. Neither a research deployment nor a customer demonstration establishes the price or benefit another organisation would obtain from moving its own applications.
Security is another part of the infrastructure choice
A separate NVIDIA announcement on October 21 introduces Confidential Bare Metal. The company says it is designed to let model builders use dedicated third-party infrastructure while retaining control over their software stack, models and data.
NVIDIA describes controls for protecting data in transit, checking platform integrity and enforcing access restrictions in hardware. It places bare metal alongside confidential containers and virtual machines as distinct deployment options, rather than treating every workload as requiring the same environment.
The design question is broader than raw processing capacity. A business may need to establish where information is handled, which parties can access it and which software it can operate itself. Those requirements can influence whether a particular cloud configuration is suitable even when ample compute is available.
For A5X buyers, the immediate milestone remains the planned November availability. Specific regions, service terms and the controls attached to a chosen instance will matter more than treating the largest announced cluster size as a ready-to-order deployment.
Written by the News Desk and edited by Samuel Rae.
Leave a Reply