Google Cloud opens Nvidia Vera Rubin to customers, targets 960,000 GPUs

Summary

• Google Cloud and Nvidia announced general availability of the Vera Rubin infrastructure at GTC Berlin.
• A5X instances, combining NVL72 systems and liquid cooling, reach general availability next month.
• Google Cloud plans to scale to 960,000 Rubin GPUs across multiple sites, including Europe.

The latest

A new generation of Nvidia AI infrastructure is now generally available on Google Cloud, with the companies announcing the rollout of Vera Rubin at Nvidia's GTC Berlin conference. The A5X instances built on it reach general availability next month. Google Cloud said it plans to expand the architecture to as many as 960,000 Rubin GPUs spread across multiple sites.

Details

• What A5X is: The A5X instances combine Nvidia Vera Rubin NVL72 systems, ConnectX-9 SuperNICs, Google Cloud's Virgo network fabric and fourth-generation liquid cooling. They are delivered inside the AI Hypercomputer platform, which packages compute, networking and storage as a single integrated system rather than separate components.
• The scale target: Google Cloud said it intends to grow the Vera Rubin footprint to up to 960,000 Rubin GPUs distributed across multiple locations, with expansion into Europe planned after availability. The company did not give a timetable for reaching that figure or name the specific European sites.
• First customer: The first customer to get access to A5X is Ineffable Intelligence, a London company founded by David Silver, the researcher behind AlphaGo. According to the announcement, it is using the platform for large-scale reinforcement learning and what it calls superlearner systems.
• Silver's framing: Silver, the company's founder and chief executive, said the firm is building "superlearners that continually improve from their own experience." The announcement did not detail the size of the models involved or how much of the Vera Rubin capacity the startup has been allocated.
• Early benchmark: Advertising group WPP presented a reference architecture for a creative agent running on A5X using Nvidia Nemotron models. In initial testing, according to the announcement, workloads that previously required a multi-node cluster ran on a single A5X instance.
• Confidential computing: Google Cloud said it is among the first cloud providers to offer Nvidia confidential computing at rack scale on bare-metal instances. Confidential computing on G4 virtual machines is now generally available, extending encrypted-in-use protections to a wider set of workloads.
• Power efficiency: The two companies said they are exploring Nvidia DSX MaxLPS technology to improve performance per watt and add capacity within existing data-center power budgets, a constraint increasingly shaping how fast AI infrastructure can be deployed.

Background

Vera Rubin is Nvidia's successor architecture to Blackwell. NVL72 refers to rack-scale systems that link 72 GPUs into a single compute domain, designed for training large models and running inference across long-context agent workloads.

Between the lines

The power-efficiency work signals the real constraint. Google Cloud's 960,000-GPU target depends on data-center power budgets, which is why the two companies are exploring DSX MaxLPS to add capacity within existing limits rather than only building new sites. The WPP result points the same way: if multi-node workloads collapse onto one instance, the economics shift from raw chip count to density per watt.

What's next

A5X general availability next month is the first checkpoint. Watch for the European expansion sites, disclosure of deployed GPU counts against the 960,000 target, and whether WPP's single-instance result holds in production workloads.

Source: NVIDIA