Google Cloud and Nvidia say Nvidia Vera Rubin infrastructure is coming to Google’s cloud as new A5X instances, with general availability set for next month. The two companies also revealed a plan to scale Google’s Vera Rubin build-out to as many as 960,000 Rubin GPUs across multiple sites.
That is a huge number, and it shows how fast the biggest clouds are racing to lock in next-generation AI chips. Here’s what A5X is, who is using it first and what else changed.
What the Google Cloud Vera Rubin A5X Instances Include
A5X is Google Cloud’s name for its Vera Rubin-based machines. Each one is built around Nvidia’s Vera Rubin NVL72 rack, which links Vera CPUs and Rubin GPUs into one large system. Google then wraps that rack in its own networking, cooling and storage.
The companies say A5X is aimed at agentic inference and large-scale reinforcement learning, where models learn by trying things and improving from the results. It plugs into Google Cloud’s AI Hypercomputer setup as a single integrated system.
| Component | What it does |
|---|---|
| Nvidia Vera Rubin NVL72 | Rack-scale system pairing Vera CPUs with Rubin GPUs |
| Nvidia ConnectX-9 SuperNICs | High-speed network cards for moving data between machines |
| Google Cloud Virgo Network | Google’s data center network fabric for AI clusters |
| Fourth-gen liquid cooling | Google’s latest cooling for dense, power-hungry racks |
| Planned scale | Up to 960,000 Rubin GPUs across multiple sites |
| Timing | General availability next month, Europe expansion planned |
Early Vera Rubin Customers on Google Cloud
The first customer to get access is Ineffable Intelligence, a London company founded by AlphaGo creator David Silver. It is using A5X to train what it calls “superlearner” systems that keep learning from their own experience.
That kind of training needs a tight loop between generating experience and updating the model. Silver says Vera Rubin’s fast links between its CPUs and GPUs let the team push its reinforcement learning research further.
Marketing giant WPP also tested a creative agent setup on A5X using Nvidia Nemotron models and OpenShell, the sandbox we covered in our story on safe AI agents. Its agents turn chat requests into multistep video jobs, with a human approving each step first. WPP says work that usually needed a multi-node cluster ran on a single A5X instance.
More Secure AI on Google Cloud
The announcement wasn’t only about raw power. Google Cloud is also expanding Nvidia Confidential Computing, which keeps models and data protected even while they run on hardware the customer doesn’t own.
- Google Cloud will be among the first to offer Nvidia Confidential Computing at rack scale on bare metal, starting with A4X Max instances built on GB300 NVL72.
- Confidential Computing is now generally available on G4 virtual machines with Nvidia RTX PRO 6000 Blackwell Server Edition GPUs.
- Early users of Confidential G4 include Vertebrae, Polaris Cloud and Confidential Core AI.
- The two companies are also exploring Nvidia DSX MaxLPS to squeeze more performance out of each watt of data center power.
Raw GPU counts grab headlines, but private, verifiable compute may matter more to companies handing their data to AI agents.
Why This Matters in the AI Compute Race
Every major AI lab is fighting for chips and power. Recent filings have shown just how hard even top model makers have to work for cloud capacity. A plan for nearly a million Rubin GPUs signals that Google intends to be a major supplier of that capacity.
It also shows Google running a two-track strategy. The company keeps building its own TPUs, including new work on PyTorch on TPUs, while offering Nvidia’s newest hardware to customers who want it. Nvidia, for its part, recently opened up its Vera Rubin designs to speed up data center builds.
What This Means for You
Most people won’t rent a Vera Rubin rack. Still, this kind of capacity shapes the AI tools you use every day.
- More compute usually means faster, cheaper access to large models over time.
- If your company runs AI on Google Cloud, A5X is the instance family to watch for heavy training and agent workloads.
- If you handle sensitive data, the Confidential G4 option is available now and is worth asking your cloud team about.
For more stories like this, check our AI news page.
Sources
Primary source: Nvidia’s Google Cloud Vera Rubin announcement.
Conclusion
Google Cloud’s A5X instances bring Nvidia Vera Rubin into its cloud next month, backed by a plan to scale to as many as 960,000 Rubin GPUs. Early users like Ineffable Intelligence and WPP show the focus on reinforcement learning and agent workloads. The expanded confidential computing options add a privacy layer that many businesses will value. Together, these moves keep Google squarely in the race to supply the compute that next-generation AI needs.
