Quick take: NVIDIA says Vera Rubin on Google Cloud is now generally available. The new A5X instances, built on Vera Rubin NVL72 racks and ConnectX-9 SuperNICs, are aimed at training, inference and reinforcement learning at very large scale.
The Next Wave of AI Hardware Goes Live
In AI, the story is often told in chips and racks. This week’s announcement from NVIDIA AI Infrastructure puts its newest platform in the hands of any customer who can rent it on Google Cloud, not just a few early partners.
A5X at a Glance
| Detail | Information |
|---|---|
| Instance | A5X (bare metal) |
| Hardware | NVIDIA Vera Rubin NVL72 racks with ConnectX-9 SuperNICs |
| Single-site scale | Up to 80,000 Rubin GPUs |
| Multi-site scale | Up to 960,000 Rubin GPUs |
| Claimed efficiency | Up to 10x lower inference cost per token and 10x higher token throughput per megawatt versus the prior generation |
| Target workloads | Training, inference and reinforcement learning |
Specifications and claims are from NVIDIA’s April announcement. The 10x figures are vendor claims, not independent benchmarks.
Who Is Already Using It?
NVIDIA highlights European innovators, including Ineffable Labs, which it says is using the platform to build the next generation of “superlearner” systems. In July, NVIDIA said Google Cloud’s first A5X instance based on Vera Rubin NVL72 was already running for the London startup.
Why It Matters
- Cost per token: If the claimed gains hold, large chatbots and agents become cheaper to run.
- Scale: Clusters approaching a million GPUs show how far AI infrastructure has stretched.
- Competition: Vera Rubin racks have also been reported at CoreWeave, Microsoft Azure and Oracle Cloud.
For background, read about NVIDIA and Google Cloud. The same compute race is driving huge spending at AI labs, as seen in our report on Anthropic’s IPO filing. More on AI and hardware is in our technology section on TechnoSports.
Sources: NVIDIA AI Infrastructure on X, NVIDIA Blog, IT Brief UK, Wccftech.
Frequently Asked Questions
What is A5X?
Google Cloud’s bare-metal instance type built on NVIDIA Vera Rubin NVL72 systems.
How big can A5X clusters get?
Up to 80,000 Rubin GPUs at one site and up to 960,000 across multiple sites.
Are the 10x claims proven?
They are NVIDIA and Google’s figures against the previous generation, so wait for independent tests.
What is an A5X instance?
A5X is Google Cloud’s bare-metal instance type built on NVIDIA Vera Rubin NVL72 systems.
