Nvidia's Vera Rubin NVL72 rack-scale system can train a 10-trillion-parameter mixture-of-experts model on 100 trillion tokens in one month while using about 75% fewer GPUs than the Grace Blackwell NVL72 configuration, according to Crypto Briefing. The system packs 72 Rubin GPUs and 36 Vera CPUs into a single NVLink domain, and each GPU carries 288 GB of HBM4 memory.
The Vera Rubin NVL72 is the direct successor to Nvidia's Blackwell NVL72 racks, which shipped in GB200 and GB300 variants. Nvidia projects that running inference on the new system will cost one-tenth the price per million tokens compared with Blackwell racks, Crypto Briefing reported.
Benchmarks behind the Vera Rubin NVL72 claims
Third-party testing offers some support for the performance claims around the Vera Rubin NVL72. MLPerf Inference v6.1, released in September 2026, recorded throughput improvements of up to 3.7 times higher than the GB300 NVL72 on the Qwen3-VL task, according to Crypto Briefing.
Separately, benchmarks from Artificial Analysis showed approximately 5.7 times higher throughput per megawatt than the GB300 NVL72 on agentic AI workloads. The figures arrive as production shipments of the platform ramp in fall 2026.
Google Cloud A5X instances and scaling plans
Google Cloud introduced A5X bare-metal instances powered by Nvidia Vera Rubin NVL72 rack-scale systems, targeting large-scale agentic inference and reinforcement learning, StorageNewsletter reported. General availability is scheduled for November 2026, with planned expansion of availability to Europe.
The A5X instances combine Nvidia Vera Rubin NVL72 systems with ConnectX-9 SuperNICs, Google Cloud's Virgo network and the cloud provider's fourth-generation liquid-cooling technology. Google Cloud plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites, with deployments reaching 80,000 Rubin GPUs within a single site.
Google Cloud has progressively integrated Nvidia's GPU generations, including Hopper, Blackwell and now Vera Rubin, into its AI Hypercomputer architecture. The A5X launch brings Vera Rubin into that architecture as an integrated system combining compute, networking and storage.
The two companies are also exploring Nvidia DSX MaxLPS capabilities for Vera Rubin-based infrastructure, an effort aimed at optimizing performance per watt and expanding capacity within available data-center power budgets, StorageNewsletter reported.
Early customers test reinforcement learning and creative agents
Ineffable Intelligence, a London-based technology company founded by AlphaGo creator David Silver, was the first customer to access A5X. The company is using the platform to advance large-scale reinforcement learning and to develop systems that learn continuously from experience.
"We're building superlearners that continually improve through their own experience," Silver said.
WPP, a London-based marketing and communications company, has also demonstrated its creative agent reference architecture on Vera Rubin-based A5X instances, using Nvidia Nemotron models and Nvidia OpenShell. In WPP's initial tests, workloads that typically required a multi-node cluster ran on a single A5X instance, StorageNewsletter reported.
Sovereign AI deployments reach the US and UAE
Sovereign AI infrastructure provider Aleria is expanding its compute capacity with Nvidia Blackwell and Vera Rubin deployments, Data Center Dynamics reported. The company will bring 8,640 Nvidia Grace Blackwell Ultra GPUs to the US, with plans to eventually expand to 16,000, and will deploy 28 racks of the Nvidia DGX Vera Rubin NVL72 in the United Arab Emirates.
Aleria said the Vera Rubin deployment will be one of the first in the region. Deploying Nvidia Blackwell Ultra in the United States – and DGX Vera Rubin to come in the UAE – is the next chapter of something already proven, said Eric Leandri, CEO of Aleria.
The company will use DDN's high-performance storage architecture to support both deployments. Ankur Arora, senior regional director of the Middle East and Africa at DDN, said the collaboration brings together "the performance and scale required for AI-first environments, where organizations cannot afford to compromise on security, reliability, or data sovereignty."
Storage and confidential computing round out the stack
DDN this week also revealed that European AI cloud company Nebul had selected DDN Infinia for storage and key-value cache to help its GPUs serve AI models. "Every time a GPU rebuilds context it has already processed, it is spending cycles repeating work," said Alex Bouzari, CEO and co-founder at DDN.
"AI sovereignty should not come at the expense of performance," added Arnold Juffer, CEO of Nebul.
Google Cloud, for its part, will be among the first cloud providers to offer Nvidia Confidential Computing at rack scale on bare metal instances such as A4X Max instances powered by Nvidia GB300 NVL72. It is also making Confidential Computing generally available on G4 VMs powered by Nvidia RTX PRO 6000 Blackwell Server Edition GPUs, targeting confidential inference and fine-tuning.
Conclusion
Nvidia's Vera Rubin NVL72 enters the market with claimed training efficiency of about 75% fewer GPUs than Grace Blackwell and projected inference costs at one-tenth per million tokens, with MLPerf and Artificial Analysis benchmarks offering partial third-party support. Production shipments are ramping in fall 2026, and the next concrete milestone is Google Cloud's A5X general availability in November 2026, alongside Aleria's planned deployments in the US and the UAE.
Frequently Asked Questions
How many fewer GPUs does the Vera Rubin NVL72 need for training than Grace Blackwell?
According to Crypto Briefing, the Vera Rubin NVL72 can train a 10-trillion-parameter mixture-of-experts model on 100 trillion tokens in one month using about 75% fewer GPUs than the Grace Blackwell NVL72 configuration. The system packs 72 Rubin GPUs and 36 Vera CPUs into a single NVLink domain.
What inference cost does Nvidia project for Vera Rubin NVL72 compared with Blackwell?
Nvidia projects that running inference on the Vera Rubin NVL72 will cost one-tenth the price per million tokens compared with Blackwell racks, Crypto Briefing reported. Third-party benchmarks, including MLPerf Inference v6.1 and Artificial Analysis tests, recorded throughput gains of up to 3.7 times and about 5.7 times per megawatt over the GB300 NVL72.
When will Google Cloud's A5X instances based on Vera Rubin NVL72 be available?
Google Cloud's A5X bare-metal instances powered by Nvidia Vera Rubin NVL72 are scheduled for general availability in November 2026, with planned expansion to Europe, StorageNewsletter reported. Google Cloud plans to scale Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites.
How many Vera Rubin GPUs does Google Cloud plan to deploy?
Google Cloud plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites, with deployments reaching 80,000 Rubin GPUs within a single site, StorageNewsletter reported. The A5X instances pair the Vera Rubin NVL72 with ConnectX-9 SuperNICs, the Virgo network and fourth-generation liquid cooling.
What is Aleria deploying in the US and UAE with Nvidia Vera Rubin?
Aleria will bring 8,640 Nvidia Grace Blackwell Ultra GPUs to the US, with plans to expand to 16,000, and will deploy 28 racks of the Nvidia DGX Vera Rubin NVL72 in the United Arab Emirates, Data Center Dynamics reported. Aleria said the UAE Vera Rubin deployment will be one of the first in the region.