The Vera Rubin era of NVIDIA's new generation AI computing platform has officially begun. This is known as the fastest AI platform in the world. It has been delivered to a number of top cloud service providers for testing and verification, paving the way for the next stage of large model and Agentic AI training deployment.
According to reports, the Vera Rubin platform has entered the mass production stage, and the first batch of complete systems has been delivered to the computer rooms of major AI cloud vendors, including companies such as Oracle and CoreWeave. Prior to this, NVIDIA had delivered the first batch of Vera CPUs to a number of leading AI companies last month, laying the foundation for the ecology of this new platform.
Mahesh Thiagarajan, executive vice president of Oracle Cloud Infrastructure, first published physical photos of the Vera Rubin NVL72 rack system on social platforms, showing this huge rack cluster composed of 72 Rubin GPUs and 36 Vera CPUs. Officially positioning it as the current top GPU and CPU combination for AI workloads, it aims to become another milestone platform in Nvidia’s AI product line after Grace Blackwell.
Thiagarajan said that Oracle is one of the first cloud providers to introduce the Vera Rubin NVL72 system in the cloud and conduct verification testing. The two parties will work closely to provide enterprise customers with a new generation of accelerated computing capabilities in large-scale cloud environments. Judging from the on-site pictures, the NVL72 cabinet is large in size and has complex wiring and cooling structures, highlighting its design orientation for ultra-large-scale data centers.
In addition to Oracle, cloud computing company CoreWeave also announced the completion of the installation of the first Vera Rubin NVL72 system and released a complete video of the entire machine being unloaded from the transport vehicle to the computer room for deployment. The screen shows that the installation of a single cabinet requires the collaborative operation of 3 to 4 engineers. This reflects to a certain extent the "data center-level" specifications of the system in terms of volume and weight. It also implies that similar cabinets may be deployed in hundreds or even thousands in AI data centers in the future.
CoreWeave also emphasized that it was not only one of the first cloud service providers to introduce Vera Rubin NVL72, but also the first to complete joint debugging and verification of the full stack of software and hardware. For this purpose, the company has created its own software-defined liquid cooling system and unified cabinet control solutions, named Valvey and Racky respectively, which are used for liquid cooling management and unified scheduling of the entire cabinet Vera Rubin platform, and through "CoreWeave Mission Control" to achieve operational monitoring at the cabinet cluster level.
In the architectural view of "AI is a five-layer cake" proposed by NVIDIA CEO Jensen Huang, Vera Rubin NVL72 is only one layer of hardware. Behind it, it also relies on sophisticated infrastructure including power supply, cooling, interconnection, network, etc., as well as software stack support built over many years. With CUDA and the CUDA-X ecosystem for AI workloads as its core, NVIDIA integrates software and hardware, making it difficult for competing manufacturers to compete head-on in terms of ecological maturity and widespread adoption.
In terms of performance, reports indicate that in the Mix of Experts (MoE) model training scenario, the Vera Rubin platform can achieve the same training speed as the previous generation Blackwell system using only a quarter of the number of GPUs, while the cost per Token in the inference stage can be reduced to one-tenth of Blackwell. This means that when deploying large-scale models and Agentic AI at the factory level, Vera Rubin is expected to bring significant advantages in computing power density and energy efficiency, providing key support for cloud vendors to control costs and expand scale.
Currently, the Vera Rubin platform has entered full mass production, and NVIDIA plans to officially launch the first batch of customer-facing production running tasks in the third quarter of this year. With leading cloud service providers such as Oracle and CoreWeave taking the lead in completing launch and verification, the industry generally expects that Vera Rubin will soon become one of the core components of the new generation of AI data center infrastructure, promoting the implementation of "Agentic AI factories" and more complex large model applications.