NVIDIA Vera Rubin NVL72 reaches production: 10x more energy efficiency for AI in scale

NVIDIA has just put into production the Vera Rubin NVL72 platform, its new bet to reduce the cost by token and bypass the biggest nightmare of AI data centers: energy consumption. The first tests with partners like CoreWeave, Google Cloud, Microsoft Azure and Oracle show a 10x jump throughput by megawatt compared to Grace Blackwell NVL72 — a metric that in practice means more inference by the same watt.

What makes NVIDIA Vera Rubin NVL72 so efficient?

The secret lies in the integrated design of the platform. By connecting 72 GPUs to a single memory and computing topology, NVIDIA eliminates communication bottlenecks and reduces latency in mass model training. With Vera Rubin architecture, each watt is converted into more mathematical operations per second, which directly drops the cost per generated token — whether for inference or fine-tuning. The company highlights that energy efficiency has become the limiting factor for the expansion of AI plants, and the new platform has been designed to maximize performance by megawatt.

Why is energy efficiency the new battlefield in AI?

AI data centers already consume the equivalent of small cities. With the demand for ever-increasing models, the cost of electricity and supply restrictions are curbing the ability to scale. By offering 10x more throughput by megawatt, Vera Rubin allows operators like CoreWeave and Microsoft Azure to double their computational capacity without expanding the electrical infrastructure — or cut the electricity bill in half to maintain the same volume of requests. This is particularly critical in regions where access to energy is limited or expensive.

Who's already running Vera Rubin NVL72?

In addition to Google Cloud and Oracle, CoreWeave — a cloud specialist for AI — has already implemented the platform into its clusters. Microsoft, which has been boosting its AI infrastructure with partners such as AMD, is also on the list of early adopters. The expectation is that broad commercial availability occurs over the next quarter, with the first customer workloads starting to migrate later this semester.

Frequently Asked Questions about Vera Rubin NVL72

When will Vera Rubin be available for joint ventures?

The platform is already in production and being used by large cloud providers. The expectation is that the expanded commercial offer will reach the market in the third quarter of 2025, with open reserves for partners.

Which worksloads benefit more from 10x gain per megawatt?

Large-scale language models, recommendation systems and computer vision applications that require continuous inference — scenarios where energy consumption is the main cost bottleneck.

How does Vera Rubin compare to Grace Blackwell in terms of cost per token?

The 10x increase in throughput by megawatt means that for a single power account, Vera Rubin can generate up to 10 times more tokens, drastically reducing the cost by inference. For AI plants with power constraints, this represents an immediate real economy.

Read the official announcement on the NVIDIA blog and check how other companies are optimizing their data centers, such as Microsoft Azure Turbine with AMD Helios and EPYC.