The first Vera Rubin samples are being delivered to hardware manufacturers and system integrators, allowing companies to begin optimizing designs around the new AI components ahead of mass production planned for the second half of 2026. The platform integrates a CPU and GPU built from the ground up for generative model workloads and AI agents that demand extremely high memory bandwidth and computing power.
A key component of the new architecture is the Rubin GPU, the successor to Blackwell, which uses HBM4 memory with capacities of up to 288 GB per chip – offering substantially higher bandwidth than Blackwell. Memory bandwidth can reportedly reach around 22 TB/s, roughly 2.5 times higher than Blackwell, which is critical for AI performance, especially with long context windows and memory-intensive models.
The new generation is not only about larger memory capacity. The Vera Rubin design tightly integrates the Vera CPU and Rubin GPU with the upgraded NVLink 6 interconnect, enabling significantly faster chip-to-chip communication compared to Blackwell. This approach is intended to support the creation of AI supercomputers at unprecedented scale and efficiency for both training and inference of generative models.
Compared with Blackwell, the Vera Rubin architecture is expected to deliver substantially higher performance and lower AI computation costs. Nvidia claims that the new platforms could reduce inference cost per million tokens by up to ten times compared with Blackwell, a crucial factor for operators running large AI models in the cloud.
In practical terms, data centers using Vera Rubin systems should be able to train and serve AI models with greater performance and energy efficiency, potentially lowering the cost of AI services and expanding the capabilities of advanced AI agents – from large language models to real-time applications.
For the broader market, this represents a significant step forward. Vera Rubin is positioned as the foundation of a new generation of AI hardware that surpasses Blackwell in both power and scalability, at a time when Blackwell is widely deployed across AI infrastructure operated by major technology companies.