At today's Consumer Electronics Show (CES), NVIDIA CEO Jensen Huang officially announced the company's new Rubin computing architecture, calling it the cutting-edge of artificial intelligence hardware. The Rubin architecture has already entered production and is expected to accelerate its rollout later this year.

Jensen Huang said, "Vera Rubin is designed to address a fundamental challenge: the explosive growth in computational requirements for AI." He added, "Today I can tell you that Vera Rubin is now fully in production."

The Rubin architecture was first announced in 2024, marking the latest achievement in NVIDIA's ongoing efforts to advance hardware development, which has made NVIDIA the world's most valuable company. The Rubin architecture will replace the previous Blackwell architecture, which itself replaced the Hopper and Lovelace architectures.

Rubin chips have been adopted by several major cloud service providers, including key collaborations between NVIDIA and companies such as Anthropic, OpenAI, and Amazon Web Services (AWS). The Rubin system will also be used in HPE's Blue Lion supercomputer and the upcoming Doudna supercomputer at Lawrence Berkeley National Laboratory.

The Rubin architecture is named after astronomer Vera Florence Cooper Rubin and consists of six independent chips working together, with the Rubin GPU as its core. The architecture also improves on bottlenecks in storage and interconnects, adding the Bluefield and NVLink systems. In addition, the Rubin architecture includes a new Vera CPU designed specifically for autonomous reasoning.

Dion Harris, Senior Director of NVIDIA's AI Infrastructure Solutions, explained the advantages of the new storage during a conference call. He mentioned that the memory requirements related to cache in modern AI systems are increasing. With the introduction of new workflows, such as autonomous AI or long-term tasks, higher demands are placed on key-value cache (KV cache). He stated, "We introduced a new storage layer that connects outside of computing devices, allowing for more efficient expansion of the storage pool."

According to NVIDIA's tests, the Rubin architecture is 3.5 times faster than the previous Blackwell architecture in model training tasks and five times faster in inference tasks, with a computational power of up to 50 petaflops. The new platform also improves inference computing capacity per watt by eight times.

The launch of the Rubin architecture comes at a time of intensified competition in AI infrastructure, as AI labs and cloud service providers race to purchase NVIDIA's chips and the required computing facilities. During a earnings call in October 2025, Huang predicted that between $300 billion and $400 billion would be invested in AI infrastructure over the next five years.

Key Points:

🌟 The Rubin architecture is NVIDIA's latest AI computing architecture, expected to significantly enhance computing power.

💻 Rubin chips have been adopted by multiple cloud service providers and will be used in several supercomputer projects.

⚡ The Rubin architecture shows significant improvements in speed and energy efficiency compared to its predecessor, bringing new momentum to the AI infrastructure competition.