Huang Renxun Explains Why Vera Rubin Was Born: The Agent Era Requires a Whole New Heterogeneous Computing Base

According to Omdena Beating monitoring, Huang Renxun, in his keynote at GTC Taipei 2026, reinterpreted the design logic of the Vera Rubin platform around Agent's computing requirements, emphasizing that "Vera Rubin is not just a chip, not just a GPU, but a whole end-to-end designed system." He referred to it as NVIDIA's most ambitious project in history, with 40,000 engineers across the company involved in its development.

Huang Renxun explained the need for such a system based on the operation of Agent. Agent is essentially an aggregated and distributed heterogeneous computing model: the large language model (brain) thinks on the GPU, with each inference activating a full row of Grace Blackwell NVLink 72; the orchestration engine (body) schedules the entire process on the CPU; tool calls (workshops) simultaneously utilize both the CPU and GPU; the security layer runs on the NVIDIA BlueField DPU, achieving end-to-end encryption at rest, in transit, and in use, following the confidential computing standard. The memory system includes KV caching, data compression and retrieval, and he predicted that this would completely revolutionize the storage industry.

The Vera Rubin full stack includes: Rubin GPU + NVLink 6 interconnect, Vera CPU, ConnectX-9 SuperNIC, BlueField-4 DPU secure processor, and the DOCA software stack. Huang Renxun also announced that all CUDA X libraries will come equipped with "skills" (Agent Skills), enabling the Agent to learn to call these libraries like reading a manual, significantly surpassing human efficiency in using CUDA X in the future.

Huang Renxun stated that NVIDIA has evolved from a GPU company to a systems company and is now transforming once again: customers do not want to buy a computer but to build an AI factory. NVIDIA's ecosystem has expanded to include industrial infrastructure such as power supply, cooling systems, and the power grid. The goal is to deliver complete full-stack systems, allowing customers to directly construct AI infrastructure.