During his GTC Taipei 2026 keynote, Jensen Huang reframed the design rationale for the Vera Rubin platform around the computational demands of AI agents. “Vera Rubin is not a chip, not just a GPU — it’s a complete end-to-end system,” he emphasized. He called it the most ambitious project in Nvidia’s history, with 40,000 engineers company-wide involved.
Huang explained why such a system is necessary by breaking down how agents work. An agent is essentially a distributed heterogeneous computing model: the large language model (the brain) thinks on GPUs, with every inference lighting up an entire rack of Grace Blackwell NVLink 72s; the orchestration engine (the body) schedules the full workflow on CPUs; tool calling (the workshop) engages both CPUs and GPUs; and the security layer runs on Nvidia’s BlueField DPU, providing encryption at rest, in transit, and in use under confidential computing standards. The memory system spans KV cache, data compression, and retrieval, which he predicted would completely revolutionize the storage industry.
Vera Rubin’s full stack includes: Rubin GPU with NVLink 6 interconnect, Vera CPU, ConnectX-9 SuperNIC, BlueField-4 DPU security processor, and the DOCA software stack. Huang also announced that all CUDA X libraries will be equipped with “agent skills,” allowing agents to learn how to invoke these libraries by reading instructions like a manual. He claimed that in the future, agents will use CUDA X far more efficiently than humans.
Huang stated that Nvidia has evolved from a GPU company to a systems company, and is now transforming again: customers don’t want to buy computers; they want to build AI factories. Nvidia’s ecosystem has expanded to include power supply, cooling systems, electrical grids, and other industrial infrastructure, with the goal of delivering a complete full-stack system so customers can build AI infrastructure directly.