NVIDIA introduced the Vera processor and Vera Rubin platform at GTC 2026, offering high-performance AI solutions with energy efficiency for training and inference.
© A. Krivonosov
At the GTC 2026 conference, NVIDIA unveiled the Vera processor and the Vera Rubin platform. This platform combines central and graphics processors, networking hardware, and data storage systems into modular racks designed for agent-based AI, reinforcement learning, and inference.
The Vera processor features 88 specialized NVIDIA Olympus cores with Spatial Multithreading technology, uses LPDDR5X memory, and delivers bandwidth of up to 1.2 Tb/s. When paired with Rubin graphics processors via the NVLink-C2C interface, it achieves coherent bandwidth of up to 1.8 Tb/s—surpassing PCIe Gen 6.
A new Vera rack can house up to 256 liquid-cooled processors, supporting over 22,500 simultaneous compute environments. The platform includes ConnectX SuperNIC network adapters and BlueField-4 graphics processors. The expanded Vera Rubin integrates NVL72 racks with 72 Rubin graphics processors and 36 Vera processors, BlueField-4 STX storage racks, Ethernet Spectrum-6 SPX racks, and LPX racks for inference.
These systems offer high energy efficiency and performance: training mixture-of-experts models requires four times fewer graphics processors than Blackwell, while inference per watt is increased tenfold with lower token costs. The platform supports DSX Max-Q and DSX Flex software for optimizing power distribution and enhancing deployment capabilities.
Vera and Vera Rubin are already backed by major cloud providers and AI labs, including Amazon Web Services, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, Alibaba, ByteDance, CoreWeave, Lambda, Nebius, OpenAI, and Anthropic. Server hardware suppliers include Dell, HPE, Lenovo, Supermicro, ASUS, Foxconn, GIGABYTE, QCT, Wistron, and Wiwynn. Production of Vera has begun, with partner shipments scheduled for the second half of 2026.