At today's GTC 2026 conference, NVIDIA took the wraps off its next-generation Vera Rubin platform. This platform stands as the most formidable AI computing platform to date and is set to be a flagship product for shipments this year. The Vera Rubin platform isn't a new concept; its core architecture is ingeniously crafted from seven distinct chips. These include the powerful Vera CPU, the high-performance Rubin GPU, the NVLink 6 Switch for seamless data transfer, the ConnectX-9 SuperNIC for enhanced network connectivity, the versatile BlueField-4 DPU, the Spectrum-6 Ethernet Switch for robust networking, and the cutting-edge Groq 3 LPU inference chip.

Through a meticulously designed "six-chip collaboration" approach (with certain configurations integrating all seven chips), the platform constructs a highly modular AI supercomputing system. Take, for instance, a typical configuration like the Vera Rubin NVL72 rack. It boasts 72 Rubin GPUs, 36 Vera CPUs, and 18 BlueField-4 DPUs, delivering an astonishing 3.6 EFLOPS of FP4 inference performance and 2.5 EFLOPS of FP8 training performance. When compared to its predecessor, the Blackwell platform, it achieves a remarkable 5-fold improvement in inference performance. Moreover, it slashes the cost of generating a single token to a mere one-tenth and offers a system memory bandwidth of a staggering 22TB/s. For the first time, it supports a massive 150TB shared contextual memory pool, effectively breaking through memory bottlenecks in long-inference scenarios.

The platform leverages TSMC's advanced 3nm process and CoWoS-L packaging technology, seamlessly integrating HBM4 high-bandwidth memory. This results in a single-card memory capacity of 288GB and a bandwidth that's 2.75 times that of HBM3e. Additionally, the Vera Rubin platform significantly reduces data center power consumption by 6% through the implementation of warm-water cooling technology. It also incorporates full-link hardware encryption, providing robust security assurance for high-privacy scenarios such as finance and healthcare.

The platform has already garnered advance orders from leading clients, including OpenAI and Google DeepMind. Deployment instances are expected to roll out through cloud service providers like Amazon AWS and Microsoft Azure in the second half of 2026, further accelerating the industrialization of AI.