The Hardware Drop: Vera Rubin Is Not a GPU. It's a Factory.
Jensen Huang walked onto the GTC 2026 stage and opened with a concept that set the tone for everything that followed. "Your data center used to be a data center for files," he told the crowd. "It's now a factory for tokens." [1] That reframe isn't marketing. It describes how NVIDIA sees its own business, and the Vera Rubin platform announcement makes the architecture explicit. Vera Rubin isn't just a GPU — it's seven chips and five rack-scale systems rolled into a single platform, architected specifically for agentic AI. The headline stat: 3.6 exaflops of compute, with 260 terabytes per second of all-to-all NVLink bandwidth. [1] The platform includes the new Vera CPU — built around LPDDR5 for extreme single-threaded performance and energy efficiency — alongside the Groq 3 LPU rack, Spectrum X co-packaged optics switches, and Rubin Ultra, which scales to 144 GPUs in a single NVLink domain. Previous Rubin systems slid in horizontally. Rubin Ultra installs vertically into the Kyber rack, with compute in the front and NVLink switches in the back, forming what Huang called "one giant computer." [1] The manufacturing improvement is notable: what used to take two days to install now takes two hours. 100% liquid cooling at 45°C. All cables gone. This is what you build when you're designing for a world where hyperscalers are ordering infrastructure at the scale of small cities.






