NVIDIA’s 1.4 Exaflop Chip

An exaflop is a measure of how many math operations a computer can do in a second, and NVIDIA says its new Vera Rubin system reaches that scale in a single rack. (blogs.nvidia.com) Jensen Huang introduced the Rubin platform at CES on January 5, 2026, then NVIDIA said on March 16, 2026 that Vera Rubin was in full production. The company describes it as a rack-scale system built from Rubin graphics processors, Vera central processors, networking chips, storage hardware, and software designed together. (blogs.nvidia.com) (nvidianews.nvidia.com) The “supercomputer room” line refers to density, not a consumer chip. NVIDIA’s Vera Rubin NVL72 product page says one rack combines 72 Rubin graphics processors and 36 Vera central processors into what the company calls “one AI supercomputer.” (nvidia.com) NVIDIA has used several different performance numbers for Rubin systems, and the number depends on the exact product and math format. In September 2025, NVIDIA said the Vera Rubin NVL144 CPX platform would deliver 8 exaflops of AI compute in one rack, while its CES material said a Rubin graphics processor delivers 50 petaflops of NVFP4 inference. (nvidianews.nvidia.com) (blogs.nvidia.com) That is why clips claiming “a 1.4 exaflop chip” compress several layers of hardware into one headline. NVIDIA’s official materials describe Rubin as a platform and a family of systems, with performance figures attached to racks, trays, and individual processors depending on configuration. (developer.nvidia.com) (nvidia.com) The comparison point is NVIDIA’s smaller desktop AI machine, DGX Spark. NVIDIA says DGX Spark uses a GB10 Grace Blackwell superchip and delivers up to 1 petaflop of FP4 AI performance with 128 gigabytes of memory in a desktop form factor. (nvidia.com) A petaflop is one-thousandth of an exaflop, so moving from a 1-petaflop desktop box to rack systems measured in exaflops shows how much of the current AI race is about packing more compute, memory, and networking into tighter data-center footprints. (blogs.nvidia.com) (nvidia.com) NVIDIA is also tying Rubin to “agentic” and long-context AI workloads, which means systems built to keep more data in fast memory while models reason across longer prompts and more steps. Its March 2026 announcement said the platform was optimized for pretraining, post-training, test-time scaling, and real-time inference. (nvidianews.nvidia.com) So the cleanest way to read the viral claim is this: NVIDIA is selling ever-denser AI systems, and its official Rubin roadmap now centers on rack-scale machines that the company says can deliver exaflop-class performance in a footprint far smaller than older supercomputing installations. (developer.nvidia.com) (nvidianews.nvidia.com)