NVIDIA has reportedly upgraded its upcoming Vera Rubin superchip multiple times since March, targeting 22 TB/s of memory bandwidth for its VR200 NVL72 system. This marks a significant jump from the initial 13 TB/s goal, driven by faster DRAM and improved interconnects. The move directly counters AMD's Instinct MI400 series, which promises 19.6 TB/s bandwidth with its new HBM4 memory.

The performance arms race reflects intense competition in the AI accelerator market. AMD's MI455X accelerator is slated to offer 40 PFLOPS of FP4 compute and 432 GB of HBM4 capacity, claiming a 1.5x advantage in scale-out bandwidth. NVIDIA's revised specs aim to close that gap, with the full VR200 system delivering roughly 100 PFLOPS of FP4 performance across its two-GPU design.

First shipments of Vera Rubin systems are expected in late summer 2026, setting the stage for a major showdown. Hyperscalers and data centers will now weigh NVIDIA's bandwidth gains against AMD's claimed capacity and efficiency wins. The outcome will shape the infrastructure powering the next wave of large language models and AI training workloads.