The chip that currently powers almost every frontier AI model — Nvidia's Blackwell — already has a successor in production. Vera Rubin, the most complex graphics-computing design Nvidia has ever shipped, is scaling up fast enough to reshape how, where and at what power level the world's largest AI factories are being built.
Rubin is not just a new graphics processor with a higher clock. Nvidia designed it as an integrated system that fuses Rubin GPUs, a new Vera CPU, Spectrum-X Ethernet networking, and BlueField data-processing units into a single fabric. The idea is to reduce every millisecond of friction between the parts that compute, route and store — a tight coupling that becomes critical when an AI factory has to coordinate tens of thousands of accelerators at once.
Industry analysts project several million Rubin accelerators reaching customers in 2026, with full production already underway and partner availability ramping through the back half of the year. That scale is the reason Rubin is now the reference architecture behind headline AI builds around the world: a Japanese consortium is deploying a Rubin-based AI factory rated at 140 megawatts, built from roughly 27,500 Rubin GPUs arranged in NVL72 racks that each house 72 GPUs alongside Vera CPUs.
The gains come with a thermal cost. A single Blackwell NVL72 system already pulls about 120 kilowatts; successor Rubin-class systems are estimated to climb toward 135 to 140 kilowatts per rack. That puts the chip in direct conversation with another frontier: the cooling, power-grid and water infrastructure that must be built to keep the silicon running.