Nvidia Vera Rubin Enters Heavy Rack Shipment Phase as Server Makers Add Capacity

Nvidia's AI platform roadmap is shifting decisively to rack-scale systems, with supply chain sources saying the company will enter a heavy shipment phase for Vera Rubin in the second half of 2026. The rollout spans the Vera CPU, the Rubin GPU, and rack platforms including NVL4, NVL8 and NVL72. That reframes the competitive unit from individual accelerators to whole racks, which pulls power delivery, liquid cooling, interconnect and rack hardware suppliers directly into the critical path alongside the silicon itself.

The rack is now a large, power-dense product in its own right. A Vera Rubin NVL72 configuration pairs 72 Rubin GPUs with 36 Vera CPUs across 18 1U compute trays, linked by nine sixth-generation NVLink switch trays for 216 TB/s of scale-up bandwidth, with 20.7 TB of HBM4 and up to 54 TB of LPDDR5X per rack. Supermicro said in a 23 September announcement that it had begun shipping the racks from expanded US manufacturing in San Jose, and Dell said in early September that it had delivered the first production Vera Rubin NVL72 racks to CoreWeave. Deployment blueprints start at a 5 MW power envelope and the racks are liquid-cooled from the chip outward, so for many buyers the building, its power delivery and its heat rejection are the binding constraint rather than GPU availability.

Demand signals further down the chain are consistent with a rack-driven cycle. Sharp began taking orders in Japan on 15 September for AI servers procured from parent Foxconn and equipped with Nvidia processors, targeting JPY250 billion of revenue by fiscal 2030 and giving the display maker a route into domestic AI infrastructure. Separately, Taiwan-listed AI supply chain companies reported August revenue that DIGITIMES describes as still accelerating, with King Slide Works posting the fastest growth among 50 tracked companies at 318% year on year as the growth driver moves from cooling to rack hardware. Together these point to the rack rather than the chip becoming the unit that gets billed and built through 2027.