Overview
- NVIDIA announced the Vera Rubin rack-scale platform on Monday, June 22, 2026, presenting a system that combines native FP64 double-precision simulation with the company’s full AI software stack.
- Supermicro published a Data Center Building Block Solutions (DCBBS) blueprint the same day that details a 3.2 MW scalable unit able to host up to 1,152 Rubin GPUs and 576 Vera CPUs in direct liquid-cooled racks.
- Vera Rubin’s architecture uses direct liquid cooling plus high-bandwidth interconnects — NVLink‑C2C, ConnectX‑9 SuperNICs and BlueField‑4 DPUs — to enable much higher power density and GPU counts per rack than air-cooled designs.
- Major system builders and labs are lining up: Supermicro, Dell, HPE, GIGABYTE and Bull have certified designs, while LRZ, NERSC/LBNL (Doudna) and LANL have chosen Rubin-based systems, with vendors targeting Q4 2026 availability as a forward-looking timeline.
- Markets reacted quickly: Supermicro shares jumped about 15.7% after its blueprint release, underscoring how turnkey rack designs, liquid-cooling integration and site services could decide which vendors win research and procurement contracts and how fast scientists can use converged FP64 and AI workflows.