NVIDIA Rubin Ultra Specs Unfinalized Amid Supply Constraints, Standard Version Delivered

Standard Vera Rubin NVL72 delivered to dozens of clients including Microsoft and OpenAI. Rubin Ultra specs remain undecided due to TSMC yield issues, shifting focus from aggressive capacity to pragmatic shipment volume.

Woofun AI reports that the standard Vera Rubin NVL72 has been delivered to dozens of customers, including Microsoft, OpenAI, and Google Cloud, with large-scale releases scheduled for fall. Dell completed the first full-scale boot test in July, reducing rack assembly time to approximately five minutes.

Conversely, specifications for the Rubin Ultra remain unfinalized due to TSMC CoWoS-L substrate warpage and yield challenges. The original four-chip design with 1TB memory is reportedly scrapped in favor of a dual-chip configuration with 384GB capacity. NVIDIA is prioritizing shipment volume and I/O speed over aggressive memory targets, with final specs expected after validation in the second half of 2026.