Nvidia quietly lowers some speed specs for its next big AI chip, Rubin
Memory and chip-to-chip speeds dropped on Nvidia's spec page, while computing power stayed the same.
Santa Clara-based Nvidia has quietly lowered some performance figures for Rubin, its next major AI chip, ServeTheHome, a tech publication, reported Saturday. Nvidia's own product page now shows the new numbers.
Each Rubin chip's memory bandwidth, how fast it can pull data from its built-in high-speed memory, dropped from 22 to 19.2 terabytes per second. NVLink, the link that lets 72 chips in one server rack work together like one giant chip, dropped from 3.6 to 3.0 terabytes per second. The networking that connects separate racks got faster, rising from 400 to 450 gigabytes per second.
Raw computing power didn't change, and neither did the amount of memory.
Why it matters
AI models constantly shuttle huge amounts of data in and out of memory, so memory speed often limits how fast a chatbot can answer. Slower memory could mean somewhat lower performance than buyers expected. Nvidia had labeled the earlier figures preliminary and has not said why they changed.
The change comes as big cloud companies begin offering the first Rubin systems. Chinese site IT Home first spotted the revision, ServeTheHome said.
Source: ServeTheHome, Oct. 31, 2026 and Nvidia Vera Rubin NVL72 product page.