Is Nvidia's Vera Rubin Really 10x More Efficient? The First Independent Test

Nvidia says its new Vera Rubin rack delivers up to 10x more AI output per megawatt. On October 15, 2026, the first independent results from Artificial Analysis found anywhere from about 1.14x to 5.7x more than Blackwell Ultra (GB300). Is Vera Rubin really 10x more efficient, and how can one test produce both numbers? This video explains the trade-off between total throughput and speed per user, why the new chip pulls ahead mostly where answers must be fast, what baseline Nvidia's 10x uses, what a chip-only power measurement leaves out, and the caveats the testers published, from time to first token to early software. It ends with Microsoft's own tokens-per-watt claims for Maia. Chapters 00:00 Two results from one test 02:27 Tokens, throughput and speed per user 03:43 How the test worked 04:53 1.14x at peak, 5.7x at speed 06:39 Why memory matters 07:46 Where Nvidia's 10x comes from 09:17 Where the line gets drawn 11:45 The caveats 14:12 Why it matters for AI agents 15:48 Is it really 10x? Sources: • Artificial Analysis — First AA-AgentPerf results for NVIDIA Vera Rubin NVL72 (2026): https://artificialanalysis.ai/article... • NVIDIA Technical Blog — Scaling Token Factory Revenue and AI Efficiency by Maximizing Performance per Watt (2026): https://developer.nvidia.com/blog/sca... • NVIDIA Blog — NVIDIA Vera Rubin Driving Performance Per Watt, Lowest Token Cost for Partners Worldwide (2026): https://blogs.nvidia.com/blog/2026/03... • Microsoft Azure Blog — Infrastructure for the frontier era: Scaling intelligence through open industry collaboration (2026): https://azure.microsoft.com/en-us/blo... Research as of 2026-10-18. The highway comparison is a simplified illustration. Charts are redrawn from the published results. Results are first measurements on early software and may change. #Nvidia #VeraRubin #AIChips #AINews