# Vera Rubin resets math behind Kimi K3 benchmark throughput _NVIDIA swept MLPerf Training with Vera Rubin and claimed 10x agentic throughput per watt. Clouds cash in first._ **Published:** 2026-11-16 **Source:** https://www.startuphub.ai/ai-news/technology/2026/vera-rubin-resets-math-behind-kimi-k3-benchmark-throughput --- [NVIDIA](https://blogs.nvidia.com/blog/mlperf-training-endpoints-vera-rubin/) says Vera Rubin NVL72 swept every MLPerf Training v6.1 test in its debut. The sharpest claim is the Kimi K3 benchmark for agentic inference, at more than 10x throughput per watt over GB300 NVL72. NVIDIA also reports up to 2x more training performance per GPU than GB300 NVL72, per MLCommons entries cited in the announcement. Power, not raw tokens, is the constraint NVIDIA wants priced. On DeepSeek-V3, NVIDIA says Rubin scaled from 256 to 1,024 GPUs at about 97 percent efficiency, cutting training time from 16 minutes to four. NVIDIA also ran GB300 systems at 10,240 GPUs, with Microsoft Azure matching that scale on GB200 systems, linked by NVLink scale up plus Quantum-X800 InfiniBand and Spectrum-X Ethernet. For Kimi K3, NVIDIA cites 41,414.2 tokens per second at 100.89 kilowatts on 32 Rubin accelerators versus 2,210.4 tokens per second at 60 kilowatts on 48 GB300 accelerators. The comparison holds end to end interactivity above 50 tokens per second per user, per the MLPerf Endpoints disclosure. For context, BenchLM lists Kimi K3 at rank 17 of 215 with a score of 68.42 from 76 displayable rows, while AI Agent Store reports 79.2 on LiveBench with reasoning at 90.7 and agentic coding at 62.2. StartupHub.ai data shows [CoreWeave](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/coreweave-6-billion-lancaster-data-center) and Lambda positioned as NVIDIA capacity channels rather than alternative silicon suppliers. That matters because 20 partners filed Training results on NVIDIA hardware, per NVIDIA, so a Rubin efficiency lead lifts incumbents first and squeezes alternative accelerator pitches on power cost. The Rubin Kimi K3 result is provisional pending MLCommons verification, and independent cloud pricing would change the read if performance per watt does not survive multi tenant overhead. Can any rival answer Rubin on tokens per watt without forcing customers to rewrite CUDA tuned pipelines? --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.