---
format: "aidr-story-markdown/v1"
id: "6d92c7d7bc866c8862150e6f7996fbd4d69a175dbf7cb36514f234ae354abd0b"
canonical_url: "https://aidr.today/6d92c7d7?lang=en"
title: "Nvidia Rubin NVL72 Delivers 67x Throughput Per Cost in SemiAnalysis Test Beating Company Claims"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-09-16T23:56:17.000Z"
category: "Chips"
topics: ["nvidia","rubin-nvl72","gpu-performance","semianalysis","throughput-cost","rubin","nvl72","throughput"]
source_urls: ["https://marketbrief.now/ai/nvidia-rubin-nvl72-delivers-67x-throughput-per-cost-in-semianalysis-test-c0598942","https://huggingnews.com/ai/nvidia-rubin-nvl72-delivers-67x-throughput-per-cost-in-semianalysis-test-c0598942"]
summary: "SemiAnalysis testing found the Nvidia Rubin NVL72 delivers 67 times the total throughput per total cost of ownership compared to the GB300 Dynamo at 170 tokens per second, exceeding the 3x gain announced at GTC 2026. The system uses TRTLLM NVFP4 Dense and was evaluated with the open source InferenceX v3 tool. The platform's gains come from a co-design of six products including the Rubin GPU, Vera CPU, and NVLink 6 Switch. SemiAnalysis spent $3 million developing InferenceX v3 and the AgentX benchmark. Some critics say the 67x figure applies only to the 170 tokens per second setting."
---

# Nvidia Rubin NVL72 Delivers 67x Throughput Per Cost in SemiAnalysis Test Beating Company Claims

> [Open the canonical story](<https://aidr.today/6d92c7d7?lang=en>)

**Published:** 2026-09-16T23:56:17.000Z
**Category:** Chips
**Topics:** nvidia, rubin\-nvl72, gpu\-performance, semianalysis, throughput\-cost, rubin, nvl72, throughput

## Summary

SemiAnalysis testing found the Nvidia Rubin NVL72 delivers 67 times the total throughput per total cost of ownership compared to the GB300 Dynamo at 170 tokens per second, exceeding the 3x gain announced at GTC 2026\. The system uses TRTLLM NVFP4 Dense and was evaluated with the open source InferenceX v3 tool\. The platform's gains come from a co\-design of six products including the Rubin GPU, Vera CPU, and NVLink 6 Switch\. SemiAnalysis spent $3 million developing InferenceX v3 and the AgentX benchmark\. Some critics say the 67x figure applies only to the 170 tokens per second setting\.

## Sources

- [Story source](<https://marketbrief.now/ai/nvidia-rubin-nvl72-delivers-67x-throughput-per-cost-in-semianalysis-test-c0598942>)
- [Story source](<https://huggingnews.com/ai/nvidia-rubin-nvl72-delivers-67x-throughput-per-cost-in-semianalysis-test-c0598942>)

