NVIDIA Shows Vera Rubin Silicon: Up to 30x Agent Throughput over GB300

NVIDIA published the first silicon-based performance results for Vera Rubin NVL72, claiming order-of-magnitude gains over GB300 NVL72 on real AI Agent workloads — the architecture's first official real-silicon validation aimed at the Agent era.

Confirmed

The test was published on NVIDIA's technical blog, using SemiAnalysis's AgentX workload and the DeepSeek V4 Pro model against real agent traces.
At matched interactivity targets, Vera Rubin delivers up to 30x higher throughput per megawatt and 35x lower agentic coding token cost than GB300 NVL72.
The benchmark replays production-grade coding agent sessions, including long context and KV-cache reuse scenarios.
The news was first posted on 08-24 by @nvidia and @nordicinst, then spread via reposts from @Scobleizer and @brianryhuang.

Why it matters

This is NVIDIA's first demonstration of Rubin's capabilities from real silicon rather than paper specs, targeting AI agents — an emerging mainstream workload — instead of traditional training/inference benchmarks.
Throughput-per-megawatt and token cost speak directly to data-center efficiency and unit economics of intelligence, offering a concrete reference for cloud providers and agent teams planning next-generation compute.
As @Crescitaly notes, the figures are NVIDIA's own claims from its blog with no independent third-party replication yet; readers should mind the test conditions such as "matched interactivity targets."

Unconfirmed

No independent replication exists yet; the 30x and 35x figures rest solely on NVIDIA's single-source methodology.

2026-08-24 ~ 2026-08-26 · 5 related posts

Full story(2 episodes)→

Episode 1: NVIDIA Shows Vera Rubin Silicon: Up to 30x Agent Throughput over GB300(2026-08-24, 5 posts)
Episode 2: NVIDIA Claims Rubin NVL72 Delivers 30x Better Efficiency on AgentX Benchmark(2026-08-29, 2 posts)

Primary sources

NVIDIA Vera Rubin NVL72 delivers 30x higher efficiency for AI agents vs GB300 — nvidia · 2026-08-24 23:02 (UTC+08:00)
NVIDIA claims up to 30× agentic throughput per MW on Vera Rubin — Crescitaly · 2026-08-26 16:45 (UTC+08:00)
Nvidia Vera Rubin NVL72 delivers 30x more work per watt than GB300 on agentic workloads — nordicinst · 2026-08-24 23:19 (UTC+08:00)

[source] NVIDIA Vera Rubin NVL72 delivers 30x higher efficiency for AI agents vs GB300 — nvidia · 2026-08-24
[source] Nvidia Vera Rubin NVL72 delivers 30x more work per watt than GB300 on agentic workloads — nordicinst · 2026-08-24
NVIDIA Vera Rubin benchmarks show 35x cheaper agentic coding tokens — brianryhuang · 2026-08-25

2 near-duplicate retellings: Scobleizer · Crescitaly