How much faster is Rubin NVL72 than GB300 Blackwell?

On SemiAnalysis's AgentX agentic-inference benchmark, published September 15, 2026, Vera Rubin NVL72 served DeepSeek V4 Pro (1.6T parameters) at 59.4 million tokens per second per megawatt at 100 tokens/second interactivity versus 28.5 million for GB300 NVL72, a 2.1x gain. At 150 TPS the gap widened to 7.2x, and at 170 TPS SemiAnalysis measured roughly 67x the throughput per dollar of TCO because GB300 collapses at that interactivity. In the 60-100 TPS range where most providers serve, Rubin delivered 1.4x to 3x the throughput per TCO.

Did Rubin beat NVIDIA's own claim?

Yes. At GTC 2026 Jensen Huang showed Rubin NVL72 at 3x Blackwell's performance per megawatt on 1-3 trillion-parameter models around 200 TPS. SemiAnalysis's pre-release measurement found up to 7x on agentic traffic and wrote that 'Jensen needs to stop sandbagging.' The same thing happened in 2024, when NVIDIA claimed 30x over Hopper for GB200 NVL72 and SemiAnalysis measured 98x.

Should I wait for Rubin or buy GB300 now?

If your workload is agentic (multi-turn, long context, high prefix reuse, sub-agent bursts) and you can wait for Rubin volume, Rubin's modeled profit per gigawatt is about $149.9 billion versus $105.3 billion for GB300 on SemiAnalysis's TCO model, roughly 2x more profit per gigawatt after costs. If you need capacity in 2026 or run chat-style traffic at 60-100 TPS, GB300 remains competitive at 1.4x to 3x less throughput per dollar, and open-source SGLang closes much of the interactivity gap.

What are the caveats in the Rubin numbers?

They come from pre-release hardware and early TensorRT-LLM software, with NVIDIA engineers helping bring up the stack; they were measured on one model (DeepSeek V4 Pro) in one scenario (AgentX); the production rack uses a 2300W GPU TDP and 1.5TB of LPDDR5X per compute tray after NVIDIA halved Vera's CPU memory; and the TCO conclusions depend on SemiAnalysis's cost model. Independent production results from cloud providers do not exist yet.

Quick Answer

NVIDIA Rubin NVL72 vs GB300 Blackwell for Agentic Inference

Published: September 16, 2026

The short answer

Rubin NVL72 is 2.1x GB300 at normal interactivity and up to 7.2x at high interactivity on agentic inference, according to the first verified numbers, published by SemiAnalysis on September 15, 2026. That is well above NVIDIA’s own 3x claim from GTC 2026. The catch is that these are pre-release results on one model with vendor help, and at the 60-100 tokens-per-second range most providers actually serve, the advantage is a still-large but less dramatic 1.4x to 3x in throughput per dollar.

	Vera Rubin NVL72	GB300 NVL72 (Blackwell Ultra)
Tokens/s per MW at 100 TPS (DeepSeek V4 Pro, AgentX)	59.4M	28.5M
Advantage at 100 TPS	2.1x	–
Advantage at 150 TPS	7.2x	–
Throughput per $ TCO at 170 TPS	~67x (TRT-LLM NVFP4 dense vs GB300 Dynamo TRT-LLM)	–
Throughput per $ TCO at 60-100 TPS	1.4x-3x	–
Max P90 interactivity (TRT-LLM)	276 TPS	172 TPS (SGLang narrows this)
Modeled profit per GW per year	~$149.9B	~$105.3B (Dynamo SGLang)
Software state	Pre-release TensorRT-LLM	Mature TRT-LLM, Dynamo, SGLang, vLLM
Platform	Rubin GPU, Vera CPU, NVLink 6 Switch, ConnectX-9, BlueField-4, Spectrum-6	Blackwell Ultra GPU, Grace CPU, NVLink 5, ConnectX-8, BlueField-3, Spectrum-4/X
Compute tray	2300W GPU TDP, 1.5TB LPDDR5X	1400W GPU TDP class, Grace LPDDR5X
Availability	Pre-release; volume ramp expected into 2027	Shipping since 2025

What was measured

SemiAnalysis ran its AgentX scenario, the agentic-inference benchmark it co-published with vLLM on September 8, 2026 and now uses across its InferenceX fleet (NVIDIA, AMD, TPUv7; SambaNova and Trainium pending). AgentX replays real agent traffic with four traits that stress hardware differently from chatbot benchmarks:

Multi-turn sessions with tens or hundreds of turns and many tool calls.
Long context that accumulates system prompts, tool schemas and prior turns.
High prefix reuse, so most input can be served from KV cache rather than recomputed, if there is enough memory to hold it.
Sub-agent bursts, short-lived agents with fresh context that create spiky KV-cache demand.

The model was DeepSeek V4 Pro, a 1.6-trillion-parameter MoE, served with NVFP4 on TensorRT-LLM. That choice matters: MoE plus long shared prefixes is exactly the workload where Rubin’s larger memory system and NVLink 6 fabric pay off.

Why Rubin wins by more at higher interactivity

Interactivity (tokens per second per user) is the axis. At low TPS, both racks are throughput-bound and Rubin’s advantage is the plain generational one, about 2x. As the target climbs to 150-170 TPS, GB300 runs out of headroom and its total throughput falls off a cliff while Rubin keeps serving; that is how a 2.1x gap at 100 TPS becomes 7.2x at 150 TPS and the ~67x per-dollar figure at 170 TPS. Rubin’s maximum P90 interactivity on TRT-LLM was 276 TPS versus 172 for GB300, about 61% higher, though SemiAnalysis notes GB300 on open-source SGLang reaches similar interactivity to Rubin.

For buyers this means the question is not “how much faster is Rubin” but “what interactivity do my agents need.” Coding agents that stream long tool outputs to a human want 100+ TPS; batch agent fleets do not.

The economics

SemiAnalysis’s AI Cloud TCO Model normalizes throughput by all-in cost per GPU-hour (capex over useful life, colocation, power, cost of capital) under two scenarios: owning at hyperscaler volume, or renting on a three-year commit. On owning assumptions, Rubin delivers roughly 2x the profit per gigawatt of Blackwell even on early software, about $149.9 billion versus $105.3 billion per GW per year in the modeled scenario. Because Rubin racks cost more per chip, the per-dollar advantage in the mainstream 60-100 TPS band is smaller than the per-megawatt one: 1.4x to 3x.

Caveats before you change a purchase order

Pre-release everything. Hardware was unreleased, TRT-LLM for Rubin was an early build, and NVIDIA’s own engineers (named in the article) helped with bring-up. SemiAnalysis expects the gap to widen as kernels mature, but the current numbers are directional.
One model, one scenario. No dense-model, no small-model and no training numbers.
Vera memory was halved. The production compute tray carries 1.5TB of LPDDR5X per tray after NVIDIA cut the CPU memory, which affects how much KV cache can spill to host memory in long agent sessions.
TCO assumptions are SemiAnalysis’s. Your power price, colocation and depreciation schedule change the per-dollar ratios.
Competitor racks are not in this cut. AMD’s MI455X UALoE72 has committed to AgentX but has no published result yet; TPUv7 numbers exist on InferenceX for other scenarios.

Bottom line by buyer

You are…	Do
A neocloud planning 2027 capacity for agent workloads	Plan Rubin; per-GW economics are ~2x even on early software
Serving chat/completions at 60-100 TPS today	GB300 remains a sound buy; 1.4x-3x is not enough to strand it
Running latency-sensitive coding agents at 150+ TPS	Rubin changes what is possible; GB300 collapses there
An inference API customer	Expect Rubin to pressure per-token prices further in 2027, on top of the 2026 cuts