NVIDIA Vera Rubin NVL72: 30× More Agentic AI Throughput, According to NVIDIA

NVIDIA is positioning Vera Rubin NVL72 as an AI infrastructure platform built for workloads in which agents repeatedly generate tokens, call tools, execute code and create more context. In results published by NVIDIA, the system delivered up to 30× higher throughput per megawatt than GB300 NVL72 and reduced token cost by up to 35× on an AgentX test using DeepSeek V4 Pro.

Those are substantial claims, but they need to be read carefully: the supplied evidence does not provide a specific dollar cost per million tokens for Vera Rubin, nor does it establish an independent, apples-to-apples verification of the headline gains.

What NVIDIA claims Vera Rubin achieved

The headline comparison is:

• Up to 30× higher agentic-AI throughput per megawatt than GB300 NVL72.
• Up to 35× lower cost per token—or per million tokens when expressed at that scale— than GB300 NVL72.
• Testing with DeepSeek V4 Pro on the SemiAnalysis AgentX workload.

The test is designed around agentic coding rather than a short, isolated question-and-answer exchange. AgentX replays production-style coding sessions that include context growth, tool calls and subagent spawning. 6

That distinction matters. An agent may generate an answer, inspect the result, call an external tool, execute code, create another task and continue reasoning with a much larger context. The infrastructure must therefore process more than the model’s initial response, increasing pressure on memory, networking, CPUs and power as well as on the accelerator itself.

How Vera Rubin compares with Hopper

The direct 30× claim compares Vera Rubin NVL72 with GB300 NVL72, not with Hopper. NVIDIA’s published AgentX-related material separately says GB300 NVL72 delivers up to 15× higher throughput per megawatt than an H200 Hopper system on DeepSeek V4 Pro.

That provides context for the generational progression, but the figures should not be multiplied into a single 450× claim. They are separate “up to” comparisons, and the available material does not supply one unified test table showing Vera Rubin, GB300 and Hopper under identical conditions.

The same limitation applies to cost. The evidence supports NVIDIA’s claim of a maximum 35× reduction versus GB300, but it does not provide a Vera Rubin dollar price per million tokens or a specific Vera-versus-Hopper cost figure.

What is inside the Vera Rubin NVL72 platform?

NVL72 is a rack-scale system combining 72 Rubin GPUs and 36 Vera CPUs with ConnectX-9 SuperNICs and BlueField-4 DPUs. NVIDIA describes NVLink 6 as the scale-up fabric and Quantum-X800 InfiniBand and Spectrum-X Ethernet as the scale-out networking layers. 23

The design treats the rack as a tightly integrated computing system rather than a collection of independent accelerator servers. That is particularly relevant for agentic workloads, where performance depends on keeping accelerators supplied with data while coordinating CPU-side execution, networking and storage activity.

Why NVIDIA’s Vera CPU is important for agents

Vera is an 88-core CPU built around NVIDIA-designed Olympus cores. It uses Spatial Multithreading and LPDDR5X memory, with up to 1.2 TB/s of memory bandwidth. NVIDIA says Vera enables up to 1.8× faster task completion than x86 CPUs across workloads including agentic AI, reinforcement learning and data processing. 15

Its role is not simply to run a conventional host operating system. Vera is intended to handle the work surrounding model inference, including:

• Tool use
• Code execution
• Sandboxing and orchestration
• Data processing and analytics
• Simulation and reinforcement-learning workloads

By accelerating these tasks, NVIDIA’s architecture aims to keep GPUs focused on inference while reducing bottlenecks in the rest of the agent loop. 159

BlueField-4 and the “Scale-In” infrastructure layer

BlueField-4 is NVIDIA’s dedicated infrastructure-processing component. The DPU combines a 64-core Grace CPU, LPDDR5X memory and ConnectX-9 networking, with connectivity of up to 800 Gb/s. 1

NVIDIA says BlueField-4 moves infrastructure functions—including control, security, data movement and orchestration—off the host CPUs and GPUs. In that model, the DPU acts as an infrastructure layer that helps provide isolated, secure and predictable operation across a large AI factory. 1

This is the logic behind NVIDIA’s “Scale-In” framing: efficiency is not limited to the GPU chip. It also depends on how effectively the rack handles communication, storage, scheduling and security without consuming accelerator capacity for those jobs.

Where DSX MaxLPS fits

DSX MaxLPS is software within NVIDIA’s broader DSX AI-factory approach, rather than a separate Vera Rubin GPU or CPU. NVIDIA describes it as a suite designed to maximize token performance per megawatt within a fixed power budget. Its stated approach combines liquid cooling and in-rack optimization so operators can run more GPUs near their energy-efficient operating points.

That makes DSX MaxLPS relevant to the 30× efficiency claim: the advertised metric is measured at the AI-factory level, where cooling, rack configuration and infrastructure software can affect tokens produced per unit of power. The available evidence does not establish how much of the reported improvement comes from the silicon, the software stack, the test configuration or the combination of all three.

SpaceXAI’s planned Vera Rubin deployments

NVIDIA and SpaceXAI said SpaceXAI plans to deploy Vera CPUs for its next generation of agentic-AI workloads and expand the infrastructure behind Grok using Vera Rubin. 45

The companies also described plans to extend an optimized Vera Rubin NVL72 system into orbit for a first-generation Starmind AI satellite. 4

These statements describe planned deployments. They are not evidence that an orbital NVL72-based system has already launched or entered service.

The practical takeaway

Vera Rubin NVL72’s significance is NVIDIA’s attempt to measure AI infrastructure against the way agents actually work: through long, iterative sessions that consume substantial context and involve tools, code and subagents. On that workload, NVIDIA reports up to 30× more throughput per megawatt and up to 35× lower token costs than GB300 NVL72.

For buyers and infrastructure teams, the key question is not whether the headline multiplier sounds large. It is whether the result holds under their own model, latency target, context length, concurrency, power budget and total-cost assumptions. The supplied evidence establishes NVIDIA’s claim and the architecture behind it, but not a final independent verdict.