NVIDIA Vera Rubin NVL72

Name: NVIDIA Vera Rubin NVL72
Brand: NVIDIA
Price: Contact Sales
Availability: InStock

Rack-scale AI system built for agentic AI.

Contact sales

View pricing

Supercomputer platform designed for agentic AI training and inference

36 Vera CPUs and 72 Rubin GPUs, co-designed to work as one: a rack-scale architecture engineered for reasoning and agentic AI.

NVIDIA Vera Rubin Platform

The Vera Rubin full-stack AI factory platform combines a rich ecosystem of developers and partners, a fungible architecture that runs every model and every workload from pre-training, post-training, inference and agentic AI to accelerated computing, and a durable asset that gets faster with every CUDA release across a decade of service.

NVIDIA Vera Rubin NVL72

NVIDIA Vera Rubin NVL72 is the foundation of the Vera Rubin AI factory platform - fungible across pretraining, post-training, inference, and agentic AI. It pairs 72 Rubin GPUs and 36 Vera CPUs with 4 NVLink, networking, and full-stack software to operate as one unit of compute with 30x higher throughput per megawatt than NVIDIA GB300 NVL72 and 45x lower cost per million tokens.

NVIDIA Vera CPU

NVIDIA Vera is the CPU for agents, delivering 80% faster agentic CPU performance for agentic AI, reinforcement learning, and data processing. By shortening tool calls, sandbox execution, queries, and evaluations, Vera helps agents think longer, RL systems learn faster, and AI factories turn more data into useful output.

Rack-scale performance for agentic AI with NVIDIA Vera Rubin NVL72

72 + 36

72 Rubin GPUs + 36 Vera CPUs
in a single rack-scale system

72 + 36

72 Rubin GPUs + 36 Vera CPUs
in a single rack-scale system

3,600 PFLOPS

NVFP4 inference

3,600 PFLOPS

NVFP4 inference

20.7 TB HBM4

High-bandwidth
GPU memory

20.7 TB HBM4

High-bandwidth
GPU memory

1,400 TB/s

HBM4 memory
bandwidth

1,400 TB/s

HBM4 memory
bandwidth

~90X

Faster compute tray assembly vs. NVIDIA GB200 NVL72, cutting 2 hours to 1 minute*

~90X

Faster compute tray assembly vs. NVIDIA GB200 NVL72, cutting 2 hours to 1 minute*

Ideal workloads for the NVIDIA Vera Rubin NVL72

Agentic AI & reasoning

72 Rubin GPUs connected by sixth-generation NVLink, purpose-built for large-scale reasoning and agentic AI workloads. Power agentic systems that plan, act, and coordinate at data center scale.

Reinforcement learning at scale

NVIDIA Vera Rubin NVL72 delivers the rack-scale throughput reinforcement learning and post-training pipelines demand, with thousands of concurrent sandboxes for agent execution and RL evaluation.

Gigascale training & inference

Train frontier models with one-fourth the GPUs and serve them at one-tenth the token cost, with a 72-GPU unified NVLink domain that operates as one rack-scale system.

Build the future faster with Crusoe Cloud

Up to 20 times faster and 81% less expensive than traditional cloud providers.

Unparalleled performance

Scale ambitiously and fearlessly knowing your AI workloads are backed by peerless infrastructure.

Accelerated breakthroughs

Focus on innovation not infrastructure with intuitive, AI-native systems designed to keep you building.

Abundant compute

Achieve your AI potential with vast compute and energy resources.

Crusoe provides the latest NVIDIA hardware, extraordinary support, and overall a great partner to enable our data, training, and inference infrastructure. Crusoe is consistently accommodating to our needs and is not afraid to dive deep in helping us solve hard problems.

Philip Petrakian

Member of Technical Staff

This was not tens of nodes or hundreds of nodes, but thousands of nodes on extremely short notice. I remember being on the call with Crusoe late evening Friday night, Saturday morning, trying to figure out the best way to get this capacity. We were able to strike a deal within a couple of days, and we got those nodes up and running within a few hours of signing.

Chenyu Zhao

Co-Founder

There are a lot of providers selling GPUs out there, but ultimately the GPU is most useful when it's backed by reliable infrastructure. We've had good success with Crusoe: the network fabric is very performant, and we're able to pump a lot of gigabits on a constant basis.

Sachin Agarwal

ML Infrastructure & Security Engineering Lead

We benchmarked the cluster, and we realized that Crusoe matches the state-of-the-art speed in terms of both the GPUs and the storage. We decided to go with Crusoe because it was the best possible choice.

David Novotný

Co-Founder + CTO

NVIDIA Vera Rubin NVL72