How NVIDIA's Vera Rubin Builds a Foundation for AI Factories

Straight from NVIDIA GTC, the tech and AI giant has set out its reference design for the NVIDIA Vera Rubin DSX AI Factory, presenting a new blueprint for the modern data centre built on co-designed AI infrastructure.

Alongside physical architecture, the company is making its Omniverse DSX digital twin blueprint available to support large-scale AI factory buildouts.

Fully compatible with NVIDIA Vera Rubin DSX, the blueprint allows operators to model entire AI factories virtually, testing layouts and power distribution before construction begins.

“In the age of AI, intelligence tokens are the new currency and AI factories are the infrastructure that generates them,” says Jensen Huang, Founder and CEO of NVIDIA.

“With the NVIDIA Vera Rubin DSX AI Factory reference design and Omniverse DSX Blueprint, we are providing the foundation to build the world’s most productive AI factories, accelerating time to first revenue and maximising scale and energy efficiency.”

Industry partners including Cadence, Dassault Systèmes, Eaton, Jacobs, Nscale, Phaidra, Procore, PTC, Schneider Electric, Siemens, Switch, Trane Technologies and Vertiv support both the blueprint and reference design, contributing tools and systems to help build and operate AI-focused data centres.

Designing data centres for power and performance

The NVIDIA Omniverse DSX Blueprint standardises how compute, storage, networking and operational systems interact. 

This standardisation enables large-scale deployments to become more predictable and efficient, which is important because inefficiencies in one layer can restrict performance across the entire facility.

The Vera Rubin DSX architecture places energy efficiency alongside performance. AI workloads demand dense, high-power environments, meaning facilities must support concentrated compute without exceeding power or cooling limits.

The modular design links compute infrastructure directly with power and cooling systems. This connection allows real-time balancing of performance against available energy, helping operators maintain output while managing constraints.

This essentially means that data centres operate with greater awareness of energy supply and thermal capacity, rather than treating them separately.

Software stack links compute and operations

The Vera Rubin DSX software stack is open, modular and composable, meaning components can be combined and adjusted without rebuilding entire systems. It connects hardware with power and cooling infrastructure to optimise performance within defined limits.

Using its DSX Max-Q library, the system aims to “maximise AI tokens per watt of available energy”. AI token is a unit of data processed by an AI model, often used as a measure of output. Maximising tokens per watt therefore reflects how efficiently a data centre converts energy into usable AI work.

DSX Flex links AI facilities to power grid services, enabling dynamic adjustment of energy use. This includes coordination with hybrid onsite generation, such as renewable sources or backup systems, helping facilities reduce consumption while maintaining grid stability.

DSX Exchange supports secure and scalable integration of compute, networking, energy, power and cooling signals across IT systems, operational technology and operations teams.

This level of integration allows data centre teams to manage infrastructure holistically, rather than in isolated silos.

Digital twins reshape deployment and operations

NVIDIA Vera Rubin’s DSX Sim and DSX SimReady tools enable high-fidelity digital twin validation. As AI infrastructure scales towards gigawatt levels, aligning power and compute are now essential rather than optional.

The ability to simulate what a data centre will look like, from the inside out, allows operators to identify inefficiencies and optimise energy and cooling configurations without making costly physical changes.

Once operational, digital twins support continuous monitoring, workload adjustment and infrastructure refinement in real time. This ongoing feedback loop allows facilities to adapt as demand changes.

Using the NVIDIA DSX Air platform, operators can model GPUs and partner infrastructure in the cloud. The three-dimensional geometry and logistics simulations of the platform can accelerate deployment and shorten time to first revenue.

Technology and engineering companies are already applying these tools. Dassault Systèmes has integrated the reference design into its Model Based Systems Engineering platform to build virtual twins of AI factories, while Schneider Electric and Eaton provide digital twin models and simulation tools for power distribution and cooling.

Cadence uses SimReady models of NVIDIA GB300 NVL72 to simulate thermal and fluid dynamics, helping plan high-density environments, and Siemens develops frameworks to balance compute and cooling. Thermal management to reduce energy use in large-scale facilities is another significant part of data centre construction, which Trane Technologies specialises in.

Across these deployments, data centres take on a new role. They operate as integrated systems where compute, energy and cooling are functioning as a single, optimised environment.