NVIDIA and Partners Transform AI Factories into Flexible Grid Assets

NVIDIA and Emerald AI have announced a collaboration to redefine AI data centers as flexible, intelligent grid assets rather than static power loads. Unveiled on March 23, 2026, at CERAWeek, the initiative aims to accelerate the deployment of AI infrastructure while enhancing the reliability and affordability of the power grid.

The partnership integrates accelerated computing, AI factory reference architectures and real-time energy orchestration. This approach allows large-scale AI deployments to connect to power grids more quickly and operate with greater efficiency by dynamically responding to grid conditions. By flexing power usage during peak demand, these factories can reduce the necessity for overbuilding infrastructure.

The Architecture of Power-Flexible AI Factories

The system is built upon the NVIDIA Vera Rubin DSX AI Factory reference design and the Emerald AI Conductor platform. This architecture combines compute, power networking, and control into a single system. The NVIDIA Vera Rubin DSX design includes the DSX Flex software library, which specifically enables the connection of AI factories to power-grid services.

View this post on Instagram

To further speed up deployment, these factories can utilize co-located energy generation and storage as bridge power for hybrid AI factories. Once connected, these resources can be used to supply the grid flexibly. Emerald AI’s Conductor platform orchestrates this computational flexibility alongside onsite generation and batteries to provide grid-responsive power while maintaining service quality for AI compute tenants.

NVIDIA founder and CEO Jensen Huang describes this computing infrastructure paradigm as a five-layer AI cake, with energy serving as the foundational layer.

Prioritizing Tokens Per Second Per Watt

As power constraints increasingly reshape data center design, NVIDIA is focusing on performance per watt—specifically the metric of tokens per second per watt. The goal is to maximize the amount of AI intelligence generated within a specific power budget to lower operating costs and increase revenue.

Power is a concern, but it’s not the only concern. That’s the reason why we’re pushing so hard on extreme codesign, so that we can improve the tokens per second per watt orders of magnitude every single year.

Jensen Huang

According to NVIDIA, the number of tokens generated within the same power budget has increased by more than one million times between the 2012 NVIDIA Kepler GPU and the current NVIDIA Vera Rubin platform.

Implementation and Grid Impact

A primary implementation of this standard is the 96MW Aurora AI Factory in Manassas, Virginia. Under construction by Digital Realty and slated to open in the first half of 2026, the facility serves as a live innovation hub to demonstrate how AI compute can align with grid needs to relieve stress during peak demand.

The broader adoption of this power-flexible reference design could potentially unlock an estimated 100 GW of capacity on the existing U.S. Electricity system, which is equivalent to 20% of the total annual U.S. Electricity consumption.

Several energy companies are collaborating to build the necessary generation capacity and optimized strategies to support these factories, including:

• AES
• Constellation
• Invenergy
• NextEra Energy
• Nscale Energy & Power
• Vistra

Ecosystem Integration and Infrastructure

The scaling of these factories involves a network of infrastructure partners focusing on the power-to-rack challenge. GE Vernova is utilizing high-fidelity digital twins aligned with the NVIDIA Omniverse DSX Blueprint to simulate grid behavior and substations before deployment. Schneider Electric has developed validated Vera Rubin reference designs and lifecycle digital twin architectures with AVEVA to optimize performance per watt.

Vertiv is providing converged, simulation-ready physical infrastructure using repeatable power and cooling building blocks integrated with the Vera Rubin DSX reference design.

Other technological advances are supporting the energy infrastructure required for the intelligence era:

• Maximo: A solar robotics company that completed a 100-megawatt robotic solar installation at the AES Bellefield site using NVIDIA Omniverse and the Isaac Sim framework.
• TerraPower: Working with SoftServe to use an NVIDIA Omniverse-powered digital twin platform to reduce the design and siting cycles for Natrium advanced nuclear plants from years to months.
• Adaptive Construction Solutions: Launching a national registered apprenticeship initiative with NVIDIA to train the skilled workforce needed for AI factories and power systems.