March 23, 2026 • 4 min read

Predictability in chaos: Engineering AI data centers for volatile demand

AI data centers increasingly operate in what appears to be chaos. Power swings, equipment operating limitations, and grid constraints all occurring simultaneously. Yet this volatility doesn’t have to translate into uncertainty. With the right system level design, the unpredictable can become predictable.

THE CHALLENGE

Managing highly volatile AI demand

The rapid growth of artificial intelligence and high-performance computing is transforming the energy profile of modern data centers. AI training clusters, GPU farms, and hyperscale compute facilities are driving unprecedented electrical demand, with individual campuses approaching hundreds of megawatts to gigawatt‑scale loads. At the same time, utilities are struggling to deliver grid capacity fast enough to support these facilities.

To address this challenge, many developers are deploying behind-the-meter or off-grid power systems incorporating on-site generation, battery storage, and cooling infrastructure. These systems are increasingly complex microgrids that must balance power stability, cooling reliability, cost optimization, and resiliency.

OUR APPROACH

Designing the data center as one integrated system

It becomes important to evaluate and design these facilities as a single system. An Omniverse enabled digital twin platform to model, simulate, visualize and optimize integrated electrical and thermal infrastructure as a single system is one way to address this. By combining physics-based simulations such as power system analysis, thermal modeling, battery optimization, and real-time visualization through the NVIDIA Omniverse 3D environment, we can evaluate how electrical and thermal infrastructure behave together under real operating conditions.

The result is a comprehensive environment that improves energy efficiency, enhances resilience, accelerates infrastructure design, and supports reliable operation of AI-scale data centers. The goal is to explore the use of a single platform to design, visualize, optimize and monitor the asset from concept to deployment.

Use cases

Integrated power and thermal system modeling

Modern AI data centers operate as tightly coupled systems where electrical infrastructure, cooling systems, and energy storage interact continuously. Traditional modeling approaches often treat these systems separately, leading to suboptimal designs and limited visibility into dynamic operating behavior.

By using NVIDIA Omniverse Libraries, we can perform physics-based simulations with sim-ready 3D assets to represent not only geometry but also system relationships and physical properties, while integrating physics engines such as PhysX and Newton. In addition, the integration of power systems simulation tools such as ETAP, PSCAD, etc. allows power flow simulations in the background.

This should provide a unified simulation environment that models:

grid interconnection and power flows (including constraints and disturbances)
on-site generation dispatch and operating limits
battery energy storage systems (BESS) performance and control
data center load dynamics and ramp profiles
microgrid control strategies (grid connected, islanded, and restoration modes).

Through this integrated framework, users can evaluate and visualize how power and cooling infrastructure respond to changing load profiles, grid disturbances, and operational constraints. For example, when GPU-based AI workloads produce rapid fluctuations in power demand, the platform simulates these load dynamics and evaluates how batteries, generators, and other storage systems respond in real time. This allows operators to design systems capable of absorbing fast load ramps while maintaining stable voltage and frequency conditions across the facility. In addition, Omniverse enables easy visualization of thermal and electric flows and impacts of design choices on facility footprint.

Data center building with adjacent electrical substation and power equipment.

Cooling and electrical storage optimization

Cooling infrastructure is one of the largest energy consumers in AI data centers. As compute density increases, maintaining stable thermal conditions becomes critical for performance and reliability.

With this in mind, we explore the use of thermal energy storage that addresses the needs of data centers and the onsite generation in an integrated manner. With its ability to integrate computation fluid dynamics (CFD) and thermal simulation tools, Omniverse introduces advanced modeling capabilities for thermal energy storage systems, enabling users to evaluate how chilled water or other storage technologies can support data center operations.

The Worley digital twin platform, based on Omniverse Libraries can simulate:

charging and discharging thermal storage tanks
interaction between cooling demand and compute workloads
coordination between chillers, cooling towers, and storage systems
optimization of cooling energy consumption.

By shifting cooling loads and pre-charging thermal storage during off-peak periods, operators can significantly reduce peak electrical demand from the grid or on-site generation assets. This in conjunction with BESS provides additional flexibility to operators. Omniverse Libraries are used to evaluate these strategies through detailed simulations and identification of optimal storage sizing and operational strategies for each facility. By validating these behaviors before interconnection studies begin, the digital twin provides utilities with greater confidence in the facility’s operational profile and the ability to serve as a demand response asset.

Rows of industrial cooling units with large circular fans.

Value offering

Worley provides designs and simulations that improve both capital efficiency and operational performance, benefiting both data center developers and operators.

For developers this means:

Early prediction of outcomes
Optimized designs that support grid connection and constructability
Development agility to quickly adapt to supply chain pressures
Optimized CAPEX
Faster time to first token.

For operators this supports:

Faster system response
Improved chip utilization
Power smoothing beyond the chip
Right sized equipment and longer asset life
Improved tokens per watt.

Enabling predictable performance in volatile AI data centers

Demand cycles in AI data centers may be unpredictable – but their behavior doesn’t have to be. When power and thermal systems are modeled and controlled as one, volatility becomes an engineering input: something you can simulate, design for, and operationalize.