Inside the AI Factory, Part 1: Simulating High-Fidelity Electrical Dynamics

An AI data center isn’t a scaled-up version of the facilities built five years ago, it’s a different engineering problem. GPU clusters like NVIDIA’s GB200 and Vera Rubin platforms have pushed rack power density up roughly 10x, from around 10 kW to well over 100 kW per rack (with talk of a full megawatt per rack on the horizon). Cooling is shifting from air to hybrid and liquid/phase-change systems. Power sourcing is shifting from grid-only to hybrid energy resilience strategies. Power distribution is shifting too: the 54 volts direct current in-rack standard that carried data centers through the last decade is out of headroom, and the industry is converging on 800 volts direct current to replace it. And design cycles, the time from concept to a live facility, keep shrinking.

Practically, that means electrical and thermal engineering can’t stay separate disciplines handed to separate teams late in the design process. They must be designed together from the start.

This is the first post in a three-part series following up on our recent webinar, Inside the AI Factory: Simulating High-Fidelity Electrical Dynamics. This post addresses the electrical side: how volatile AI workloads stress a facility’s power system, and how system-level simulation exposes those problems before they appear on a live floor.

The Challenge

Traditional data center loads are relatively steady, but AI workloads are very transient. AI workloads come in two major flavors: training loads, which occur when large language models (LLMs) are being trained, and inference loads, which occur when users interact with LLMs. Training loads are notoriously transient, quickly transitioning back and forth from low power to high power in a matter of milliseconds. For a deeper dive on the different training load types, National Laboratory of the Rockies published an analysis and raw datasets of Generative AI workload power profiles here. Using this dataset for inspiration, we set up a training load and an inference load to test our electrical system with:

Every electrical component in the system is only built to handle a narrow range of voltage before overvoltage or undervoltage lockouts occur. The intense transitions in power demand from AI training loads, combined with the resistance and inductance of the power system, cause voltage sag and ripple that stress the system. At hyperscaler densities, with racks now pulling 100+ kW each and busways running the length of a data hall, that adds up to a unique challenge: making sure voltage stays within tolerance everywhere in the building, including at racks far down the busway.

Building an 800 VDC facility model

To ground the simulation in something realistic, the facility model draws on Heron Power’s published blueprint for 800 VDC data centers, which lays out an architecture built around a solid-state transformer (SST) with integrated battery backup. Using that as a reference, a representative row of IT racks was built in GT-SUITE, including:

• A grid-following rectifier: converting medium-voltage three-phase AC down to 800 VDC
• Long DC runs: busways and wires connecting the bus to each rack, modeled with per-meter resistance and inductance pulled from a real busway manufacturer’s spec sheet
• High-power IT loads: representing the compute racks themselvesFour simulation cases were then defined: steady-state full load, a training-load transient, an inference burst, and a power outage.

What the results showed

For the AI rack closest to the main DC bus, the results were relatively stable. Focusing on the spike in power at 400 milliseconds of the AI training load, which jumps power from 10% to 100% in a few milliseconds, the rack saw a voltage sag under 1%, settling in less than 10 milliseconds, and a current overshoot under 1%, also settling within a few milliseconds.The everyday-use case (inference-burst case) was gentler, which is what we would expect: sub-1% voltage sag and under 0.5% current overshoot, both settling faster than in the training case.
As mentioned earlier, we also wanted to study the effect of rack position on voltage sag and ripple because the resistance and inductance of the busways accumulate over distance. Racks located farther down the central power (busway), farther from the rectifier and any local energy storage, showed noticeably more voltage sag and more ripple than racks close to the source. Once workloads move this fast, layout matters just as much as component sizing.

Why this matters

None of these effects would show up in a simple steady-state power calculation. They appear only in simulation of the actual millisecond-scale transient behavior of the system, and that analysis gets harder to skip as rack densities climb and data halls grow longer.

Part 2 examines how adding a battery backup unit to this same model changes the picture, specifically to utilize the battery to smooth out the power demand on the grid.
Check out our webinar introducing this model or contact us for a live walkthrough and see how this approach could apply to your own facility design!

Visit our Data Center Solutions page to explore more data center simulation topics, learn more about our Digital Twin Solutions and Machine Learning capabilities, and browse our blog page for additional insights on digital twins and data center innovation. Follow our LinkedIn channel to stay updated on the latest advancements in our solutions.  Subscribe to the GT blog to receive updates on new AI capabilities, engineering solutions, and industry trends.