AI load in data centers: new aspects of uninterruptible power supply system development

Analyzing Vertiv's latest developments, we would like to highlight three key mechanisms for adapting UPSs to these conditions:

1. Battery Shield: Battery resource protection

In typical modern UPSs, any sudden change in output load causes the system to instantly “turn” to the batteries to compensate for the energy deficit. Since the AI load changes thousands of times per hour, the battery is constantly in micro-discharge mode. This destroys its chemical structure and overheats the plates.

Vertiv's Battery Shield technology works on a different principle:

• Energy storage in the DC bus: The UPS has a bank of powerful capacitors that maintain a stable voltage between the rectifier and the inverter. Capacitors are capable of delivering and receiving energy almost instantaneously, unlike batteries, where a slow chemical reaction takes place.

• Dynamic voltage window: The system allows the voltage on the DC bus to fluctuate within certain limits. When a load surge occurs, energy is taken from the electric field of the capacitors.

• Filtering effect: High-frequency AI pulses are “drowned out” in the capacitance of the capacitors before reaching the battery terminals.

Thus, the battery remains physically disconnected from these fluctuations (or operates in “idle” mode), which is critical for preserving the life of both classic lead-acid and lithium batteries.

2. Input Power Smoothing (IPS): Input network stabilization

Input Power Smoothing (IPS) is an algorithm that turns the UPS battery system into an active power filter. Its main task is to ensure that the input network (or generator) does not “notice” sudden jumps in AI equipment consumption.
Here is how this process is physically implemented, according to Vertiv documentation:

• Determining demand: The system analyzes the average consumption of the facility. Anything that falls outside this “norm” (up or down) becomes subject to regulation.

• Peak Shaving Mode: When the load increases sharply (the same transient from 10% to 130%), the UPS does not take this difference from the network. Instead, the inverter takes the necessary deficit energy from its own UPS storage system (capacitors, batteries). Thus, the input current remains stable.

• Pre-charge mode (Valley Filling): As soon as the peak passes and consumption drops, the UPS begins to smoothly recharge its internal energy storage devices. It is important that this does not happen instantly, but at a controlled rate, so as not to create sudden fluctuations in the grid.

• FR% (Frequency Range) parameter: This is a “tolerance window.” You can configure the system to activate IPS only when power fluctuations threaten the stability of the grid frequency.

This algorithm is critical for facilities with limited connection power or where the grid has low inertia.

3. Input Power Ramp. Safe operation with generators

This is a critically important feature for stable system operation in conditions where the data center switches to power from local generation (diesel or gas units).
The main technical problem lies in the $dI/dt$ parameter—the rate of change of current. Most generators have mechanical inertia and cannot instantly adapt to the sudden load surge generated by AI.

Here's how Input Power Ramp solves this problem, according to Vertiv's technical description:

1. Load “angle” control: Instead of supplying 100% power to the generator in a fraction of a second, the UPS limits the rate of increase of the input current. This creates a “smooth rise” instead of a vertical jump.

2. Using batteries as a temporary bridge: While the generator gradually ramps up, the UPS takes the missing energy from the batteries. This means that the load on the generator increases evenly, and the AI equipment receives the necessary power instantly.

3. Protection against “drops”: This approach prevents a sharp drop in frequency and voltage at the generator output, which could trigger protective automation and a complete system shutdown.

This algorithm actually gives the generator time to “spin up” without creating a shock state for it.

Analyzing the above information, we can conclude that the introduction of AI loads changes the very model of designing uninterruptible power supply systems. The critical factor is not only ensuring the necessary power and autonomy time, but also the ability of the system to effectively cope with highly dynamic changes in consumption.

• Millisecond transients with large amplitudes create increased requirements for:
• the speed of power electronics;
• the permissible charging/discharging currents of batteries;
• input power smoothing algorithms;
• the thermal stability of components;
• load transfer control to generator sets.

Technologies such as DC bus energy buffering, Input Power Smoothing, and Input Power Ramp demonstrate a transition from passive backup to active control of the energy dynamics of an object.

Therefore, the compliance of the infrastructure with AI requirements is determined not by the nominal characteristics of the UPS, but by the system's ability to control transients, limit $dI/dt$, and maintain the stability of network parameters under rapidly changing load profiles.

* Based on an analysis of Vertiv's technical document “Advanced UPS controls for AI workloads management,” an article was prepared for Alpha Grissin Infotech Ukraine.