Addressing Power Quality Challenges in AI Data Centers

The rise of AI data centers (AIDCs) has introduced unprecedented load density and dynamic behavior into electrical power systems, creating new challenges for an industry built on reliable, stable power. While data centers have long served as the backbone of the internet — housing vast amounts of IT and telecommunications equipment — today’s AI-driven facilities are pushing power consumption and power quality demands to new levels. Rapid load fluctuations, harmonic distortion, and voltage instability are emerging as critical concerns for electrical professionals. To understand these challenges in practice, it’s important to first examine how data center loads have evolved.

The early large loads

Small data centers have been in corporate facilities for more than eight decades. Internet-based data centers (IDCs) began showing up in the mid-‘90s. IDCs are much larger than corporate data centers, contain a high density of electronic loads (HDELs), and are located in stand-alone buildings. During the late ‘90s to early 2000s, requests to design and construct IDCs started popping up everywhere as the internet was being built out to provide data to millions of end users. This sudden buildout precipitated a concern among electric utilities that stemmed from the requested size of electric services needed to power these new HDELs.

IDC developers were requesting electric services ranging in size from tens of megawatts (MWs) to well over 100MW. However, electrical distribution networks weren’t designed to provide this range of power to individual customers in downtown districts without some serious and costly modifications to their systems. While electric utilities are eager to sell power, they must be able to justify their investment to cover the design, hardware, installation, maintenance, and other costs to support all customers.

The power demand requests that came from IDC developers stemmed from the power calculations that IDC designers put on their drawings. These new power requirements generated new research to look deeper into how, why, when, and where the need to power what many thought were just commercial buildings filled with IT equipment. Electric utilities were, of course, very familiar with the load requirements that other customers needed (like hospitals with HDELs) to operate their businesses and had no problem providing electric services to meet the power requirements. But when it came to defining the load requirements for IDCs, a new sheet of paper and a sharpened pencil were needed to go back to the basics and determine why IDC developers were asking for such large electrical services.

Defining high-density electronic loads for IDCs

Power requirements for the density of electronic loads are defined by a metric called power per unit area (or power density). Power can be defined as real power (in watts) and apparent power (in volt-amperes, or VA). The unit of area used in the United States is the square foot (ft²). So, for IDCs, we use the ‘watts per square foot (W/ft²)’ metric, or the ‘VA per square foot (VA/ft²) metric. Electronic IT loads have active power factor correction already designed into their power supplies, so the power factor is close to unity. This allowed engineers to discuss these new load requirements using the W/ft² metric.

With new discussions about HDELs came a focus on which HDELs were being discussed — the HDELs across the IDC’s floor space or the HDELs installed in a single server rack. This needed to be well defined to avoid confusion about which HDELs developers and engineers were discussing. The fact that two HDEL numbers were floating around in discussions between developers and electric utility engineers did create some confusion. The per-server rack power requirements were around 12kW. HDELs as high as 3,000 W/ft² were presented to electric utilities. At the same time, HDELs as low as 50 W/ft² were also requested. This wide range created even more confusion.

Research helped to define two types of HDELs: the IDC floor space HDEL and the server rack HDEL. Server rack HDELs, on average, were requested between 1,000 and 3,000 W/ft², while floor space HDELs (ranging from less than 50 W/ft² to more than 100 W/ft² for premium IDCs) were also requested. These wide ranges created electric service requests ranging from less than 100MW to more than 500MW for hosting-type IDCs.

Defining HDELs for AIDCs

Today’s discussions are all about AIDCs. The HDELs in these centers are even higher than 3,000 W/ft² per server rack and 100 W/ft² across their floor spaces. This stems from IT equipment designed for AI requiring power over 100kW per rack. Such higher power levels are based on powering high-performance hardware that uses graphics processing units (GPUs) and tensor processing units (TPUs), field-programmable gate arrays (FPGAs), advanced cooling systems, and massive data storage devices.

Running AI models requires processing large amounts of data very quickly and “re-training” and fine-tuning AI algorithms to support complex tasks such as machine learning, deep learning, and language processing. AI processes require massive data storage devices such as non-volatile memory express (NVMe) units, high-speed solid-state drives (HSSSDs), and parallel file systems. Air cooling, using computer room air conditioning (CRACs), can’t keep IT equipment cool enough, so thermal management engineers looked toward liquid-cooling AC systems to boost performance and efficiency. AIDCs must also be scalable to keep up with the demands for processing large amounts of data. As the demand for AI processes grows, so will the HDELs. A single AIDC can require one to 3GWs of power, equivalent to powering 750,000 to 2 million homes.

Powering traditional industrial customers & AIDCs

Powering traditional IDCs from electric utility distribution systems was, at that time, challenging and presented interesting distribution power quality (DPQ) problems. This was mainly because utility distribution systems are more exposed. After all, they have a lot of circuit miles routed through densely populated areas with lots of tree cover.

Traditional IDCs present relatively steady loads, while AIDCs introduce highly dynamic demand that challenges transmission system stability. Utility engineers are more concerned with keeping transmission systems stable because larger amounts of power flow across them. Power management on and across transmission systems is a much more delicate process than on utility distribution systems.

Utility transmission systems operate under different and more complex protection and control schemes than utility distribution systems. It is not uncommon for the load on a specific transmission circuit to be suddenly transferred to another transmission circuit in the event that a problem occurs on the original circuit. This type of behavior is expected by utility engineers, and the effects of such a load transfer are known and planned for.

One goal in managing transmission power quality (TPQ) is to properly manage the shift of a large load or group of loads powered by a utility transmission system. Distribution substations receive power from transmission systems. Large distribution transformers step the voltage down and current up to support utility distribution circuits. Most industrial customers are also powered directly from utility distribution substations, and transferring their load from one substation to another is also known and expected.

However, because of the larger power requirements, some industrial customers are directly fed by utility transmission systems. It is also common for a utility transmission substation to power multiple distribution circuits and a few industrial customers directly fed from one or more substation transformers. In other words, it’s common for an industrial customer to share the same substation bus as other commercial distribution customers. However, the load characteristics of most industrial customers, including the ones that are directly fed at the transmission level, are well known and predictable. Although these are industrial load customers, the ramping up and down of these loads is much more gradual. Because of this, starting up an industrial plant, altering production levels, and shutting down a plant are load behaviors that have known effects on the stability of the utility power system. Utility power systems are typically able to manage their power delivery when an industrial customer shuts down and when their load suddenly trips offline. Protection and control schemes designed into the utility power system help protect, control, and successfully manage load changes without impacting distribution or transmission systems.

Load behavior of AIDCs

Unlike traditional IDCs, the load behavior of an AIDC is much more dynamic because of the way AI processes work. IT equipment is designed and programmed to quickly ramp up the processing power that IT equipment requires to support AI processes demanded by end users. Large banks of processors and data storage devices must operate in parallel to produce AI results. The fast ramping and fluctuating nature of AI-based IT equipment directly correlates to the fast ramping and fluctuation of the power to operate an AIDC.

The dynamic load behavior of AIDCs has been described as “volatile and demanding.” This means that they are essentially unpredictable and can change levels very quickly — within a few seconds. While a few seconds can be an “eternity” as compared to the 1/60th of a second (i.e., 60 hertz) power frequency, the stability of utility transmission systems can be challenged when AIDC loads exhibit this dynamic behavior. For example, if an AIDC is operating at a baseline of 50% to 70% of its power requirement, its dynamic load can quickly ramp up to 130% within seconds.

Power quality challenges for utilities and AIDCs

Power quality challenges surfaced for utilities and IDC customers when traditional IDCs were placed on utility distribution systems. Utilities had to plan for the increased load on the distribution systems that did not normally experience these loads. It is a well-known fact that increased load on any circuit — utility and customer — will cause additional and new PQ problems to surface. Although the load of IDCs was higher, it was a steady load. This doesn’t mean the traditional IDC load didn’t vary; it just didn’t vary a lot in a short amount of time.

Server loads would decrease as well as increase in a predictable and steady manner. Similarly, IDC customers had to plan to experience the typical PQ problems that occur on utility distribution systems (i.e., common everyday disturbances), including voltage sags, voltage swells, harmonic distortion, and voltage transients, among other disturbances. Utility power systems are designed to protect their circuits as well as avoid long interruptions and outage periods to help their customers “keep the lights on.”

Most IDC customers had no problem preparing for the typical PQ characteristic of distribution circuits. Like most other commercial and industrial customers, IDCs used a bank of uninterruptible power supplies (UPSs), fast transfer switches, emergency generators, and other typical PQ mitigation equipment and systems. Surge protection devices (SPDs) were also frequently used to help protect expensive IT and communications equipment from voltage surges caused by thunderstorm lightning and transients generated by the normal operation of UDSs. Some traditional IDCs engaged in PQ monitoring to record their baseline PQ for voltage (that belongs to the utility) and current (that “belongs” to the customer). However, most IDCs were not prepared to keep track of their PQ, much less investigate the cause of a PQ problem when one occurred. The traditional IDC customers who didn’t engage in PQ monitoring quickly learned the value of permanent PQ monitoring. These shifting load characteristics become even more evident when you look at how power demand fluctuates:

Rapid changes in power usage

AIDCs use server racks that consume much more power than the racks in traditional IDCs. This increased power usage means that if a bank of 10 x 10 server racks (i.e., 100 racks) that draw 100kW per rack have servers that quickly ramp down in power usage or suddenly trip offline, the system will see a decrease in power by as much as 10MW. This conservative number isn’t often experienced by utility power system circuits. In fact, a decrease in power by as much as 10MW can easily grow to 1,000MW (or 1GW). Sudden changes in power levels like this will cause utility power systems to experience PQ problems that must be carefully managed.

A sudden increase in power level in AIDCs can also mean a sudden increase in power level for AIDC cooling systems, which must respond quickly to prevent the AIDC temperature from reaching levels that could cause servers to be tripped offline to avoid overheating. Such increases in cooling demand will also cause cooling systems to experience sudden increases in power levels, adding to the increase in power levels caused by AI servers ramping up very quickly.

Sudden increases in AIDC power levels can range from a millisecond to a few seconds and take the shape of a sawtooth waveform. Traditional utility transmission systems face challenges in remaining stable in frequency and voltage when such load excursions occur. Not only do increased steady power levels cause new PQ problems to surface, but increasing power levels within such a short time frame will cause other new PQ problems to surface.

Transmission systems must have time to respond to sudden changes in power levels. Multiple AIDCs on a single transmission system could also cause multiple frequency and voltage stability issues on the same circuit. Such challenges must be carefully managed and mitigated to avoid creating low-frequency oscillating voltages that could cause widespread grid instabilities and cascading outages. The fast-changing load profiles of AIDCs will necessitate the installation of sophisticated PQ mitigation systems to help compensate for the effects of such large power swings that AIDCs can cause on transmission systems.

Voltage and current harmonic distortion

Facilities like AIDCs that produce a significant amount of heat, which require rapid and continuous cooling, will use a significant number of variable-frequency drives (VFDs). They are available in multiple sizes and configurations ranging from a few horsepower (HP) to 1,000’s of HP. Unfortunately, they are also available in various configurations, which can cause a significant amount of harmonic current. Six-pulse VFDs are known for high harmonic currents, which can easily cause parts of the AIDC’s power system to overheat, as well as substation transformers. Fuses used in utility substations to protect their transformers are not rated for harmonics. High harmonic current content can cause premature fuse blowing. Increased inrush currents caused by rapid calls for additional cooling in AIDCs will also cause increased fuse temperatures as well as premature blowing.

The effects of high harmonic currents can easily be high harmonic voltages. Power systems with a soft source (i.e., a source that has a higher impedance) will allow more harmonic voltage distortion to occur. High harmonic voltage distortion can cause severe flat-topping of the source voltage, which is passed downstream to electronic power supplies; thus, increasing the strain on power supply operations, as well as temperature increases and shortened life. Highly distorted voltages can also cause elevated VFD temperatures and reduced VFD efficiency. High harmonic voltage distortion can also impact the voltage quality of other customers on the same transmission circuit.

When voltage oscillations caused by sudden changes in AIDC load occur, subharmonics will develop. Subharmonics are harmonics below the power frequency (60 HZ in the U.S.). Unfortunately, most active harmonic filters cannot mitigate against subharmonics and can destabilize local generators and DC-to-DC converters.

Power quality monitoring

PQ monitoring allows the characteristics of the voltage and current to be recorded with respect to time and frequency. Customers who engage in permanent PQ monitoring establish a baseline of PQ performance that can be used to: 1) develop a baseline of voltage and current quality, 2) investigate unexpected PQ problems, and 3) provide data to characterize the large load characteristics of AIDCs.

Utility companies are very familiar with PQ monitoring. Most utilities engage in some form of PQ monitoring at the substation level or other levels in their transmission and distribution systems. Engaging in PQ monitoring allows utilities to “keep an eye” on their voltage quality (delivered to the customer) and current quality (formed by the customer for the utility to provide). Without PQ monitoring, utilities would not be able to understand their own internal power quality status within the grid, much less the PQ delivered to their customers.

Utilities should require that their AIDC customers install PQ monitoring at least at the switchgear level and at the point of common coupling (PCC) where the power is handed off to the customer. Utilities will benefit from having access to PQ data recorded by the AIDC customer as well as that recorded at the substation level.

Smart PQ monitoring should have the capability to notify the utility company of changing PQ conditions that could impact the performance of the utility power system. Such prior notification may be able to allow the utility to prepare for potential PQ problems. For example, if a PQ monitoring system could notify a utility company that the AIDC load was about to experience a significant change, specific actions could be taken to avoid unfavorable PQ effects on the utility power system. Remote PQ monitoring is a must for the AIDC as well as for the utility company, so that both parties are kept informed of continuous PQ conditions as well as sudden changes in PQ conditions that could impact the performance of the AIDC as well as the utility power system.

Conclusion

The bottom line is AIDCs are here to stay. Although their load profiles will continue to change, the PQ challenges associated with operating them will only intensify. Utility companies and AIDC designers, owners, and operators should prepare to take on the PQ challenges they face head-on. Predictive PQ performance is one area of power quality that can provide significant advantages to the utility company and AIDC customer. This can prepare and position both parties to adapt to specific system operations that minimize the impact of sudden changes in power quality.

If AIDCs can operate from advanced learning algorithms, then PQ mitigation and solution strategies can also benefit from smart algorithms that predict upcoming changes in system status and prepare the energy supplier and end user for actions that will keep the utility power system online and the AIDC customer up and running.