Training and Saving Checkpoint Progress: NERC's White Paper on Emerging Large Loads

A white paper of NERC's Large Loads Task Force, dated July 2025, carries the title "Characteristics and Risks of Emerging Large Loads." A NERC committee set up the task force, the paper says, to analyze the reliability impacts of emerging large loads such as data centers, among them cryptocurrency mining and AI, industrial facilities and hydrogen production facilities. The paper describes large loads and the reliability risks they may pose to the bulk power system. Parts of it follow, not the whole.

Training and Inference

The paper defines AI training as "the creation and modification of an AI model and its associated model weights." The electrical demand, it says, "is different for training an AI model versus using an AI model to create inferences." Generally, in its words, AI training executed in large clusters "has rapid fluctuations during training periods and while saving checkpoint progress," and the transition from training to saving checkpoint progress, or back again, "may happen in under one second."

Its Figures 2.2 and 2.3 show the demand of one block of a larger AI training data center that starts a training run, recorded with a high-speed recorder, from data the paper says was supplied by a company it names. In that recording, the paper says, "jittery spikes up and down occur" while the training is active. Under the heading of AI inference data centers, the paper says it is common "for data centers to be purpose-built for either inference or training workloads," and "the rapid ramps up and down seen in AI training/checkpoint cycles are not observed in current AI inference methods."

Backup Power and Checkpoints

The power delivery systems of a data center, the paper says, may include emergency backup systems, among them UPS and on-site generation. According to data center industry experts, the paper says, power delivery systems for "HPC loads (AI data centers) commonly exclude UPS protection systems for the IT equipment." Instead, in the paper's words, "they employ a checkpoint process where checkpoints are saved at regular intervals," and if the power to the IT equipment is interrupted, "the checkpoint may be restored after the power interruption ends." Of its Figure 2.1, which shows how a data center's components and other features, such as UPS and backup generation, may be arranged, the paper says that it is "more relevant to traditional data centers." "For AI training loads, there may be no UPS for the IT equipment," it adds.

Risks to the Bulk Power System

In its section on forced oscillations, the paper says that, broadly speaking, the reliability risks associated with them "are not always well defined." The bulk power system, it adds, "was not designed to operate with large and persistent subsynchronous oscillations." Some large loads, it says, "can produce forced oscillations at frequencies below 60 Hz," either "as a natural byproduct of the load's end use or as the result of unexpected control interactions." In the same paragraph, the paper says that AI data centers "can have load profiles that are periodic, repetitive, and sustained in nature," and points to an example load profile of an AI data center during training.

In its section on voltage fluctuations, the paper says that some new large electric loads, "such as cryptocurrency miners and AI data centers, also have the potential to introduce significant voltage fluctuations to the supply system." Of large loads in general, its conclusion says that "the fast ramp rates and variability of the loads could exhaust reserves for balancing and contribute to voltage and frequency instability." The paper says that later work of the task force, "like the next white paper and reliability guideline, will need to address the risks noted in this white paper."