Large Computational Load Growth: A literature review and first-order assessment of load characteristics and its impacts on emissions, costs, and decarbonization on CAISO

Abstract

The proliferation of artificial intelligence (AI) is supercharging the development of computational loads, particularly data centers, in the United States. Their increasing demand coincides with the onshoring of manufacturing and increased electrification, driving total load growth to levels not seen in decades. Computational loads pose novel challenges to the electric industry. The last few years have been defined by a frenzy to secure power for new development—data center operators are flooding utilities with requests for reliable, firm service. In some cases, these are duplicative or speculative requests inflating load growth projections, which are key to determining grid investments. Moreover, data center operators are seeking service at unprecedented large single-facility power draws. The scale of computational loads, some reaching into multi-hundred megawatt or nearing a gigawatt, presents both reliability and regulatory challenges. A review of how they impact the reliability of the bulk power system is already underway but there is less focus on what these loads mean for the electric industry’s regulatory regime, which requires publicly regulated monopolies to serve all customers in a territory. Just one or a handful of computational facilities can completely reorder the power sales of a utility. This reordering means that at disproportionate share of the utility’s capital expenditures are dedicated to serving those customers, increasing the risks to all ratepayers if that customer does not materialize in demand or duration as expected. This masters project presents a review of the academic and grey literature regarding computational load behavior, with a focus on AI loads. While exploring how different cost allocation procedures can protect incumbent customers against cost-shifting is out of the scope, the review summarizes the literature on concerns raised by computational load growth and the investment needed to serve it. In addition, this project extracts from the literature review several assumptions needed to obtain a first-order estimate of the emissions and cost savings that will be caused by computational load growth in California’s power system CAISO. A streamlined Economic-Dispatch model is used to simulate CAISO’s hourly operations and estimating the costs and emissions that would result from serving an additional curtailable load without increasing the power generation or transmission capacity of the system. The literature review offers two findings. First, computational loads, which are likely to export a vast majority of their digital products and services out of that utility's service territory, do not bring with them the same suite of economic benefits (e.g., jobs) as other large loads, and hence, the costs of serving that may be more than the benefits those loads provide to that territory. And second, load growth driven by an emerging technology like AI is inherently uncertain. If that forecasted load were to be inflated or does not materialize in magnitude and duration as expected, the costs of the infrastructure built to serve it will be borne by all the utility’s ratepayers. Protective ratemaking and cost allocation processes can avoid this outcome but are likely to require regulatory intervention. The simulation of CAISO’s operation assuming the integration of new computational load offers the following conclusions:

Increased utilization of the existing generation system can lead to material cost savings, $0.10-0.13 per MWh, when load growth is low to moderate. While emissions are harder to discern, it's clear that computational loads which serve personal uses and therefore behave similarly to residential load, are likely to lead to worse environmental outcomes. To serve a constant and permanent demand from a computational task (i.e., 24/7 flat consumption) with carbon-free power will significantly impact the feasibility and costs of decarbonization. For example, California would need to more than double its existing nuclear capacity to replace the additional generation from its existing natural gas resources to serve 3.5 gigawatts of data center growth. Doing so would abate only ~86% of the additional emissions relative to a no-load-growth scenario.

Finally, this paper culminates with several recommendations to ensure that the grid can integrate data center load growth sustainably:

Regulators and grid operators should seek to standardize the large load interconnection process to minimize speculative and duplicative requests. Regulators and grid operators should create legal and market incentives for loads to become flexible. Doing so would be a win-win-win that: a. Better utilizes the capabilities of large-scale data centers and rewards them with a quicker time to energization; b. Better utilizes the existing generation resources such that the same power plant costs are spread over more MWhs, reducing costs for all ratepayers; c. Hedges against overinvesting in thirty-year infrastructure built to serve AI-driven load with significant uncertainty.