In the US, electricity demand growth driven by AI and cloud services is outpacing the development of new power supply and transmission infrastructure. While it typically takes two to three years to design, permit and build a datacenter, development timelines for power generation resources often extend to five years or more. 

Moreover, utilities have had minimal time to plan for the expected surge in demand. For instance, OpenAI released ChatGPT 3.5 to the public in November 2022, but utilities only began broadly assessing generative AI’s impacts on grid demand toward the end of 2023. Without adequate lead time to obtain project approvals, new utility infrastructure will reflect a regulatory lag.

As a result, supporting rapid, near-term growth in electricity demand will require squeezing more from the existing power generation fleet and delivery infrastructure. Regions with greater load growth will scramble for capacity in the near term and rely on delayed fossil plant retirement. The remaining fossil generation fleet will run harder. And utilities will look to “grid enhancing technologies” such as dynamic line ratings, which often use machine learning, to increase utilization of the existing grid. Utilities in Texas, New York, Ohio and Indiana have already begun using these technologies.

While new datacenters present significant growth opportunities for utilities, uncertainty over the scale, timing and location of new datacenters brings new risks. Opinions vary on how far and how fast datacenter power consumption will grow. A survey of industry stakeholders suggests that datacenter-driven demand growth through 2030, expressed in terms of equivalent state-level electricity demand, could range from “Maryland” to “Texas.”

On the one hand, there is no question that datacenters will drive electricity demand growth. The US has 5-6 GW of datacenter capacity under construction, which will expand the existing fleet by roughly a quarter. Meanwhile, vacancy rates, especially in primary markets, have fallen over the past five years.

On the other hand, the timing and scale of new datacenter-driven electricity demand will depend on several factors and emerging trends:

• Evolution of AI demand and commercial prospects for AI technologies

• Efforts to improve grid connection bottlenecks

• Sufficiency of skilled labor in regions where development is planned

• Evolution of hardware and AI computational and algorithmic energy efficiency

• Adequacy of grid infrastructure and datacenter hardware supply chains

• Datacenters' use of behind-the-meter generation (electricity generated, stored and consumed directly by datacenters)

• Incrementality of new AI workloads — i.e., adding new workloads versus displacing others 

• Capability of devices to perform AI tasks without data transmission to and from a datacenter 

Skepticism over electricity demand forecasts is warranted. For decades, utilities and grid operators have over-forecast load as the industry has underestimated the impacts of energy efficiency improvements and shifts in industrial activity that pushed power demand below expectations.

Similar concerns about the IT sector’s appetite for electricity emerged during the rise of internet services. In 2018, researchers found that demand for internet services rose much faster than associated electricity consumption.

While US technology executives expect significant growth in datacenter-driven electricity demand, they also signal that AI's energy demands may be overestimated. Executives from Google and NVIDIA have suggested that efficiency gains in AI will surpass expectations as energy-efficient hardware and optimized models continue to develop. (Of course, such statements should be taken with a grain of salt, as the big tech players continue to release increasingly dense and power-hungry computing infrastructure.) Some have also indicated that certain forecasts double-count projects, while others point to the likelihood that US datacenter capacity will be overbuilt, potentially leading to overestimation of the associated power demand. 

Further, as datacenters underpin an increasing share of global economic activity, they are reshaping electricity consumption in other economic subsectors. Datacenters are the key infrastructure supporting a large and growing portion of the economy (i.e., the "digital economy"). The US digital economy has seen robust growth for decades and is now one of the largest economic subsectors. The digital economy is reshaping how we shop, work, and spend our time and money, which affects patterns of electricity consumption. From 2018 to 2023, overall retail sales of electricity to US commercial and industrial customers, including datacenters, only rose by about 1 kWh for every 2 kWh of increased datacenter sales, indicating reductions in electricity consumption for other commercial and industrial categories.

Timelines for new datacenter-driven demand may be longer than anticipated. Already, the need for incremental grid infrastructure has delayed datacenter development timelines in key markets such as Northern Virginia and Silicon Valley. As spare grid capacity dwindles, further delays are anticipated, making regional grids a key bottleneck for datacenter development. Queues for large loads have emerged in Texas and Ohio. American Electric Power Co. (AEP), one of the largest utilities in the country, with customers across 11 states, recently informed investors that no spare capacity is available to support new datacenter requests in key subregions such as central Ohio.

In pursuit of utility power, adaptation is also manifesting in datacenter operators' site selections. For example, xAI plans to build a large AI datacenter in Memphis, while Amazon Web Services is planning two hyperscale AI facilities in Mississippi — markets that have yet to see any datacenter development at scale.

Some are seeking to minimize their reliance on the grid through demand response and flexibility, while others are looking to avoid reliance on the grid altogether by colocating with an existing power supply resource. Owners of nuclear plants in competitive markets could strike deals with datacenters like the one made between Talen and Amazon Web Services in early 2024. Others have envisioned more dramatic solutions such as the development of new, private utilities and grids that can develop faster to match technology-sector demand. 

While Google, Amazon and Oracle have all made recent headlines with their plans to purchase electricity supplied by new small modular reactors (SMRs) and enhanced geothermal systems, construction will take years and involve higher levels of financial and technological risks.

Datacenter development will remain "hyperlocalized" in regional clusters in Northern Virginia, Phoenix, Atlanta, Dallas and Columbus, Ohio. However, even in these regions, grid constraints or local pushback may counteract ambitions. Regional distribution will be shaped not only by power availability and lead times but also by connectivity and fiber, latency needs and customer proximity, state and local development policies and tax incentives, water resource availability, skilled labor availability, cost of land and electricity, and natural disaster risk. 

State policies, in particular, present a shifting landscape. In Maryland, Aligned Data Centers canceled a major development project in 2023 after the state's Public Service Commission rejected its request for an exemption to install 168 diesel-fueled backup generators. In mid-2024, the state passed the Critical Infrastructure Streamlining Act to ease future approval of such projects. In Georgia, a bill that would have suspended tax abatements for new datacenters passed through the legislature, but the governor vetoed the measure in May 2024 to continue promoting datacenter development. In the same month, Tennessee passed legislation expanding datacenter tax breaks.

S&P Global Commodity Insights continues to expect that new datacenter-driven demand will increase the near-term outlook for gas- and coal-fired electricity generation, as technology companies appear to be prioritizing competitive pressures over environmental goals. With demand growing faster than new supply, the power sector must lean more heavily on existing fossil generation resources to supply incremental demand. Data from Google and Microsoft reveals that rising datacenter demand has already increased power supply from carbon-emitting resources.

Over time, in addition to uncertainty regarding datacenter electricity demand, power sector fossil generation demand will depend on the pace of adding renewables and other clean energy resources. Most companies driving the current datacenter build-out are directing major investments toward clean energy, accounting for about one-half of all US corporate renewables procurement. At the same time, challenges with building transmission, interconnection backlogs, local opposition to development and permitting delays continue to threaten the pace of renewable additions. Residual demand not met by renewables will primarily be supplied by gas-fired generation. It is not surprising, then, that technology companies have recently expanded investments into clean energy technologies such as advanced geothermal and SMRs to avoid longer-term reliance on fossil generation.

Overall, S&P Global Commodity Insights does not expect that datacenters will produce a significant and sustained increase in total US natural gas demand. By 2040, if all 450 TWh of new datacenter-driven electricity demand (high-end forecast) were incremental and supplied entirely by gas-fired generation, the outlook for total demand for US natural gas would increase by 9 Bcf/d, or 8%, according to data from S&P Global Commodity Insights. However, S&P Global Commodity Insights expects clean energy to supply a large share of incremental datacenter demand over the longer term. If one-quarter of the projected datacenter load were supplied by gas-fired generation, this would translate to a 2% increase in total US gas demand in 2040.

New AI tools for the power sector may help but will not fully resolve issues of power grid adequacy and rising emissions from AI workloads. Initial applications of AI for the power sector do not provide the step-change improvements needed to meet the scale of new power demands. For example, applications such as accelerating security-constrained unit commitment calculations, improving forecasts for electric vehicle charging and nondispatchable power supply technologies, and enhancing vegetation management practices are useful but only offer incremental improvements. Moreover, the electricity required to train and run such models would partially offset the efficiency gains. Ambitions to use AI to "operate" the grid may hold greater potential but will require more time to develop and may introduce new security risks. For instance, using AI in power system operations introduces an additional attack vector for breaches into critical systems, such as through training data "poisoning" that could cause models to learn incorrect behaviors. 

US leadership in AI is bolstered by one of the largest fleets of datacenters globally. According to data from S&P Global Market Intelligence 451 Research’s Datacenter KnowledgeBase, the US hosts approximately 38% of the world's operational datacenter capacity, providing the computational resources necessary to manage complex AI algorithms and large datasets. Though the US is poised to maintain its lead in AI in the coming years, this depends on meeting datacenters' increasing power demands.

In support of this goal, efforts to expand power infrastructure will likely remain a focus of federal action. The US government has announced plans — and in some cases has already acted — to bolster grid development through Federal Energy Regulatory Commission orders targeting transmission planning, cost allocation and interconnection queue backlogs, in addition to directives and financial support from the US Energy Department. However, overbuilding to meet new AI-driven electricity demand could impose financial burdens on utilities and consumers. Striking a balance between meeting demand and avoiding excess capacity will be a crucial challenge for the US power sector in the coming years.