Generative AI is set to usher in an era of remarkable economic growth and efficiency. By automating complex tasks, enhancing decision-making, and driving innovation across industries, AI systems have the potential to revolutionize sectors like healthcare, finance, and manufacturing.
But AI’s expansion will require immense energy, far surpassing the current needs of modern computing. For instance, while a single Google search uses about 0.3 watt-hours of energy, a ChatGPT search consumes around 2.9 watt-hours – nearly ten times as much.
Meeting these rising energy demands will spur increased processor development and substantial growth in supporting infrastructure, including physical data centers, power generation, and data center cooling.
And this growth is expected to accelerate. Power demand for data centers is projected to triple from 2022 to 2030. By the next decade, data centers will consume around 9% of the US power grid’s energy, with over half of that supporting generative AI.
This surge in energy consumption poses critical questions for investors: How will we generate the energy required to power data centers, how do we secure the necessary real estate for additional equipment, and what hurdles will we need to navigate during this transition?
How Are Data Centers Powered?
Data centers are crucial to the advancement of AI. Not only do these facilities provide the massive computing power, data storage, and high-speed processing required to run complex AI algorithms, but they also train models and handle the vast amounts of data that AI applications rely on.
Nearly all data centers run on electricity, making it essential to have electric utilities nearby. These utility companies procure their power from a mix of renewables and natural gas.
Despite many large data center operators committing to carbon neutrality, the high demands of these facilities – which run near-to-peak capacity throughout the day – means that nearby renewables are not enough to satisfy power demands. To compensate, operators often turn to natural gas, the most common fossil fuel technology and one which is accessible in most of the locations in the U.S.
As a result, solar, wind, and gas will play an outsized role in the AI revolution. By 2030, its estimated that solar and wind demand will grow by 20% and 39%, respectively, with natural gas demand growing around 4%. The net result is a shared energy burden over the medium-term, with gas and renewables each making up around 50% of power needs by 2030.
However, the concrete net emissions goals at major tech companies are expected to accelerate the growth of renewable energy over the long term. For these major players, simply buying unclean electricity from the grid will not be enough to satisfy clean energy targets.
To meet ESG goals, firms will have to develop a comprehensive carbon neutral strategy that offsets emissions with renewables. While this might involve installing onsite generation or purchasing renewable energy certificates (RECs) or green energy tariffs from utilities, virtual power purchase agreements (PPAs) have emerged as a favorite option for tech firms.
This practice – which occurs when firms purchase renewable power credits from assets that might be in a different grid – was utilized by 70% of all corporates in the first half of 2024.
Where Can We Build Them?
Determining where to build a data center is dependent on several factors: timely permits for power connections, proximity to urban hubs and power sources for generative AI queries, state subsidies, access to fiber infrastructure, and net-zero goals.
Despite these diverse considerations, electricity generation costs and availability remain paramount. If electricity isn't accessible, a data center cannot be built. Water sources are also significant – a typical mid-sized data center in the U.S. consumes around 300,000 gallons of water daily, equivalent to the consumption of 300 households.
Due to its proximity to major cities, access to reliable, affordable power, and business-friendly environment, the Pennsylvania-New Jersey-Maryland (PJM) region remains a key area for data center development. Dominated by Northern Virginia, this region housed 25% of U.S. data centers as of 2022 and is projected to drive 41% of total data center growth.
But key PJM limitations – such as surging real estate costs and the northeast’s limited access to low-cost renewables – could see data center development shift to other areas in the future. For example, the Midwest does not need to source natural gas exclusively from Appalachia, like the PJM region, and could look to multiple areas, such as the Rockies or Canada, for gas resources. Texas and the Southeast region, while often lacking proximity to major networks, could benefit from governments sympathetic to long-haul pipelines and oil and gas infrastructure development.
However, comparing regional differences often fails to weigh the considerations of citizens who are tasked with shouldering the financial burden of grid upgrades. Typically, when new generation assets are added, such as those providing power exclusively to data centers, utilities spread the costs across all customers.
This has led to a growing question of fairness: Should only the data centers bear these costs, or should they be socialized across the general customer base? Additionally, adding more generators could drive up commodity prices, indirectly raising costs for everyone. Balancing these factors is key to determining who should ultimately pay and where data centers should be placed.
The Impact of Bottlenecks
While grid upgrades are necessary, they will not be easy. Strong opposition from anti-fossil fuel state and local stakeholders continues to pose a significant challenge to gas infrastructure development. Other regions are hesitant to add solar and wind projects, citing environmental and reliability concerns.
However, grid bottlenecks and equipment supply issues – which lead to longer approval and lead times – remain the main obstacles to data center growth. Currently, critical grid equipment like transformers and generators are backlogged two to three years. Assuming these parts are ordered in time, it takes approximately three to five years to build new renewable or natural gas power projects from scratch.
For example, a company seeking to build a data center in the Northeast must consider the shortage of pipelines built to cross the Appalachian Mountains. Even though Western Pennsylvania has abundant supplies of natural gas, this bottleneck will make the construction of a new data center difficult – meaning that the firm in question would have to search for operators with spare capacity.
While there’s no question that AI is here to stay, much is still unknown about its rollout and future potential. Slower adoption of AI applications, reduced energy needs for AI tasks, energy efficiency improvements at data centers, and the emergence of alternative power sources could all influence generative AI power demand and the number of data centers needed to be built.
Additionally, the outcome of a hotly-contested presidential election is still uncertain; one which could impact federal subsidies for renewables and permitting and interconnection for natural gas projects. Whatever developments emerge in the macroeconomic or political space, it’s clear that data centers will play an increasingly critical role in supporting this technology, making it essential for firms and investors to carefully plan their energy strategies and balance the growing demand for data processing with sustainable energy.