Challenges and Opportunities at the Intersection of AI and Energy

The global race to secure AI leadership is radically reshaping the U.S. electricity sector and the contours of global energy policy. This race has commercial and geopolitical dimensions, with private firms pursuing an expanding market and governments securing a critical element of national security. Both perceive potentially existential stakes in victory or failure. Competitiveness for both states and firms requires ever-larger datacenters, now routinely housing tens of thousands of state-of-the-art GPUs and consuming electricity at a scale unmatched by all but the most energy-intensive industries on the planet. Paradoxically, though AI is digital in nature, the fundamental constraint on technological progress is physical infrastructure.

This paper explores the evolving forecasts of U.S. electricity demand, with a particular focus on the explosive growth of datacenters. It introduces the concept of “speed-to-power”—the speed at which a given datacenter can secure enough electricity supply to operate—and analyzes the key bottlenecks shaping the commercial power landscape. We also examine the emerging technologies and business models that could help overcome these challenges. Ultimately, we argue that the U.S. power sector must strike a careful balance: racing to add capacity in the near term while laying the foundation for “sustainable scaling” over the long run. Meeting these rising energy demands is not only a matter of economic efficiency or opportunity, it represents the foundation of U.S. advantage in an AI-driven global economy.

In the United States, a wide range of data, studies, and reports point to staggering growth in datacenter electricity demand. CSIS modeling of the sector indicates AI-dedicated datacenters could jump from 4 gigawatts (GW) of demand in 2024 to 84 GW by 2030, representing an astonishing 2,100% increase. Today the entire datacenter sector represents 20 GW of capacity and consumes an estimated 4.4% of US electricity, which could grow to 12% by 2028 according to a recent study^*1 from Lawrence Berkeley National Laboratory (LBNL). Estimates vary widely given the rapid evolution in AI technology, uncertainty about the commercial value of AI, and questions over the sustainability of capital expenditure at leading technology firms like Microsoft, Meta, Google, Amazon and Oracle, collectively referred to as hyperscalers. But despite this uncertainty, growth for the power sector is clear. Even the lowest plausible estimates for datacenter growth given already underway investment, roughly 50 GW by 2030, would represent a more than doubling of US datacenter demand.

*1

https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf

Beneath the topline national figures, the pattern of geographic distribution will determine how this boom affects the electric-power sector. Virginia, and the northern Virginian computing cluster in particular, remains the largest market for datacenters in the world and is slated to continue to see rapid expansion, with datacenter demand in the state growing from 5 GW in 2024 to upwards of 20 GW by 2030. However, with this market increasingly power supply-constrained, datacenter projects are reportedly facing 7 years or longer supply wait times, other markets are attracting large volumes of datacenter investment. Texas stands out positioned to attract over 10 GW of datacenters. States like Louisiana, Mississippi, and New Mexico are attracting large gigawatt scale datacenter investments starting from a baseline of zero. Datacenter investment is clearly flowing to sites and states where permitting and power-sector policies enable rapid investment in grid and generation assets to serve the hyperscalers. By contrast, the Northeast states (New England and New York) combined represent 15% of US GDP but are slated to receive less than 5% of datacenter investment.

The ongoing rapid evolution in the domain of AI model design translates into uncertainty about the structure of the datacenter fleet expansion. For example, if large pre-training runs continue to be crucial to performance progress, then ultra large, multi-gigawatt scale training optimized facilities are crucial. If instead, performance progress shifts to post-training and reinforcement learning techniques, then the inference-oriented facilities would grow relatively more valuable. Meanwhile all expectations are for provision of service oriented “inference” computation demand to grow to dwarf computation dedicated to model training. As the commercial landscape for AI applications and services matures, the line between post-training inference and provision of service inference could grow blurry.

While the overall commercial environment at the datacenter-power intersection reflects a rush to add capacity as quickly as possible, we believe the expansion will ultimately proceed at a moderated, slower for longer pace. The political and physical limitations of deploying power-sector infrastructure act as a moderating brake on demand growth. The “speed-to-power” era will evolve into an era of long-term scaling with both private and federal government investment oriented towards long term competitive advantage in the power sector.

In the immediate moment the “speed-to-power” mandate is pushing a handful of proven technologies to the front of the line. Solar, gas turbines, and battery storage stand out as the leading choices because of siting flexibility, scaling to meet large loads, and integration into existing grid structures with minimal delays.

Wind power, historically a major contributor to renewable capacity growth in the U.S., is increasingly facing a divergent path from solar. Transmission system congestion in the nation’s most attractive wind production regions – the Plains and Midwest markets of SPP (Southwest Power Pool) and MISO (Midcontinent Independent System Operator) – create significant barriers in terms of interconnection costs and delays. Wind deployment is increasingly contingent on the completion of large-scale high-voltage transmission projects and ,as a result, will struggle to scale in response to the near-term demands for capacity from datacenters.

Nuclear power, though essential to long-term decarbonization plans and the subject of considerable public attention, is likewise poorly positioned to serve near-term demands. Though two nuclear reactor restarts are underway with a third under consideration – this represents the limit of that opportunity in the U.S. given that all other reactors are too far along in deactivation. Advanced reactor designs, such as those supported by major AI firms like Google and Amazon, remain a nascent technology still in the development phase. Several first-of-a-kind deployments of such reactors will likely be completed by 2030; after accounting for some design iteration based on construction and operational lessons, real commercial scaling of this technology is unlikely until 2035 or beyond. The only nuclear technology licensed and ready to deploy in the U.S. is the AP1000, but the sheer price tag of such investments, risk of cost overruns, and long construction timeline have hindered commercial prospects.

These realities leave utilities and independent power producers focused on what’s fast, financeable, and already proven. Solar and battery storage show immense promise with deployment rates growing across the country. In 2024 solar and battery projects represented 75% of all projects in interconnection queues across the country (by nameplate capacity), with hybrid and co-located projects representing nearly half this volume. Across multiple states, utilities, and policy paradigms the U.S. power sector is relying on natural gas, a key source of firm power capacity, deployed in a portfolio with solar and storage.

Datacenter developers and the power suppliers are exploring new commercial models that might deliver faster speed to power for these preferred technologies. A leading candidate is co-location, where datacenters are built adjacent to large generation capacity. Co-location of datacenters at existing nuclear power plants was attempted by at least one firm, but federal regulators intervened as they consider reliability and cost allocation implications for other consumers. Co-location of datacenters with new build gas, solar, and storage in various combinations is an increasingly common pathway. Fully islanded gigawatt-scale “microgrids”, again employing gas, solar and batteries in various combinations are under development and may prove successful because they avoid traditional grid interconnection and policy risk altogether.

Another emerging commercial tool to deliver faster speed to power is demand-side flexibility where datacenters shift workloads to off-peak times to ease grid stress. Historically datacenter operators have declined to participate in wholesale demand response programs because the value of their services is orders of magnitude larger than the price signal sent by the power market. But as the value of speed-to-power grows, curtailment commitments may be justifiable. For example, a commitment by datacenters to reduce demand during peak load periods could obviate the need for grid upgrades or new generation resources, which could allow faster interconnection by the electric utility.

Looking beyond the next five years, a broader array of power supply solutions will become available. By this time, the commercialization of advanced nuclear reactors will have progressed. Advanced geothermal power also shows promise in overcoming its initial commercialization hurdles, potentially offering a clean dispatchable alternative to gas generation. Gas plants equipped with carbon capture technology are another avenue for reducing the sector’s carbon footprint without sacrificing the flexibility and reliability of natural gas. At least two large oil and gas companies have signaled plans or willingness to build co-located, gigawatt-scale, gas generation with integrated carbon capture.

As these approaches mature, datacenter operators and utilities will be able to choose from a wider solution set that can support continued demand growth even as efficiency gains, hardware constraints, and business realities reshape the pace of AI’s expansion.

The electric-power sector is defined by long-lead time investments with very long operational lifetimes. High-voltage transmission lines or nuclear power plants can comfortably operate for 60 years or more. This reality clashes with an AI technology sector desperate to scale datacenters at the pace of AI technology advancement, where performance breakthroughs are achieved month-by-month. This temporal mismatch poses an acute challenge for the power sector in the near term.

These near-term challenges risk obscuring a profound shift in the relationship between electricity and overall economic growth. AI technology creates a new vector for continuous expansion of computation demand, and by extension demand for electricity, in the economy. As AI-enabled productivity gains drive economic growth, and the demand for computation scales, the electricity intensity of the U.S. economy will likely reverse a long-term decline. The AI boom is but one of several factors contributing to this trend. Semiconductor fabrication and battery manufacturing are two highly electricity intensive, strategically vital industries slated for growth and backed by strong policy support. New semiconductor fab clusters in Arizona and New York could each consume nearly one gigawatt of electricity at full operational scale. Finally, across the economy an underlying trend of electrification—in transport, heating, and industrial applications—is shifting more energy consumption into electricity. All together these factors point to an electric-power sector increasingly central to economic outcomes and U.S. policy objectives.

Utilities, independent power producers, and supply chain players all need to adapt to a new paradigm of long-term sectoral growth. Alongside obvious investments necessary to support such growth, a deeper challenge will be shifting cultural norms towards innovation and dynamism at utilities, regulatory bodies, and independent grid operators. At the level of the federal government, a core question is whether decades of energy security policy primarily oriented around oil markets can adapt to a new era in which electricity supply is the cornerstone of long-term national strength. The strategic thesis in electricity that AI makes electricity the new central domain of global energy advantage implies bold new thinking in the policy domain is needed.

The first and most pressing national policy challenge is the high-voltage electric transmission system. Without expanded grid capacity it is impossible to add new generation and demand at the rate required to secure AI leadership. In many parts of the country, generation is done in a competitive market, or it is otherwise simple to allocate new generation to new demand of specific customers. In contrast, transmission is a quasi–public good. Investments in high-voltage transmission spread benefits to all customers which make cost allocation, be it to a specific datacenter customer, or to a specific state’s ratepayers, a politically challenging process. High-voltage transmission is flexible, supporting future demand expansions, new generation preferences, and economic growth patterns to proceed unconstrained despite the inability of system planners to predict specific outcomes. This strategic option value is a public good and is undervalued by existing regulatory procedures. New long-term transmission planning and greatly increased investment must be undertaken with the strategic thesis in mind.

The second policy challenge is affordability. Clearly investment in both generation and transmission must grow to serve growing datacenter demand. In general, rising investment equates to rising rates for customers, at least in the near term. The exact mechanism by which generation and grid investment costs are socialized across ratepayers depends on state policy frameworks. But wherever it occurs, political backlash from rising electricity rates can undermine social license to pursue AI leadership. Increased federal funding for upgrading and expanding high-voltage lines serves as a solution to this problem. Such funding works to “buy down” infrastructure costs that would otherwise burden ratepayers. By keeping electricity rates more stable, enabling investment to proceed, and ensuring smooth long-term expansion of the power sector, strategic investment in the high-voltage transmission grid becomes a linchpin of an AI leadership strategy.

Finally, the long-term challenges of reducing emissions and responding to climate change will remain. The immediate needs of AI will take precedence over emissions reductions, there is no way around that, and it is the right choice for the country. But resolving the load growth challenges of AI will catalyze a new era of power sector investment and open up pathways for long-term decarbonization that have been imagined, but rarely plausible. AI technology could solve grid optimization problems that have long plagued renewables deployment. If we cannot scale the system for AI, we will never succeed in decarbonizing the economy. In contrast, a growing economy and rapidly advancing technology frontier bring real decarbonization progress within reach.