AI's Thirst for Power: Nuclear Deals Signal a Shift in Energy Strategy Amidst Rising Costs

In a significant development, two of the world's top three cloud computing companies have secured long-term contracts for nuclear electricity. Amazon has committed to purchasing 690 megawatts from Constellation's Calvert Cliffs plant in Maryland, which includes an additional 190 megawatts of generating capacity expected between 2030 and 2032. Google has also finalized an agreement to bring 890 megawatts of new nuclear capacity into the United States. Constellation is investing over $4.3 billion to upgrade 11 reactors across Illinois, Pennsylvania, and New Jersey. Notably, these agreements do not involve the construction of new reactors but rather aim to increase output from existing facilities, some of which predate the contracting companies.

This trend marks a notable reversal from previous years when nuclear power faced significant public and political opposition following accidents at Three Mile Island, Chornobyl, and Fukushima. Germany and Japan, for instance, had phased out or significantly reduced their reliance on nuclear energy. However, the landscape is shifting, with even a previously damaged reactor at Three Mile Island being prepared for service, with Microsoft agreeing to purchase its output for two decades. The US Department of Energy is supporting the restart of three shuttered plants with up to $4.4 billion in loans, aiming to restore approximately 2.25 gigawatts of capacity. The Palisades plant in Michigan, which ceased operations in 2022, is slated to resume generating power by March 2027, potentially becoming the first US commercial nuclear plant to restart after entering decommissioning.

The industry's initial approach to AI's burgeoning energy demands focused on efficiency improvements, such as more advanced chips, smaller models, and smarter cooling systems. While these advancements increased computational performance per watt, they did not curb overall electricity consumption. Instead, efficiency gains were absorbed by larger, more demanding AI applications, including continuous background agents and video generation. This phenomenon echoes William Jevons' 1865 observation that more efficient steam engines led to increased coal consumption due to newly viable applications. The International Energy Agency projects that global data center electricity demand will more than double to around 945 terawatt-hours by 2030, even with ongoing efficiency improvements.

Faced with the inability to shrink demand, tech companies are now focusing on securing energy supply. However, scaling up electricity generation is a slow process. New large reactors can take over a decade to become operational, and while tech companies are investing in small modular reactors, many are not expected to deliver power until the early 2030s. Reactor restarts and upgrades offer a faster solution, but even these timelines extend to 2027, 2028, or beyond, creating a mismatch with the rapid pace of AI hardware development and capital expenditure.

The financial implications of these long-term energy contracts are significant, particularly within grid operator regions like PJM, which serves 13 US states and Washington, DC. PJM utilizes capacity auctions to ensure future electricity supply, with costs passed on to consumers. Data center demand has substantially contributed to these capacity costs, with an estimated $29.4 billion across the last four auctions attributed to data centers, including projected demand from future facilities. This has led to increased electricity bills for households, with projections suggesting a potential rise of $70 per month for typical households in the PJM region by 2028 if current trends persist.

These rising costs are fueling public dissatisfaction and prompting regulatory responses. Virginia has introduced a separate electricity rate class for data centers, Ohio has implemented minimum billing requirements, and Illinois is considering new legislation. Google's agreement aims to address these concerns by ensuring new major electricity consumers contribute additional generation capacity rather than straining existing supply. While this voluntary approach is a positive step, it follows a rejection by US federal regulators of PJM's proposal to mandate such requirements. The sustainability of these voluntary commitments remains uncertain if costs continue to escalate.

For Gulf countries planning their own AI infrastructure expansion, the US experience serves as a cautionary tale. While abundant energy resources offer a competitive advantage, the scale of proposed AI data center developments warrants careful consideration of opportunity costs. Qatar, Abu Dhabi, and Saudi Arabia are making significant investments in AI infrastructure, but the electricity demand for these projects could represent a substantial portion of their national generating capacity. The Gulf region has a unique opportunity to define the terms of AI infrastructure development, including industrial tariffs, funding for new generation, and long-term economic benefits, before large-scale commitments limit their options.