Is Nuclear Energy Big Tech’s “Silver Bullet” to Power Energy-Intensive AI Data Centers?

Essential components to successfully scale High-Performance Computing (HPC) and Artificial Intelligence (AI) are big data, technology experts, advanced processing microchips, and terawatts of reliable energy. Energy is becoming the biggest bottleneck to building the next generation of AI Data Centers (DCs). This is resulting in the search for low carbon and reliable energy sources. Nuclear energy, including nuclear fusion and Small Modular Reactors (SMRs), is considered carbon free, reliable, and consistent (24x7) compared to solar or wind energy, which relies on the weather. Thus, gaining Venture Capital (VC) investment is a viable option to support big tech’s AI ambitions.

• Constellation Energy and Microsoft have signed a power deal to restart the 837 Megawatt (MW) Three Mile Island nuclear plant in Pennsylvania by 2028. The Power Purchase Agreement (PPA) will last 20 years, longer than a traditional solar and wind PPA, and showcases Microsoft’s efforts to decarbonize the grid.
• Amazon Web Services (AWS) acquired Talen Energy’s 960 MW Cumulus DCs for US$650 million earlier this year, a 10-year agreement to power AWS’s new DC. Cumulus Data Assets is part of a 1,200-acre Susquehanna Steam Electric Station campus powered by 2.5 Gigawatts (GW) of nuclear power.
• SMRs are gaining popularity and being explored by the likes of Oracle, OpenAI, and Azure. Kärnfull Next, a Swedish nuclear company announced plans of a commercial campus at Nyköping, while Google announced plans to purchase 500 MW of nuclear energy from Kairos Power’s SMR by 2030.

According to the U.S. Department of Energy (DOE), nuclear power has been used for the past 60 years to support the energy grid and national defense under strict regulations. The invention of SMRs—compact, cheap to manufacture and transport nuclear reactors—is changing the game. They can typically fuel submarines and aircraft carriers for up to 20 years with few maintenance requirements. The International Atomic Energy Agency (IAEA) defines “micro” reactors as up to 20 Megawatts Electric (Mwe), “small” as 300 MWe and “medium” up to 400 Mwe, with DCs piloting series of micro and small SMRs. Partnerships between DC operators and nuclear plants are proving to be mutually beneficial through tax rebates, utilizing existing infrastructure, and reducing pressure on the energy grid.

On the contrary, lawmakers and regulators are raising concerns about the cascading impact of diverting nuclear energy for DCs. Energy experts stress that the partnership with DC operators will reduce grid reliability, increase costs for consumers, and increase overall emissions. For example, in Philadelphia, Exelon and American Electric Power challenged the Federal Energy Regulatory Commission that AWS and Talen Energy’s partnership would result in US$140 million in annual transmission costs being transferred to local consumers. Besides the trickle down on cost, climate activists are stressing a proactive approach to increasing investment in renewable PPAs, battery storage, and geothermal to increase grid capacity sustainably.

ABI Research forecasts that hyperscalers will consume 381 Terawatt Hours (TWh) of low carbon energy (renewables and nuclear energy) by 2030, in line with the self-regulatory Climate Neutral Data Centre Pact (CNDCP), with a pledge from industry players, trade associations of cloud infrastructure services, and DC operators to achieve climate neutrality by 2030. Currently, the top 20 hyperscalers consume or offset 88% of their total energy consumption with renewables. Amazon, Alibaba Cloud, Equinix, Google, Meta, and Apple operate on 100% low carbon energy with Global Switch, Equinix, and Microsoft Azure operating with 96% of low carbon energy capacity. The CNDCP signatories have pledged a 75% renewable energy consumption match by 2025 and 100% by 2030. Hyperscalers are piggy backing on SMR technologies from Oklo, BWX Technologies, Kairos Power, NuScale Power, X-Energy, Seaborg, and Rolls-Royce to meet their ambitious climate neutrality targets.

Powering an existing DC with nuclear power is not “plug-n-play,” it requires an infrastructure redesign to handle the increased energy demand, regulatory compliance, safety compliance, and stringent waste disposal. Cautious and proactive measures should be taken to address concerns and complexities around public safety, transparency, and long-term environmental impact. To navigate the obstacles posed by nuclear power, a multidimensional effort from Public-Private Partnerships (PPPs) should be achieved with technology experts, industry collaboration, community acceptance, and academia to address the following:

• Public Opinion: One of the biggest hurdles for nuclear power is winning over public opinion on nuclear power. A longstanding perception of safety and waste disposal is a long-standing concern.
• Licensing and Regulatory Risk: Deploying and operating nuclear plants requires the plant operation to adhere to stringent regulations and safety standards. This can pose a logistic challenge and can result in delays.
• Initial Investment: Upfront investment required for nuclear infrastructure is massive, but operating cost is lower due to high Return on Investment (ROI), continuous operation, low maintenance, and comparatively lower carbon emissions.
• Waste Disposal: Effective waste disposal and storage is a paramount concern for public health and the environment. Nuclear waste is currently stored in barrels, which will be safe for reuse in potentially 5 to 10 years, but this process is currently expensive.

To gain public trust, nuclear-powered DC operators should stress transparency, independently verified Environmental, Safety, and Governance (ESG) evaluation, reporting challenges and success to build the case for long-term impact. For more information on the latest energy trends in DCs, see the forthcoming projects from ABI Research: Optimizing Data Center infrastructure: Market Sizing Thermal Management & Energy Consumption (MD-CSDC-101) and its accompanying Market Data Overview: 4Q 2024  (PT-3252).