AI to go nuclear? Data center deals say it’s inevitable

🌈 Abstract

The article discusses how hyperscalers and major data center operators are turning to nuclear power to meet the growing energy demands of compute-intensive AI workloads. It highlights the partnerships and deals being made between tech giants like AWS, Microsoft, Google, and nuclear power providers to access nuclear energy and develop small modular nuclear reactors (SMRs) to power their data centers.

🙋 Q&A

[01] Hyperscalers and Nuclear Power

1. What are the key reasons why hyperscalers are turning to nuclear power?

• The growing energy demands of compute-intensive AI workloads are outpacing the capacity of existing electric grids, leading hyperscalers to seek alternative sustainable and carbon-free energy sources.
• Nuclear power is seen as a viable option to meet the massive power requirements of large-scale data centers needed for generative AI applications.
• Hyperscalers like AWS, Microsoft, and Google are partnering with nuclear power providers and investing in the development of small modular nuclear reactors (SMRs) to power their data centers.

2. What are some of the specific deals and partnerships between hyperscalers and nuclear power providers?

• AWS purchased the Cumulus Data Assets, a 960-megawatt nuclear-powered data center, from Talen Energy.
• Microsoft, Google, and Nucor issued a request for information (RFI) on clean energy, and Constellation Energy responded with a proposal for using advanced nuclear power at existing nuclear sites.
• Microsoft signed a deal with Constellation to receive up to 35% of its power from nuclear sources for its Boydton, Virginia data center.
• AWS also has a 10-year Power Purchase Agreement (PPA) to access nuclear energy from the Susquehanna site.

3. What are the challenges and concerns around using nuclear power for data centers?

• The Nuclear Regulatory Commission has not yet approved any of the proposed nuclear power projects for data centers.
• There are concerns about managing nuclear waste and spent fuel, as well as the potential for catastrophic incidents associated with nuclear power.
• Some experts believe that small modular nuclear reactors (SMRs) may not be operational until 2030, which could delay the adoption of nuclear power for data centers.

[02] Innovations in Data Center Design and Power Efficiency

1. What innovations are data center operators exploring to address power demands?

• In addition to nuclear power, data center operators are looking at other innovations to improve power efficiency and sustainability, such as:
  • Advances in hardware and chip efficiency
  • Liquid cooling technologies
  • Edge computing to reduce dependency on large, centralized data centers
  • Composable and disaggregated infrastructure to enable more agile and efficient resource allocation

2. How can AI and data center innovations help address power and sustainability challenges?

• AI can be used to more efficiently manage smart grids and technology infrastructure, helping to solve some of the power demand issues it creates.
• Distributed AI architectures and other innovations in data center design and operations are on the horizon to address the power and sustainability challenges.
• There is an expectation that as the true use patterns and value of generative AI are determined, the demand for energy and computing will become more predictable and balanced.

3. What is the role of CIOs in addressing the power and sustainability challenges for data centers?

• CIOs are dependent on the large cloud providers to be the experts in producing ample, redundant, and reliable power, allowing CIOs to focus on their core mission.
• CIOs acknowledge the need for sustainable, carbon-free alternatives to electricity, and support the exploration of solutions like nuclear power and other innovations.
• However, CIOs also note that the hype around generative AI has created artificially high demand, and expect organizations to "right-size" their AI computational power as the true use patterns and value are determined.