Economic Viability of Nuclear-Powered AI Data Centers : A Study Examining Hybrid Energy Strategies in Sweden

Sammanfattning

The large-scale introduction of artificial intelligence is rapidly increasing electricity demand from data centers, particularly in Sweden, where hyperscale connection requests are placing increasing pressure on the electricity system. Limited grid capacity and long connection queues create uncertainty for new establishments, which has increased interest in hybrid energy solutions. Nuclear power is relevant in this context due to its ability to provide stable, low-carbon electricity with high supply reliability. The purpose of this thesis is therefore to examine whether nuclear-powered hybrid solutions are economically attractive for AI training data center operators in Sweden compared with electricity procurement from the spot market. This thesis evaluates three electricity sourcing strategies for hyperscale AI training data centers in Sweden: a Mankala ownership structure, a Corporate Power Purchase Agreement (PPA), and direct electricity procurement from the wholesale spot market. The analysis is limited to Sweden’s SE3 electricity price area and considers the AP1000 reactor design by Westinghouse. More specifically, the aim is addressed by calculating the levelized cost of electricity (LCOE) and risk measures for the different electricity sourcing strategies in order to determine which alternative is most favourable for a hyperscale AI training data center. The LCOE is calculated from the data center’s perspective and includes all costs related to each electricity sourcing alternative. These costs include, for example, spot market purchases, nuclear construction costs, operation and maintenance costs, grid fees, electricity taxes, balancing fees, Guarantees of Origin, and grid connection fees. Since there is considerable uncertainty regarding several input parameters, such as WACC and nuclear construction costs, the LCOE is calculated for 1,000 future scenarios. Based on these scenarios, the expected LCOE and risk measures are calculated and used to compare the three alternatives. Sensitivity analyses are also performed for electricity tax assumptions, electricity price volatility, and weighted average cost of capital. The results show that the Mankala case provides the lowest expected LCOE, while the Corporate PPA case performs similarly but is slightly more expensive. The spot market case has the highest expected cost, but lower downside risk, making it more attractive for risk-averse data center operators. The results also show that the reduced electricity tax on electricity supplied from nuclear power is crucial for making the nuclear-based alternatives competitive with spot market procurement.