Isik C. Kizilyalli

Abstract Title

Data Centers and Artificial (General) Intelligence Energy Consumption in the Next Decade
Energy consumption in data centers accounts for approximately 4% of electricity use in the US in 2022 and is expected to reach 6% by 2026 [IEA 2024]. However, there has been an order-of-magnitude increase in training compute for deep learning models every 2 years since 2010. By some accounts (and based on trendlines in AI progress and development), training clusters requiring 1–100 GW of electrical power buildout for $10–$1000s of billions will be deployed by 2030 [L. Aschenbrenner, 2024], with power appearing to be the major constraint on the supply side. Investments in AI data centers are already expected to exceed $580B in 2025 (this surpasses investments in global oil supply) [IEA 2025] and to reach $1.7T globally in 2027 [The Economist]. By the end of 2027, $3T will have been invested in sum worldwide. Demand for power contracts, power sources (new and old; coal, oil, natural gas, solar PV, wind, nuclear fission/fusion, and geothermal), power transformers, and switchgear will grow, and total electricity consumption for all AI may require doubling of the total electricity generation in the USA by 2030. The power delivery architecture of most modern data centers consists of a line frequency transformer, a low-voltage power distribution network, a centralized backup unit, and chip-level voltage regulators. Strategies to improve energy efficiency range from integrating lower-loss power converters to a complete redesign of the power delivery network. The latter approach often involves down-converting higher voltages at the rack level to reduce transmission losses and the number of conversion stages, but significant technical challenges must be overcome due to congested rack space and concentrated heat dissipation. In any case, over the next decade, all areas of electrical and energy science (power generation, power systems, storage, power electronics, optoelectronics, computer science, semiconductor device technologies, magnetic materials) will be severely challenged in their efforts to meet the ever-growing needs of computing.

Contact

ikizilyalli@gmail.com