Avoiding Waste Heat through AI Infrastructure Thermal Integration

Globally, data center energy use is projected to grow from about 415 TWh, equivalent to 1.5% of global electricity demand in 2024, to 945 TWh, equivalent to 3%, by 2030 (IEA 2024). Similarly, worldwide annual water consumption associated with the growth of data centers and artificial intelligence (AI) is expected to reach 4.2–6.6 billion m³ by 2027, roughly four to six times Denmark’s annual water demand (Privette 2024). In the United States, data centers could account for 6.7–12% of total electricity use by 2028 and consume 0.14–0.28 billion m³ of water by 2028 (Figure 1) (Shehabi et al. 2024).

Data centers are high-power-density facilities in which nearly all electricity is converted into heat. Servers and chips act as large resistive heaters, meaning each kilowatt-hour of electricity produces roughly 1 kWh of heat, of which approximately 70–80% could be recoverable (IEA 2024). However, most facilities expend additional energy and water to dissipate this heat to the environment instead of reusing it. Although AI growth is driving higher power densities, non-AI expansion such as additional cooling infrastructure adds to energy and water demand. Power availability and water stress are emerging as primary bottlenecks to data center development, highlighting the need to rethink data center efficiency and the related socioeconomic impacts. Encouraging heat reuse can lower energy and water consumption, reduce permitting burdens, improve public perception, and benefit local communities and businesses.

The industry average power usage effectiveness (PUE) in 2023 was approximately 1.4, down from 1.6 in 2014 (Shehabi et al. 2024). This means that only about 70% of electricity consumed is used to power computation (1 ÷ 1.4 ≈ 0.7), with the remainder used primarily for cooling, which can account for up to 80% of the noncomputational load. Advanced cooling technologies, spurred by high-power AI chips whose power densities have increased from nearly 200 W to more than 1,000 W (Díaz-Marín and Berquist 2025), aim to reduce thermal resistance through liquid and two-phase cooling and improved thermal interfaces. These approaches both lower PUE and raise the temperature of dissipated heat from about 30°C (typical of air cooling) to more than 50°C, making waste heat more suitable for reuse. As AI-driven data centers continue to generate heat at increasingly higher temperatures, opportunities to integrate them into broader energy- and water-efficient ecosystems also grow.