USA – WASHINGTON Trends and Developments Contributed by: John Pierce and Patrick Njeim, Kilpatrick Townsend & Stockton
(Blackwell) GPUs or Google’s TPU v4 clusters, which can draw several kilowatts per server rack. As AI workloads scale, traditional gains from Moore’s Law and energy-saving algorithms are increasingly insufficient to offset the power draw. The associated thermal loads from AI operations have surpassed the capabilities of many legacy cooling systems, prompting a rapid transition to liquid-based solutions. Water cooling, which uses chilled water circulated through pipes, plates or direct-to-chip systems, provides far greater heat removal capacity. However, it also raises significant environmental and operational challenges. Hence, water usage in data centres has become a major concern in arid regions such as the southwestern United States, where competing demands from agriculture, residential use and industry already stress local water sup- plies. For example, some hyperscale data cen- tres can consume millions of gallons of water per day for cooling, especially during peak load conditions or in older facilities without closed- loop systems. Furthermore, sourcing, treating and discharging large volumes of water require infrastructure that not all regions possess or can sustainably support. To address these issues, some operators are investing in waterless cooling technologies, such as direct-to-chip liquid cooling, refriger- ant-based cooling and phase-change materials, which offer improved sustainability. Additionally, AI-driven facility management and digital twins have been deployed to fine-tune energy usage and thermal flows in real-time. Despite these innovations, the balance between performance, energy efficiency and environmental sustainabil- ity remains a concern.
Thus, while historical efficiency gains in serv- er hardware and cooling enabled data centre expansion without parallel growth in energy use, the rise of AI has disrupted this equilibrium. Meeting future AI demands will require a new generation of energy- and water-efficient tech- nologies, initiative-taking policy frameworks, and sustainable infrastructure planning to avoid exacerbating ecological and utility grid pres- sures. Regional impacts and grid stress The rapid expansion of AI, particularly through data centres housing powerful computing clus- ters, is placing an unprecedented load on the United States’ electrical grid, a system origi- nally designed for a far less energy-intensive economy. These demands are not only increas- ing the total amount of electricity required but are also altering when and where that electricity is needed, exacerbating existing strain points across the grid. The United States’ power grid is essentially composed of three major interconnections – Eastern, Western and Texas (ERCOT) – which were largely constructed in the mid-20th cen- tury and have seen only incremental upgrades in the decades since. While these systems have served traditional industrial and residential needs well, they were not engineered to sup- port the explosive, concentrated energy draw characteristic of modern AI data centres. Unlike conventional load growth, which occurs gradu- ally and is geographically dispersed, AI-related energy demands are sudden, large-scale and often localised, overwhelming local distribution systems and creating severe capacity shortfalls. One of the most pressing challenges lies in the limited capacity of existing transmission lines. High-voltage transmission corridors are essen-
374 CHAMBERS.COM
Powered by FlippingBook