USA – WASHINGTON Trends and Developments Contributed by: John Pierce and Patrick Njeim, Kilpatrick Townsend & Stockton
using graphics processing units (GPUs), opti- mised for large data volumes. Specialised mod- els, by contrast, are tailored for specific tasks and can be more efficient, running on GPUs or customised chipsets. Generic models use much more energy than specialised ones – often ten to 50 times more – during both training and deploy- ment. As both types of AI models become more preva- lent, the demand for electricity driven by GPUs and specialised hardware continues to increase, raising important considerations for energy infra- structure, regulatory compliance and policy. The rapid rise of AI-driven energy demand and electricity consumption trends As of 2023, US data centres consumed approx- imately 4.4% of the country’s total electricity production. Projections by the Department of Energy (DOE) suggest this figure could rise to between 6.7% and 12% by 2028, primarily driv- en by the computing intensity of AI workloads. LLMs, natural language processing engines, and other generative AI tools require vast amounts of data and continuous high-performance process- ing, contributing heavily to this growth. From a historical lens, energy consumption by data centres in the United States has almost tripled in less than a decade, from 76 terawatt- hours (TWh) of electricity in 2022 to more than 176 TWh in 2023 (marking a 131.6% increase). Looking forwards, AI-related electricity demand could increase to more than 580 TWh by 2028. This reflects a significant shift and challenge in the country’s electricity production, transmis- sion grid and consumption landscape, as well as the energy planning and policies of the state, regional and national policy makers (for perspec- tive, the global data centre electricity usage in 2022 was about 460 TWh and was projected to
exceed 1,000 TWh by 2026, more than doubling the 2022 figure). Efficiency plateau and cooling requirements For more than a decade, data centre operators have relied on advances in server performance, virtualisation, workload optimisation and cool- ing technologies to meet growing demand with- out a proportional rise in energy consumption. Innovations in chip design – such as improved instruction sets, energy-efficient central pro- cessing units (CPUs) and GPUs, and dynamic workload scaling – allowed data centres to man- age increasingly complex tasks while maintain- ing a stable power profile. Similarly, improve- ments in infrastructure – such as high-density server configurations, airflow management, and the adoption of hot and cold aisle containment systems – helped reduce waste and maximise thermal efficiency. Cooling technologies have also evolved. Tradi- tional air-cooling systems, which rely on ambient air and fans to maintain safe operating tempera- tures, were gradually supplemented or replaced by more efficient systems, including liquid cool- ing, immersion cooling and evaporative systems. Among these, water cooling emerged as a lead- ing solution, offering better thermal conductivity than air and the ability to cool high-density serv- ers more effectively, especially in AI workloads that generate concentrated heat over sustained periods. However, these efficiency improvements have plateaued since 2020. The underlying reason is the exponential increase in computational intensity introduced by AI. AI training, particu- larly for LLMs and generative networks, involves processing massive datasets through complex neural networks. These models require custom hardware accelerators, such as NVIDIA’s B100
373 CHAMBERS.COM
Powered by FlippingBook