USA – WASHINGTON Trends and Developments Contributed by: John Pierce and Patrick Njeim, Kilpatrick Townsend & Stockton
Kilpatrick Townsend & Stockton LLP 1420 Fifth Avenue Suite 3700 Seattle, WA 98101 USA
Tel: +1 206 626 7726 Fax: +1 206 374 8224 Email: JFPierce@ktslaw.com Web: ktslaw.com
AI Data Centres and the Looming Energy Crisis in the United States The accelerated proliferation of artificial intelli- gence (AI) technologies has ushered in a new era of innovation and productivity across vari- ous sectors. However, alongside these trans- formative advancements lies a growing chal- lenge – the rapidly increasing energy demand stemming from AI data centres. These facilities, which power large language models (LLMs) and generative AI applications, are becoming major consumers of electricity in the United States and are threatening to outpace the country’s current power infrastructure, the necessary grid infra- structure, and national and state policy capa- bilities. Use case example A prompt to prepare a graphic for a presentation was as follows: “Hello AI model, I would like to generate an image of a data centre showing a layout of servers.” The energy required for this simple task can range from 0.01 to 0.29 kilo- watt-hours (kWh), depending on the AI model, according to recent studies. To contextualise, this is comparable to charging a smartphone from empty to full.
Text-based queries exhibit similar patterns. For instance, prompting a generative AI model with “AI model, tell me how much energy AI uses” consumes far more energy than a typical online search. While a Google search uses about 0.3 watt-hours, generative AI models consume ten to 30 times more per prompt. This highlights the substantial increase in power needed for advanced AI computation. The scale of AI usage is staggering. Over 30 million new images and 3.5 billion searches are generated daily, rapidly escalating electricity demand. In addition to operational use, training these models also requires significant energy. For example, training the GPT-3 model is esti- mated to have used about 1,300 megawatt- hours (MWh) of electricity in a month. For the more complex GPT-4, greater energy require- ments are expected, with estimates suggesting that the hardware footprint was 20 times greater, and the training spanned three months. These figures underscore the growing environmental and infrastructural challenges of large-scale AI. AI models fall into “generic” and “specialised” categories. Generic models, like LLMs, are designed for a wide range of tasks, typically
372 CHAMBERS.COM
Powered by FlippingBook