AI: The Energy Saga

Generative AI relies on supercomputers to perform complex calculations and answer increasingly intricate questions. It is thus highly dependent on advanced semiconductors (chips), large data sets and powerful algorithms (models). The energy requirements of these processors, which are typically housed in datacentres, are enormous and require stable and secure power sources.

According to the International Energy Agency (IEA), energy for data centres, cryptocurrency, and artificial intelligence accounted for roughly 2% of global energy demand in 2022. The IEA forecasts that the energy demand could almost double from 460 TWh in 2022 to about 800 TWh by 2026. However, these estimates may vary significantly depending on the actual demand for generative AI products and services over the next two years.

Cloud computing giants like Amazon, Microsoft, and Alphabet are heavily investing in their data centres to enhance their capabilities for running complex AI applications. This involves training and inferencing (producing outputs) with AI models. Consequently, their energy consumption and greenhouse gas emissions have risen. According to Alphabet’s 2024 Environmental Report, their greenhouse gas emissions have increased by 48% since 2019. The growing demand for AI, combined with Alphabet’s termination of some clean energy projects, has led to this overall increase in its emissions.

Why haven’t companies simply expanded their energy capacity? Data centres require a stable energy supply that can be adjusted instantly. Renewable energy is not always reliable due to its dependence on unpredictable elements. Nuclear power can provide the needed stability, but building nuclear plants takes more than five years. In contrast, gas-fired plants take about two years to build.

The cloud providers are exploring various ways to increase their access to stable energy sources. Earlier this year, Amazon bought a data centre from US power generator Talen Energy, located next to a nuclear power station. The data centre will be powered with energy from the station.

Microsoft on the other hand, signed a power purchase agreement with Brookfield Asset Management, to support the development of 10.5 gigawatts of renewable energy and to secure their energy requirements and offset their carbon footprint.

At the core of these new data centres are semiconductors responsible for all processing tasks. Nvidia claims that their flagship Blackwell Platform will enable organizations to “build and run real-time generative AI on trillion-parameter large language models at up to 25x less cost and energy consumption than its predecessor.” No wonder everyone is racing to get their hands on these chips.