BONUS MATERIAL - The Staggering Cost of A.I.

This MIT Technology Review article examines the significant and growing energy demands of artificial intelligence, analyzing its current footprint and future projections.

Overview

This MIT Technology Review article examines the significant and growing energy demands of artificial intelligence, analyzing its current footprint and future projections.

It highlights how the energy required for AI is rapidly increasing due to widespread adoption and more complex tasks, contrasting this with previous data center efficiency gains.

The piece explores the challenges in accurately measuring AI's energy use due to a lack of transparency from leading companies and discusses the environmental impact as data centers often rely on carbon-intensive power sources.

Ultimately, the article suggests that the energy costs of AI are poised to reshape power grids and may lead to increased costs for consumers.

How will the exponential growth of AI usage impact global energy consumption and carbon emissions?

Based on the sources, the exponential growth of AI usage is expected to have a staggering impact on global energy consumption and carbon emissions.

This is described as a significant shift in online life and a marked departure from Big Tech's electricity appetite in the recent past**.**

Here's a breakdown of the key impacts according to the sources:

Massive Increase in Energy Demand: Data centers, which house AI models, are already using a significant amount of electricity, consuming 4.4% of all energy in the US as of the latest reports2. This is set to increase dramatically. New projections estimate that by 2028, AI alone could consume as much electricity annually as 22% of all US households, representing more than half of the electricity going to data centers. The share of US electricity going to data centers is projected to triple between 2024 and 2028, rising from 4.4% to 12%. Leading AI companies are making harnessing ever more energy a top priority and aiming to reshape energy grids. They are planning and constructing data centers of unprecedented scale, with multi-gigawatt power requirements**.... For example, Google expects to spend $75 billion on AI infrastructure alone in 2025, and initiatives like OpenAI and President Trump's Stargate aim to spend $500 billion to build data centers that could require more power than some states6**.

Shift in Energy Demand within AI: While training AI models is energy-intensive (e.g., GPT-4 training consumed 50 gigawatt-hours of energy), inference (running models for user queries) represents the increasing majority of AI's energy demands and is estimated to use 80-90% of computing power for AI currently. The energy used for a single query varies greatly depending on the model size, type of output, location, and time of day. For instance, generating a high-quality image with a leading open-source model can require over 4,402 joules, while generating a five-second video with a high-quality open-source model can require about 3.4 million joules, significantly more than text or image generation.

Future AI Usage Will Greatly Increase Demands: Current energy estimates for individual queries only shed a "tiny sliver of light" on future demands**.** The direction AI is headed – becoming more personalized, able to reason and solve complex problems, and integrated everywhere – means the AI footprint today is likely the smallest it will ever be**.** The future involves AI "agents" performing tasks unsupervised, models used for hours daily in voice or video mode, and "reasoning models" that require significantly more energy for complex problems.

These future applications cannot be understood by simply extrapolating from current query energy usage.