Is ChatGPT raising my electric bill?

Quick Overview

The energy cost of an individual ChatGPT text query is negligible compared to the training cost of the AI model, yet the massive scale of data centers, which AI significantly drives, is projected to consume 12% of US total electricity by 2028, leading to increased electricity bills for regular households because utilities must build new infrastructure to meet this demand.

Key Points: A single ChatGPT text query uses approximately 0.34 watt-hours (Wh) of energy, equivalent to running an incandescent light bulb for 18 seconds. The energy cost for generating an image via AI can be 10 times more, and generating video can be 45,000 times more than a text query, according to research cited. US data centers accounted for 56% of US electricity consumption derived from fossil fuels, with a carbon intensity 48% higher than the national average. Data center electricity use is projected to reach 6% of total US electricity consumption by 2025 and 12% by 2028, according to Lawrence Berkeley National Laboratory projections. The high energy demand from data centers forces utility companies to build expensive new infrastructure, the costs of which are socialized across all customer electricity bills. The ML.ENERGY Leaderboard ranks large language models based on energy efficiency (Energy per Token) and latency.

Context: The video explores the hidden environmental and economic costs associated with the rapid growth of Artificial Intelligence (AI), particularly large language models (LLMs) like ChatGPT. The presenter investigates the energy consumption of AI inference (generating answers) versus training, and contrasts these figures with the broader energy demands of the data centers that host these systems, which are increasingly impacting local communities and household electricity costs.

Detailed Analysis

The video investigates whether using ChatGPT is raising individual electric bills, concluding that while the energy cost of a single query is small (0.34 Wh for text), the massive overall electricity demand from the data centers powering these AI tools is driving up utility costs for everyone. The presenter cites data showing that a single text query uses about 0.34 Wh, comparable to running a 60-watt light bulb for 18 seconds, or brewing coffee for 10 seconds. However, image and video generation consume vastly more energy—up to 45,000 times more than text generation for video. More significantly, data centers themselves account for a growing portion of US electricity consumption, projected to hit 12% of the US total by 2028, with 56% of that coming from fossil fuels. The high demand forces utility companies to build new, expensive infrastructure, and these infrastructure costs are socialized, meaning regular households end up paying for the build-out through higher electric bills. The video references the ML.ENERGY Leaderboard for ranking LLM efficiency and features interviews with experts like Jae-Won Chung and Ari Peskoe to contextualize these infrastructure and cost issues.