Urgent.News

What's breaking now, across thousands of outlets.

AI

How much energy does agentic AI actually use? One scientist tracked every prompt he sent

As a climate scientist, Zeke Hausfather uses artificial intelligence every day for research and data analysis. But he kept wondering how much energy all that AI use was consuming. So over eight weeks this summer, he tracked the 1,138 prompts he typed into Claude Code and estimated how much electricity they consumed at data centers. Tech companies have shared little data about the energy use of…

How much energy does agentic AI actually use? One scientist tracked every prompt he sent

Climate scientist Zeke Hausfather explored the energy consumption of artificial intelligence (AI) prompts during a study conducted over eight weeks this summer. He tracked 1,138 prompts he inputted into Claude Code and estimated the electricity consumed at data centers. While tech companies have been secretive about the energy use of their AI models, a few estimates exist.

Google reported that a single text prompt on Gemini uses approximately 0.24 watt-hours, while OpenAI CEO Sam Altman suggested a ChatGPT query consumes around 0.34 watt-hours. However, Hausfather believed these figures may no longer accurately represent how individuals currently utilize AI.

Hausfather argued that AI usage has evolved from simple chatbot conversations to complex "agentic AI," where a single instruction leads to the creation of numerous subagents, resulting in significantly higher energy consumption. He estimated that his median prompt utilized about 150 kilowatt-hours of energy, which is approximately 600 times more than a typical chatbot query.

While this estimate is based on token usage and correlation with energy pricing, it remains relatively rough. Nonetheless, even this substantial amount of electricity consumption is still insignificant compared to running an electric dryer for a year. However, when considering the broader impact, AI is poised to become a major new source of energy demand. By some estimates, AI data centers could consume up to 12% of all U.S. electricity by 2030.

Currently, many tech companies are relying on fossil fuels to meet the energy demands of new data centers. For instance, Meta is constructing seven natural gas plants for a massive data center in Louisiana. Nevertheless, this decision is not inevitable. Hausfather contends that the AI boom could potentially accelerate the transition to clean energy if tech companies allocate the same resources and urgency towards building a greener, more capable grid.

He emphasizes that electrifying nearly all aspects of society, including vehicles and home heating, is crucial to electrifying current fossil fuel-dependent processes. This shift would require a significant expansion of the grid, creating challenges but also providing an opportunity to address energy demand and emissions.

Written by urgent.news from Fast Company's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at fastcompany.com →

More in AI

AWS Lets Agent Builders Restrict Web Search to Approved, Recent Sources

AWS has added domain and publish-date filtering to the Web Search tool inside Amazon Bedrock AgentCore, according to the AWS Machine Learning Blog .

  • AWS introduces domain and publish-date filtering for Web Search in AgentCore.
  • Filters limit search to predetermined domains and recent sources.
  • Enhances relevance, reliability, and security for AI agents.

Why Most Multi-Agent Systems Fail in Production (And How to Fix It)

Most multi-agent demos look impressive on stage. Then they hit production and fall apart. Here's the pattern: agents that "worked" in a Jupyter notebook start conflicting, retrying infinitely, or…

  • Multi-agent systems fail in production due to lack of structured communication.
  • Agents use plain text messages leading to misinterpretations and errors.
  • Absence of retry mechanisms causes failures to halt or loop endlessly.

How to Stop Fighting Your AI Coding Assistant

I used to hate AI coding assistants. They felt like talking to a junior developer who had read the documentation once and forgot the rest.

  • Treat AI assistants as probabilistic text generators, not oracles
  • Break down complex tasks into smaller, manageable steps
  • Provide clear context, style preferences, and constraints to guide AI

More from Thursday 3 September →