AI needs a ton of energy—but it can save lots of power, too

Today’s biggest AI models burn more power than entire cities, when accounting for the costs of training and operation.

• The energy cost of training a single frontier AI model—like OpenAI’s GPT-4o or xAI’s Grok-3—currently sits at ~100 megawatts, or enough energy to power 100,000 US households.
• That figure is also projected to skyrocket in the coming years to reach 5-10 gigawatts (equivalent to 5-10 million homes).

By 2028, AI alone is projected to consume as much electricity as 22% of all US households annually.

A take and give situation

At the same time, some experts say power-hungry AI technology can also help use a whole lot less energy by creating more efficiencies. Examples include:

• Buildings: Research suggests AI-enhanced HVAC, lighting, and shading systems can save energy and reduce building-sector emissions by up to 19% over the next 25 years. AI can also manage a building’s batteries to shift grid use to off-peak times.
• Transportation: AI route optimization can help freight companies cut fuel use by 10%-15%, help airlines reduce emissions by 3%-10%, and cut ships’ fuel use by up to 20%.
• Materials: AI can accelerate material-science innovation and the discovery of more energy-efficient products to replace—or add to—materials like cement, steel, and plastic.

However…A future where AI enables significant energy savings is far from guaranteed. Many of the efficiency gains promised by AI also require massive data collection, and raise major privacy concerns (like tech companies knowing when consumers are or aren’t home).