Close All Tabs The Real Cost of AI Slop
Jan 28, 2026
James O'Donnell, senior AI reporter at MIT Technology Review, covers AI systems and data centers. Casey Crownhart, senior climate reporter at MIT Technology Review, investigates tech's environmental impacts. They translate AI energy into microwave-time analogies. They explain training versus inference, per-query energy for text, images and video, reasoning models' higher costs, data center cooling and water use, and local grid and infrastructure effects.
AI Snips
Chapters
Transcript
Episode notes
Viral Fake Video That Revealed Costs
- Morgan Sung describes being fooled by a viral AI-generated "bunnies on a trampoline" video that disappeared midair.
- That five-second fake video used about 3.4 million joules, roughly an hour of microwave time.
Inference Is Now The Big Energy Driver
- James O’Donnell explains inference (every query) now often dominates energy budgets compared with training.
- More complex prompts and bigger models consume noticeably more energy per query.
Huge Energy Gap Between Text And Video
- Casey Crownhart and James found small text queries can use ~114 joules while larger text models use thousands of joules.
- Video generation can use millions of joules, thousands of times more than small text queries.