AI data centers consumed 415 terawatt hours of electricity in 2024. That is more than 10% of total US power production, and the International Energy Agency says demand will double by 2030. Everyone in the industry knows this is unsustainable. Almost nobody has a solution.

A team at Tufts University might have just found one.

Matthias Scheutz, a professor at Tufts' School of Engineering, has built a proof-of-concept AI system that cuts energy consumption by up to 100 times while actually improving accuracy. The research will be presented at the International Conference of Robotics and Automation in Vienna this May.

The Core Insight: Stop Guessing, Start Reasoning

The approach is called neuro-symbolic AI, and it combines traditional neural networks with symbolic reasoning. Instead of throwing billions of parameters at a problem and hoping the statistics work out (which is basically how every large language model operates), the system breaks problems into logical steps and applies rules.

Think of it this way: a conventional AI robot asked to stack blocks will analyze every pixel, try hundreds of placements, fail repeatedly, and eventually learn through brute force. A neuro-symbolic system understands what "balance" means as a concept, reasons about which block goes where, and gets it right faster with a fraction of the compute.

Scheutz's team focuses on visual-language-action (VLA) models, which power robots that need to see, understand language, and physically act. These are the systems behind every humanoid robot you have seen in the news. And right now, they are wildly inefficient.

Why This Could Matter More Than Any New Model Launch

OpenAI is about to burn $85 billion in a single year on compute. Anthropic expects 2026 to be its biggest year of losses. Every major AI company is locked in an arms race to build bigger models that consume more power. The entire industry strategy is predicated on the assumption that more compute equals better AI.

This research says that assumption might be wrong. Not slightly wrong. 100x wrong.

Now, the caveat: this is a proof-of-concept tested on classic planning puzzles like the Tower of Hanoi, not on frontier language tasks. The leap from "stacking blocks efficiently" to "replacing GPT" is enormous. But the principle is sound. The same hallucination problems that plague chatbots also plague robot AI. If symbolic reasoning can fix the robot version while using 100x less energy, someone will eventually try it on language models too.

The AI industry is spending trillions to scale brute-force computing. The most interesting research right now is the stuff that asks whether we need to.

Published in the proceedings of the International Conference of Robotics and Automation 2026.