In early 2026, a research team at Tufts University published a paper with a title that sounds like a tax complaint: a paper whose title challenges the cost assumptions behind modern AI training. The subject was not taxes. It was the single most expensive computational arms race in modern history, the race to build smarter AI, and the paper’s central claim was that almost everyone running that race had been doing it wrong.
The numbers behind that claim are worth sitting with for a moment. According to the International Energy Agency, US data centers and AI systems consumed hundreds of terawatt-hours of electricity in 2024. That is roughly 10% of the entire country’s electricity generation. To put that in terms that don’t require an engineering degree: the machines thinking on behalf of American industry now draw more power than most countries generate in a year. And the IEA projects that figure will grow substantially by 2030.
The Tufts team, which includes researcher Timothy Duggan, wasn’t arguing that AI should be slower or dumber. They were arguing something more unsettling: that the architecture driving all that consumption might be fundamentally misaligned with what intelligence actually requires.
The Tower and the Test
The standard benchmark Duggan’s team used was the Tower of Hanoi, a classic recursive puzzle used in cognitive science and computer science education. The goal is to move a stack of differently-sized disks from one peg to another, following simple rules: only one disk at a time, never put a larger disk on a smaller one. It sounds trivial. For modern AI systems built on pattern recognition alone, it is not.
Standard vision-language-action models, the kind of large-scale AI being deployed in robotics and manufacturing, achieved a significantly lower success rate on the task, according to the paper. The Tufts neuro-symbolic hybrid achieved a substantially higher success rate, which the paper reports at 95%. That gap, between 34 and 95, is not a marginal improvement. It is a different category of result entirely.
And here’s the part that should make every AI lab director uncomfortable: the neuro-symbolic system used a fraction of the training energy required to reach that result, a figure the paper puts at roughly 1% of the comparable baseline. One percent. Which sounds insane until you realize it follows directly from what the system was actually doing.
What “Neuro-Symbolic” Actually Means
The phrase sounds technical, but the idea behind it is old and intuitive. Current mainstream AI, the kind that powers large language models, image generators, and most commercial robotics, learns by ingesting enormous quantities of data and finding statistical patterns within it. It does not reason in any formal sense. It predicts. Given enough examples of what a correct answer looks like, it gets very good at approximating one.
The neuro-symbolic approach does something different. It combines that pattern-recognition layer with a symbolic reasoning system: a set of explicit logical rules the machine can apply step by step. Think of it as the difference between a person who has read ten thousand chess games and can usually guess a good move, versus a person who actually understands the rules of chess and can calculate forward from any position.
For problems like the Tower of Hanoi, the distinction matters enormously. The puzzle is governed by clear rules. A system that can represent and apply those rules will always outperform one that is guessing from patterns, especially when it encounters a configuration it has never seen before. That is precisely what the Tufts results showed.
The cognitive science community had been making this argument, in various forms, since the 1980s. The AI field largely moved away from symbolic approaches when deep learning arrived and proved itself on perception tasks, image recognition, speech, and translation. Pattern machines turned out to be extraordinarily capable at those things. The assumption, never quite stated but widely held, was that scale would eventually solve everything else too.
The Scale Assumption and Its Costs
That assumption is now running into physics. Training a frontier AI model requires not just enormous amounts of data, but enormous amounts of compute, and compute requires electricity. The economics have been absorbed quietly by the industry, passed along to customers and investors, and largely invisible to the public. But the IEA numbers make the trajectory plain: the current path is not sustainable at scale.
The Tufts paper, scheduled for presentation at the ICRA 2026 robotics conference, does not claim to have solved general AI. Neuro-symbolic systems have their own limitations; they struggle when the rules governing a domain are unclear or incomplete, which describes most of the real world most of the time. A robot navigating a crowded kitchen is dealing with a far messier problem than a Tower of Hanoi puzzle, and no one is suggesting symbolic logic alone handles that.
What the research does suggest is that the field’s instinct to throw more data and more electricity at every problem may be costing far more than it returns. For a specific class of structured tasks, the kind that dominate industrial robotics, logistics, and manufacturing, a hybrid approach appears to beat brute force on both performance and efficiency simultaneously.
That combination, better results at a fraction of the cost, is not the kind of finding that stays in academic journals. It tends to become a product strategy.
The 415 terawatt-hours from 2024 will look modest by 2030 if current trajectories hold. Whether the machines consuming all that power will still be built the same way is the question that “The Price Is Not Right” is quietly asking everyone in the field to reconsider. Whether anyone listens before the next round of data centers breaks ground is a different matter entirely.
This article was created with AI assistance and reviewed for clarity and accuracy.
