How Many Questions Can the World Afford to Ask AI? | Chicago Booth Review
Home-->
Chicago Booth Review
Monetary Policy
Inequality
Health Care
Climate Change
Artificial Intelligence
Motivation
Jobs
All Chicago Booth Review Topics
How Many Questions Can the World Afford to Ask AI?
Chicago Booth ReviewHow Many Questions Can the World Afford to Ask AI?
Monetary Policy
Inequality
Health Care
Climate Change
Artificial Intelligence
Motivation
Jobs
All Chicago Booth Review Topics
Mariaelena Caputi
How Many Questions Can the World Afford to Ask AI?
Every query incurs energy costs, raising concerns about resource allocation.
By<br>Monika Brown
April 24, 2026
CBR - Artificial Intelligence
Share This Page
Share This Page -->
Chances are, in the past few days you’ve posed a question to a chatbot or search engine. Writing the question may have taken a few seconds. The response required a sliver of electricity from a power plant, a pulse through copper wiring, and enough waste heat to warm a thimble of water.
Multiply that by trillions of queries a day, and the thimble becomes an ocean.
We talk about artificial intelligence as if it lives in a weightless cloud, but it doesn’t. Every token—AI’s basic unit of text, typically three to four characters—initiates a process with an energy cost as real and resource intensive as burning coal.
Most of us never see that meter running, but research by QStar Capital’s Alec Litowitz, Chicago Booth’s Nicholas Polson, and George Mason University’s Vadim Sokolov tries to measure it. Debates about AI capabilities and risks, they contend, shouldn’t proceed without quantitative grounding.
Given AI’s dependence on earth’s limited resources—especially the copper and other minerals, fossil-fuel-based energy supply, and water needed for data centers—the researchers contend that we face a finite “question budget.” They seek to calculate a fundamental puzzle: How many questions can be directed at AI systems given the planet’s physical constraints? Precisely because the meaningful questions we ask AI, the kind that could help answer real problems, are as abundant as the inane ones, the answer has big implications for policymakers, investors, and the general public—and, let’s face it, humanity too.
The physics of AI
The researchers started with an insight from the early-20th-century polymath John von Neumann: Computation— whether by a brain or a computer—is a physical process that consumes energy. They coupled it with the approach the late physicist Sir David MacKay used to cut through energy debates. MacKay’s 2009 book, Sustainable Energy—Without the Hot Air, “reframed energy policy as a problem of arithmetic,” write Litowitz, Polson, and Sokolov.
Much like MacKay did for energy, the researchers turned the real-world system by which AI transforms copper, electricity, and other materials into answers into a math problem. They converted AI supply and demand into a common unit (the token), then followed AI’s supply chain from mining site to data center to computer screen. By mapping the world’s energy supply, material resources, and computing capacity against current AI consumption, they were able to define the question budget. The researchers anchored the idea that it takes a measurable amount of energy to generate a snippet of AI text using Landauer’s principle, named for the late physicist Rolf Landauer, who established that computation is never free—processing or discarding information always releases some heat.
The universe, in other words, charges a small tax on every calculation. The principle allowed the researchers to set a physical floor on how efficient AI can ever become. Translating Landauer’s floor to a single token, which they estimate to be 12 bits, they calculated the theoretical minimum for producing a token as about 3.4 × 10-20 joules, a tiny number. But today’s AI chips are staggeringly wasteful by comparison, burning roughly 50 quintillion times more energy per token than that minimum, the study finds. The gap is so large, it’s like needing a tanker of fuel to push a bicycle one block. Yet that gap also represents runway for hardware improvement, suggesting the industry is still in its steam-engine era.
Two more physical limits tighten the picture. The late mathematician Claude Shannon’s channel capacity theorem sets a speed limit on how fast data can travel through any wire or chip. (Shannon pioneered information theory.)
The Bekenstein bound, named for the late physicist Jacob Bekenstein, caps how much information any physical object can store or process.
You can build a bigger data center<br>or write smarter software,<br>but you cannot negotiate with the physical limits of the universe.
Together, these three constraints—fuel, speed, and storage—form the hard physics of the token economy. You can build a...