How Many Questions Can the World Afford to Ask AI?

Together, these three constraints—fuel, speed, and storage—form the hard physics of the token economy. You can build a bigger data center or write smarter software, but you cannot negotiate with the physical limits of the universe.

Next come economic constraints

With the physical constraints on tokens established, the researchers built a balance sheet for token supply and demand, converting current and projected energy figures into token capacity, which translates to the capacity for answering questions. Drawing on the work of the late mathematician Richard Cox (who made inquiry a branch of mathematics) and Shannon (whose entropy framework put a number on uncertainty), the researchers formalized the connection between inquiry, or questions, and computation.

A question, in their terms, is an operation that reduces uncertainty. Think of any problem as a room full of fog. A good question clears some of that fog; a great question dissipates half of it—the way “Is it alive?” as an opening move in the game Twenty Questions instantly and significantly reduces the possible answers. By contrast, asking for a cat video clears no fog. In a finite question budget, that difference matters.

The researchers calculated the energy supply available to the token economy using data from the International Energy Agency (an intergovernmental organization that works with countries to shape energy policy) and the US Energy Information Administration. In 2024, roughly 1.6 percent of the US national grid was used by data centers processing AI. By 2028, the researchers project, that figure could climb to about 7 percent. Using today’s efficiency levels (a conservative assumption), they estimate that 2028 US allocation alone could support about 225,000 tokens per person per day globally—approximately 169,000 words, or a novel’s length, for every human on the planet.

The researchers converted that budget of 225,000 tokens per person into the finite number of questions AI can be asked. One hundred tokens could get you the answer to a short factual query such as “how tall is the average person?” Meanwhile, you might spend 1,000 tokens to make a request such as “explain how a vaccine works.” If each query averaged 100 tokens, the 2028 US energy allocation would be able to support roughly 2,200 questions per person per day. Stretch each question to a 1,000-token exchange, and the budget would drop to about 225 questions.

As of mid-2024, global usage sat at only about 125 tokens per person per day, barely a paragraph, according to provider data and estimates from the research institute Epoch AI. The bottleneck right now, the analysis suggests, is not electricity shortage but a shortage of deployed chips and data centers. We have the fuel. We have not yet built enough engines. In fact, providers of large language models are starting to make tough choices about how to use the computing power they have available. OpenAI discontinued its computation-intensive video generation tool, Sora, to reallocate resources toward higher-priority initiatives.

Cheaper tokens, bigger bills

Litowitz, Polson, and Sokolov note a shift in how the industry is burning electricity. Training models used to be the bulk of AI’s electricity bill. Inference—the computation used to answer prompts—represented roughly one third of AI electricity use in 2023. But the International Energy Agency and others indicate that answering questions now accounts for an estimated 60–70 percent of the electricity AI consumes.

This shift reflects falling AI costs. Between late 2021 and late 2024, the cost per million tokens fell from $60 to 6 cents—a thousandfold reduction in three years, causing a response by the market that follows what economists call the Jevons paradox. In 1865, William Stanley Jevons noticed something counterintuitive about James Watt’s improved steam engine: It didn’t conserve coal. By making steam power cheaper, Watt’s engine made coal useful in industries that had never touched it, and Great Britain’s total coal consumption soared.

The AI economy, the researchers suggest, follows the same pattern. When the Chinese AI company DeepSeek released a frontier-class model in January 2025 at a fraction of prevailing costs, the result was not less energy use but exploding demand. Cheaper tokens made AI viable for bulk document processing, real-time translation, and continuous code generation—applications previously priced out.

In its steam-engine era, AI is creating skyrocketing demand for not coal but copper, which is used to help both power and cool the massive data centers crucial to question answering. A single hyperscale AI data center can require up to 50,000 tons for wiring and power systems, according to the Copper Development Association, and the research-and-data provider and consulting firm Wood Mackenzie says that global data-center copper demand is projected to reach 1.1 million metric tons per year by 2030. Meanwhile, electric vehicles, renewable energy, and grid electrification also compete for copper. The consultancy S&P Global projects that annual demand for the metal could hit 42 million metric tons by 2040, up from 28 million metric tons last year.

That would strain global supply. “The world needs six giant tier-one mines to come online every year through 2050 just to keep up with baseline needs, let alone the demand from electrification, data centers, and grids,” writes Frank Giustra, CEO of the investment firm Fiore Group. Yet, opening a copper mine takes 15–20 years.

Giustra says recycling copper could help by adding 4 million tons annually, but “it’s nowhere near enough, and much of the historical stock would require dismantling entire infrastructures to access. Without unforeseen breakthroughs, catching up to demand in the next decade is mathematically near-zero.”

It’s hard to know exactly when a resource problem will arise, Polson says, noting that “many have tried and failed” to predict that timeline.

In 1980, for example, the late economist Julian Simon, who argued that human ingenuity would ensure resources never run out in any economically meaningful sense, famously challenged Stanford biologist Paul Ehrlich to choose any raw materials and a future date beyond a year and he would bet that the inflation-adjusted prices would fall (assuming rising prices as an indication of scarcity). Ehrlich picked five materials, including copper, and bet the length of a decade. Ten years later, Ehrlich mailed Simon a check for $576.07, after prices fell across the board for each of the resources he chose.

Had the same wager run longer, the results would have changed, however. “Let’s take an equally arbitrary but much more satisfactory bet: from then, 1980, until now, and include all of the most important commodities,” wrote British investor Jeremy Grantham in 2011. “Simon would have lost posthumously, and by a lot!”

The Simon-Ehrlich bet illustrates one measurement problem facing AI: knowing when resource constraints will bind. But the race to build better AI has a second, more immediate measurement problem built into the technology itself. In this realm, the researchers invoked Goodhart’s law, coined by London School of Economics’ Charles Goodhart, which states that when a measure becomes a target, it ceases to be a good measure. They offer an illustration, attributed to Anthropic’s cofounder John Clark and its CEO Dario Amodei: An AI agent trained to play a boat-racing video game called CoastRunners was given the goal of maximizing its score. Instead of finishing the race, it discovered that driving in circles and hitting regenerating bonus targets earned more points. It crashed into other boats, caught fire, and drove the wrong way. But it got the high score, beating those that completed the course.

Optimizing harder doesn’t fix a flawed target, it just pursues the wrong goal more efficiently. More computation doesn’t solve the problem; it amplifies it. Thus, AI faces two different kinds of limits: one physical and one related to measurement. The first will become a big issue at some unknown point in the future. The second, however, is already here.

The AI token economy

The physical constraints of the supply chain explain what AI can produce, while economics explains who benefits and where the money goes.

The late economist Ronald Coase said companies exist because of transaction costs. The researchers borrowed from Coase in defining the AI value chain, which stretches from the photons and atoms at the bottom; up through chips, electricity, and tokens; to the questions humans pose at the very top. At the bottom sit slow, capital-heavy industries such as mineral mining and semiconductor fabrication. At the top sit fast, nimble software layers that convert electricity into responses. Litowitz, Polson, and Sokolov argue that economic value migrates upward as the lower layers commoditize.

The mechanism driving the upward migration is straightforward: Every GPU sold today competes with yesterday’s. Buyers who know that a better generation of GPU will ship in 18 months have reason to wait for it, putting long-term downward pressure on hardware margins. Tokens face no such self-competition, however. They vanish the instant that answers get read.

The researchers identified the competitive variable as tokens per watt per dollar. Whoever squeezes the most intelligence from a single cent of electricity wins. Those profits are multiplied by a silent subsidy embedded in the system. AI companies buy electricity at cheap industrial rates and sell AI responses at unregulated market prices. “The implicit subsidy flows from ratepayers and the environment (which bears the unpriced externalities of generation) to AI shareholders,” Litowitz, Polson, and Sokolov write, explaining that the public bears the grid strain and the environmental cost. Likewise, every ton of copper committed to today’s infrastructure comes at an opportunity cost to future generations.

Can policy help?

Markets will always steer tokens toward whoever can pay. Currently, though, just a few AI model providers essentially decide how to dole out access to LLMs and AI, thus are deciding how the question budget is spent based on their “pricing tiers, rate limits, acceptable use policies, and content moderation rules,” the researchers argue. One alternative approach would be to regulate AI like a utility or public infrastructure.

Right now the most profitable uses of AI tend to win at the expense of social uses. The companies are free to favor things that make money, such as ad copy, over rare-disease research or other information of social value. The information paradox developed by the late economist Kenneth Arrow compounds the problem: You cannot judge the value of an answer until you have already received it, making efficient pricing of AI queries structurally impossible, say the researchers. The market for intelligence, the analysis suggests, will always carry a whiff of the casino.

All of this funnels toward a blunt policy question. If the question budget is finite, who decides how to spend it? “The choice between a medical diagnosis and a disposable video is not a technical question,” the researchers write. “It is a political one—and it should be made with the numbers in hand.”

When resources are fixed, every question displaces another that might have mattered more. The researchers leave us with a call to move toward a mindset of deliberate, informed resource management where human judgment dictates the direction of the technology. In an age drowning in answers, the rarest commodity may be the judgment to know when posing the question is worth the electricity it will burn.