The Body of Artificial Intelligence

I hit a wall a few weeks ago. I was in the middle of building out a system I'd been working on for days. Not casually browsing. Actual work, the kind where you're deep in context and losing it means starting over.

Then the message appeared: you've reached your limit. Come back later.

My first instinct was to blame my subscription. I figured I was on the wrong plan, or I'd used up whatever my tier allowed. But that nagging feeling stayed. So I started looking into what was actually happening behind that message. What I found changed how I think about every AI tool I use.

What a token actually is

When I started digging, the first thing I learned is that AI doesn't think in words. It thinks in tokens. A token is a chunk of language, roughly three to four characters of English text. The word "fantastic" might be one token. A space or a comma can be a token too.

Every message I send gets broken into a stream of tokens. Every response comes back the same way, generated one token at a time. Every exchange I have with an AI tool is thousands of these tokens being processed by a machine running somewhere I can't see.

That machine isn't virtual. It's physical. And running it isn't free. That's where the story gets much bigger than my monthly plan.

I didn't expect the electricity numbers

Processing tokens at this scale requires enormous computing power. That power lives in data centers, massive warehouses packed with specialized chips that handle the calculations behind every AI response.

In 2024, data centers consumed around 415 terawatt-hours of electricity globally, according to the International Energy Agency. That's roughly 1.5% of all the electricity used on earth. In the U.S. alone, data centers used more than 4% of the country's total, comparable to the annual demand of a nation like Pakistan.

One data center campus with peak demand of one gigawatt would use more electricity in a year than the entire state of Vermont. By 2030, the IEA projects data center consumption will nearly double, reaching roughly the equivalent of Japan's entire annual electricity use today.

The cloud isn't a cloud. It's a building. It's land. It's a power grid. And it's thirsty.

Then I looked at the water

Electricity generates heat. Heat has to go somewhere. In most data centers, the answer is water.

A typical data center uses around 300,000 gallons per day, according to Brookings Institution research. That's the equivalent of about 1,000 households. Google's thirstiest single facility, in Iowa, consumed 1 billion gallons of water in 2024, peaking at 2.7 million gallons on a single summer day.

Much of this water evaporates during cooling. It doesn't go back into the system. And two-thirds of new data centers built since 2022 sit in regions already facing water stress, according to Bloomberg News data.

I'd never thought about my AI usage in terms of water before. That was the moment this stopped feeling abstract.

The chips behind all of it

The hardware story surprised me most. The chips that power AI aren't like the chips in your laptop. They're specialized processors (graphics processing units, or GPUs) that handle the parallel calculations AI requires at speed. One company, Nvidia, controls somewhere between 80 and 92% of the market.

But Nvidia doesn't manufacture its own chips. Almost no one does anymore. A handful of facilities fabricate the most advanced chips in the world, with Taiwan Semiconductor Manufacturing Company at the center of the supply chain. The precision required (processes measured at the scale of atoms) means only a few places on earth can do it at all.

This is why the tension between the U.S. and China over semiconductor access has become one of the most consequential economic conflicts of this era. It's not about trade. It's about who controls the physical hardware that intelligence runs on.

This pattern isn't new

The more I read, the more familiar the shape of the story became. Every major technological revolution has promised to free us from the constraints of the previous one. The industrial revolution replaced human muscle with machines, and coal became the resource everyone fought over. The digital revolution replaced physical objects with software, and oil, rare earth metals, and bandwidth became the new battlegrounds.

AI was supposed to be different. Intelligence itself, scalable and weightless. Available to everyone. But intelligence turns out to need a body. It needs chips. Chips need water and rare materials. Data centers need electricity and more water. Electricity, in many regions, still needs fossil fuels.

We didn't escape scarcity. We relocated it, and made it bigger.

How fast the math changed

When I first tried ChatGPT in late 2022, I used it a few times and moved on. It was interesting but not essential. I started using AI seriously about a year ago, after I got laid off. That's when it became a daily tool, not a novelty.

My timing wasn't unusual. In 2023, the central challenge for these companies was still adoption. ChatGPT had reached 100 million users faster than any consumer app in history, but most people hadn't touched it. Companies burned through capital to offer free access and generous limits. The scarce resource was human attention.

By 2025, ChatGPT had gone from 200 million to 800 million weekly active users in a single year. Claude, Gemini, and a wave of coding tools scaled alongside it. The adoption problem was solved. What replaced it was a cost problem no one had priced for.

The response came fast. OpenAI introduced a $200-per-month tier in December 2024. Anthropic followed with its own in April 2025, then added weekly rate limits that summer. Google published hard daily prompt caps for Gemini. Coding tools like Cursor and Replit repriced to limit power users.

This month, Anthropic went further. It blocked a popular third-party agent framework called OpenClaw from using Claude subscriptions entirely, affecting over 135,000 users. The company said its flat-rate plans "weren't built for the usage patterns of these third-party tools."

Usage outran the business model. The people who once taught you how to prompt are now teaching you how to conserve. That shift didn't happen by accident.

What this changed for me

The promise of AI, the version repeated in every keynote and pitch, is that it will keep getting better, keep getting cheaper, and keep becoming more accessible. That billions of people will use it for everything.

Maybe all of that is true, eventually. But there's a version of this story that doesn't get talked about much: the one where the energy and resource infrastructure can't keep pace with the ambition. Where the bottleneck isn't the intelligence. It's the planet.

The people promising that AI will replace most of human knowledge work are also implicitly promising that the energy to run it will exist, at scale, affordably, reliably, everywhere. That's a second bet. And it's a large one. Right now, natural gas still powers over 40% of U.S. data center electricity, with coal covering another 15%. The clean energy transition and the AI scaling race are happening simultaneously, and they're competing for the same grid.

We may get there. The engineers working on more efficient chips, smarter cooling, and nuclear microreactors aren't unaware of the problem. The search for a sustainable path is real and serious. But history suggests that breakthrough technology and resource pressure tend to arrive together, and the pressure usually comes first.

The next time you hit a usage limit, or notice that a tool you relied on has gotten more expensive, that's not a glitch. That's the edge of a very real resource equation working itself out in real time.

I don't have the answer. I'm not sure anyone does yet. But I wanted to understand what the question actually was. And it turned out to be a lot bigger than my monthly plan.