AI Is Thirsty (and not just for knowledge)

In honor of Earth Day, I wanted to tackle a subject that’s been hovering in the background of my AI workflows for a while. We’re diving into the world of AI data centers: what they actually are, why their environmental footprint is becoming impossible to ignore, and what we can do about it.

First, AI isn’t going anywhere. And, we can’t holistically deem it villainous and turn away from it because it’s “bad.” Well, at least that’s not what I’m going to do, and I’m guessing most of you won’t either. Some of the best innovations in human history — cars, planes, spacecrafts — contribute to pollution, and we didn’t abandon them. We eventually (albeit imperfectly, inconsistently, and rather ambiguously at times) worked to understand the cost and do better.

AI is the same deal. We need to understand what we’re working with. So I went looking.

Through my work in communications for an industrial hemp fiber startup, I’ve spent a lot of time inside of the textile industry (not the glossy Anna Wintour version, the supply chain version).

I’ve seen the mountains of discarded clothing sitting in open-air dumps in Chile’s Atacama Desert, the kilotons of microplastics leeching into waterways, and the many gallons of water it takes to produce one single shirt. Before that, “fast fashion is bad,” was just a thing I politely agreed with, but now I actually get it.

AI feels similar.

For most of us, using AI feels a lot like shopping at Zara. It’s fast, a bit addictive, and so frictionless that the cost becomes invisible — especially when you're just tapping away on your phone. Meanwhile, far from our screens, the physical world is quietly paying the price in diverted water and strained power grids

AI feels like magic. You type, it responds. Done. 🪄

But behind every prompt you send is a very real, very physical building, somewhere in the world, filled with servers, GPU and CPU processors, cooling systems, batteries, backup power generators, security infrastructures …. aaaaand a host of environmental problems.

⚡ The Electricity Problem

U.S. data centers used about 4.4% of total electricity in 2023. Which sounds manageable until you realize the entire U.S. lighting system — every home, office, and streetlight — uses about 5%.

By 2028, projections land somewhere between 6.7% and 12%, the high end of which is roughly the equivalent to the total residential electricity use of California, Texas, and New York combined.

💧 The Water Problem

This is the one that got me.

In 2023, U.S. data centers consumed an estimated 17 billion gallons of water, which is the entire annual water supply of the city of Houston. Used in one year just to keep servers cool. And what’s worse, a lot of that water doesn’t come back. It evaporates. It’s just … gone.

By 2028, projections for large-scale facilities alone hit up to 33 billion gallons annually.

To bring it down to a more personal level, scientists at UC Riverside estimate that a 100-word AI prompt uses about one standard water bottle’s worth of water, and there are hundreds of millions of prompts sent every single day.

🌫️ The Carbon Emissions Problem

Data centers run on energy, and much of that energy comes from fossil fuels. One estimate I read puts the increase in emissions tied to data centers’ demands at around 220 million tons of carbon. For context: driving one gas car 5,000 miles produces about one ton of CO₂.

Also worth noting: those diesel backup generators that kick in when the grid strains? Not exactly clean energy heroes.

We’re currently at about 11,000 data centers globally. At the current pace of AI growth, data center capacity will need to grow by 130% by 2030.

For a sense of how seriously the industry is taking this, U.S. spending on data center construction has now officially surpassed spending on traditional office buildings for the first time in history. By the end of 2025, monthly data center construction spending reached over $3.6 billion.

And nobody can tell you exactly how many data centers we need, because nobody fully knows how much AI we’re going to use or whether efficiency gains will outpace demand. What we do know is that right now, we’re building as fast as we possibly can and just hoping the math works out. 😰

This is a systemic issue that requires policy, regulation, and corporate accountability to meaningfully move the needle. Individual action isn’t going to fix this kind of structural problem.

But informed users turn into informed employees, and those people ask better questions in the rooms where decisions actually get made.

For Individuals🧍🏻‍♀️

✅ Use the lighter model

Most of us pick one AI tool, use it for everything, and never think about it again. But every major platform has model options and the lighter, faster versions use a fraction of the energy of the big flagship ones.

Inside Claude: there’s a model selector at the bottom of the chat window. Haiku is the lightest and great for quick summaries, rewrites, and short Q&As. Sonnet is the middle ground, Opus is for genuinely heavy lifting.

Inside ChatGPT: GPT-4o mini is the lighter option. There’s also an “Auto” mode that picks the right model size based on your request.

The rule of thumb: if you’re about to hit send on something routine such as summarizing a document, fixing a sentence, generating a quick list, try the lighter model first. You’ll probably get 95% of the quality at a fraction of the footprint.

✅ Match the tool to the job

Think of your AI tools less like a Swiss Army knife and more like a small team with different specialties:

• Claude → deeper writing, editing, long-form thinking

• ChatGPT → drafts, ideation, quick iterations

• Gemini → research inside Google’s ecosystem

• Perplexity → fast, cited lookups

Better tool selection = fewer retries = less wasted energy

✅ Prompt smarter

Every additional prompt costs something. Better prompts with clear, specific, with context upfront reduce trial and error, cut down on follow-up questions, and shorten conversation length. Getting to the point the first time is both a productivity win and, it turns out, a small environmental one.

✅ Ask questions at work

If your company is rolling out AI aggressively, it’s reasonable to ask what that actually costs environmentally. You don’t need to be the expert, you just need to be the person willing to raise the question.

You may have heard the advice to choose AI providers with strong renewable energy commitments. It’s good advice in theory. In practice, it’s complicated.

Here’s a breakdown on the major players:

Google (Gemini) is the most vocal. They’ve matched 100% of their global electricity use with renewable energy purchases since 2017. But their own 2025 environmental report acknowledged that total emissions increased 11% year-on-year due to data center growth. Their emissions are up 51% from their 2019 baseline.

Anthropic (Claude) has not publicly committed to specific climate goals, has not reported carbon emissions figures, and has no documented reduction targets on record. They run on Amazon Web Services, which reached 100% renewable energy in 2023, so there’s an indirect green claim via the cloud provider. But Anthropic itself has published essentially nothing on its own environmental footprint.

OpenAI (ChatGPT) scores a 23 out of 100 on DitchCarbon’s sustainability rating reflecting a near-total absence of disclosed environmental data. No annual sustainability report. No third-party audited emissions figures. Their primary defense is that they run on Microsoft Azure, which has its own green commitments.

The uncomfortable truth is that right now, choosing the “greenest” AI provider is a bit like trying to pick the most ethical fast fashion brand. The options are limited, the data is murky, and the most important thing might just be using less of all of it and more thoughtfully.

Companies are racing to adopt AI and many have made public climate commitments. Those two things are increasingly at odds. Simultaneously, employees are noticing and the tolerance for vague sustainability language or greenwashing is dropping fast.

According to S&P Global’s 2025 Corporate Sustainability Assessment only 36% of companies have a dedicated AI governance policy at all, and fewer than 1 in 5 can actually quantify the impact of their AI use on sustainability goals. It’s not a great foundation for credible communication, but it’s exactly where comms professionals can add real value.

What employees actually want to know

They’re not asking for perfection. They’re asking for clarity on these main questions:

• Do we understand what AI is costing us environmentally?

• Is leadership making decisions about AI adoption with sustainability factored in, or is it just speed and efficiency?

• What can I do, in my own role, that makes a difference?

• Are we being honest about the tension between our AI growth goals and our climate commitments?

The worst outcome isn’t that your organization doesn’t have all the answers yet, it’s that employees sense the answers exist but aren’t being shared.

What this looks liks is clear internal communication that involves stating the tension out loud, sharing what you know even if it’s incomplete, and giving people something practical to do or contribute to as you all figure it out together.

The companies that handle this well won’t be the ones with perfect numbers, they’ll be the ones telling the truth while they figure it out.

The infrastructure behind AI is heavy, physical, and resource-intensive in ways most of us—myself included—rarely stop to consider. The water evaporates. The fossil fuels burn. Data centers are rising faster than renewable energy can keep up. I’m not sharing this to catastrophize; I simply looked under the hood, and this is what I found.

But, none of thsis means we log off and call it a day. It just means that we use these tools with a little more awareness, a little more deliberately, and maybe a little more proportionately, and we stay in the conversation about what better looks like.

Knowledge is power, and the people who understand the cost are the ones who eventually will help to change it.

Happy Earth Day. 💚

📬 If this made you think differently, send it to someone who uses AI all day without thinking about what’s behind it.

Did this change how you’ll use these tools?