Every time you ask an AI a question, something physical happens far away: servers in a data center draw power and, often, water to stay cool. Individually it's tiny. Multiplied by billions of queries and a global building boom, it adds up to a real environmental footprint — one that moved into the spotlight again as Google announced plans to raise roughly $80 billion for AI infrastructure alongside a pledge to become water-positive by 2030.

So how big is AI's hidden cost, really? Here's an honest look at what the 2026 estimates say, why it's growing, what's being done — and what it means for the rest of us.

The scale (with appropriate caution)

Exact numbers are hard to pin down and estimates vary widely, but the direction is clear. According to a United Nations University report issued in June 2026, the world's data centers consumed on the order of hundreds of terawatt-hours of electricity last year — more than the total consumption of all but a handful of countries — along with roughly a trillion-plus gallons of water and emissions comparable to those of a mid-sized nation (PBS NewsHour, UN University).

Treat any single figure as an estimate, not gospel — methodologies differ and companies don't always disclose. But across sources, the takeaway is consistent: AI's physical footprint is large and rising.

Why water, of all things?

The energy part is intuitive — computers need electricity. The water part surprises people. Large data centers generate enormous heat, and many use water for cooling (directly, or indirectly via the power plants that supply them). In water-stressed regions, that draw can compete with local needs, which is exactly why "water-positive" pledges have become a talking point.

Why it's growing so fast

The generative-AI boom is the accelerant. Training and running large models is computationally heavy, and demand is climbing. Multiple analyses — including projections referenced by the International Energy Agency — expect data-center energy use to roughly double within a few years as AI scales. More models, more users, more "AI built into everything" all push the same direction.

This is the flip side of the trend we covered in why big tech is building its own AI models: the race for AI capability is, underneath, a race for power, chips, and cooling.

What's actually being done

It's not all grim — there's real work on the problem:

  • Efficiency gains. Newer chips and smarter cooling keep cutting the energy per computation, and "good enough, cheaper" models reduce waste.
  • Cleaner power and water pledges. Commitments like water-positive-by-2030 and renewable-powered data centers are now competitive table stakes — though pledges are promises, and delivery is what counts.
  • Reusing waste heat. Some projects capture data-center heat for nearby uses instead of dumping it.
  • Smaller, on-device models. Running AI locally on your own device — instead of a distant data center — can cut the cloud burden for many tasks. We looked at this shift in on-device AI and local models.

What it means for you

You don't control a data center, but the footprint is worth factoring into how you use AI:

  • Don't waste compute. Firing off the same heavy request ten times, or using a giant model for a trivial task, has a real (if small) cost. Pick the right-sized tool for the job — our framework for choosing the right AI tool helps.
  • Prefer efficiency where it exists. Smaller and on-device models are often plenty for everyday tasks.
  • Ask vendors for transparency. As a customer, demand for disclosure is part of what drives improvement.

The balanced view

None of this means AI is an environmental villain to be avoided. AI also helps cut energy use — optimizing power grids, improving logistics, accelerating climate and materials research. The honest framing isn't "AI bad," it's: a powerful, resource-hungry technology whose footprint we should measure, disclose, and work to shrink — not ignore in the excitement.

FAQ

Does using AI really consume water? Indirectly and directly, yes. Many data centers use water for cooling, and the power plants supplying them often use water too. A single query's share is tiny, but at global scale the totals are significant, according to 2026 reports.

How much energy do data centers use? Estimates put global data-center electricity use in the hundreds of terawatt-hours per year — comparable to entire countries — and several analyses project it to roughly double within a few years as AI grows. Exact figures vary by source and methodology.

Is AI's environmental impact getting better or worse? Both, in a sense. Efficiency per computation keeps improving, but total demand is rising faster, so the overall footprint is growing. Efficiency gains and cleaner energy help, but scale is the dominant force right now.

What can I personally do about it? Use appropriately sized tools rather than the largest model for everything, avoid needlessly repeating heavy requests, consider on-device AI for routine tasks, and favor providers that disclose and reduce their footprint.

The bottom line

AI's intelligence runs on a very physical foundation of electricity and water, and as AI scales, that foundation is growing fast — which is why infrastructure and sustainability pledges are now front-page news. The numbers are big and the estimates imperfect, but the responsible stance is the same: measure it, demand transparency, prefer efficiency, and keep the benefits and costs in view together.

This is analysis based on third-party reports; specific figures are estimates that vary by source and are attributed, not independently verified here.