How Much Energy and Water Does AI Actually Use?

When you send a prompt to an AI, the answer comes back as a tiny puff of steam off a roomful of red-hot computer chips. Each query runs on power-hungry GPUs that turn electricity into heat, and that heat has to be carried away — often with water. At global scale the numbers are large: data centers used roughly 448 terawatt-hours of electricity in 2025 — more than the entire country of Saudi Arabia — and that's projected to nearly double to about 945 TWh by 2030, with AI the fastest-growing slice. A June 2026 United Nations report warned the water side could rival the consumption of over a billion people.

This isn't a reason to feel guilty about asking a chatbot a question. It's worth understanding because data centers are now showing up on your electric bill and in your town's water supply. Here's the physical chain, from your keystroke to the cooling tower.

Table of Contents

1. What Actually Happens When You Send a Prompt
2. Why Chips Need Water
3. The Numbers, in Context
4. Training vs. Asking: Where the Energy Really Goes
5. Why It's Showing Up on Your Bill
6. What's Being Done About It
7. Frequently Asked Questions

---

What Actually Happens When You Send a Prompt

Your phone or laptop barely lifts a finger. The work happens somewhere else:

1. Your prompt travels to a data center — a warehouse full of specialized computers. (If you're fuzzy on how it gets there, see how the internet works.)
2. GPUs do the math. Generating an answer means billions of multiplications across the model's parameters. These run on graphics processors (GPUs) packed by the thousands.
3. Electricity becomes heat. A law of physics you can't dodge: nearly all the electricity a chip draws comes back out as heat. A rack of AI GPUs can pull as much power as several homes — concentrated into a box the size of a fridge.
4. The heat has to leave. If you don't remove it, the chips throttle or fail. So the data center cools them — and cooling is where the water comes in.

Why Chips Need Water

There are two main ways to carry heat away, and both have a cost:

• Air cooling blows air across the chips, but moving that much air takes a lot of electricity.
• Water cooling is more efficient, so big data centers often use evaporative cooling: water is evaporated to shed heat, the same way sweating cools you down. That evaporated water is consumed — it leaves as vapor and doesn't come back to the local supply.

So the water isn't "inside the computer." It's the sweat. Every hot chip is a small furnace, and water is the cheapest way to keep the room from cooking itself.

There's also a hidden, second helping of water: the power plants that generate the electricity often use water for their cooling too. So an AI query draws water twice — once at the data center, once at the power station.

The Numbers, in Context

The verified 2026 picture:

Sources: UN News, June 2026, PBS NewsHour.

A useful gut-check: a single AI query is small on its own — far less than, say, running your oven. The footprint comes from scale: billions of queries a day, plus the enormous one-time cost of training the models in the first place.

Training vs. Asking: Where the Energy Really Goes

It helps to split AI's energy into two buckets:

• Training is the one-time (per model) marathon: running thousands of GPUs for weeks to teach the model. This is the giant, headline-grabbing number — but it's paid once and amortized across everyone who ever uses the model.
• Inference is every individual question you ask afterward. Each one is cheap, but they never stop, and there are billions of them. In 2026, inference is now the larger share of total AI energy, precisely because everyone uses these tools all day.

This is also why efficiency research matters so much. New methods that shrink the compute per query — better quantization, or neuro-symbolic approaches that cut energy up to 100x — attack the inference bucket directly. "Stop throwing more compute at it" is becoming a research goal, not just a slogan.

Why It's Showing Up on Your Bill

Here's the part that turned AI energy from an abstract worry into a kitchen-table issue: when a hyperscale data center moves into a region, it can draw as much power as a small city. That new demand can push up wholesale electricity prices and strain the local grid — and in many places, ordinary customers see it as higher bills. In 2026, dozens of US states introduced legislation specifically about data-center power and water use (Consumer Reports).

So the honest framing isn't "your chatbot question boils a lake." It's "the aggregate buildout of AI infrastructure has real, local costs in power, water, and price — and those costs are increasingly visible to people who never use AI at all."

What's Being Done About It

Three levers are being pulled at once:

• More efficient chips and cooling — liquid cooling that recirculates instead of evaporating, and chips that do more math per watt.
• More efficient models — the 2026 research wave on quantization and structured reasoning cuts the energy per answer without losing quality.
• Cleaner and better-sited power — putting data centers where there's surplus renewable energy and cooler climates that need less cooling.

None of these is a silver bullet, but together they're why the per-query footprint is falling even as total usage climbs.

Frequently Asked Questions

How much energy does one AI query use?

A single query is small — on the order of running a household appliance briefly, far less than heating your home. The environmental story is about scale: billions of queries per day plus the large one-time cost of training each model. Exact per-query figures vary by model size and aren't consistently disclosed, so treat any single number with caution.

Why does AI use water at all?

Because AI runs on chips that turn electricity into heat, and the cheapest way to remove that heat at scale is evaporative cooling — essentially making the data center "sweat." The evaporated water is consumed rather than returned. Power plants that supply the electricity often use water for cooling too, so there's a second, indirect draw.

How much electricity do data centers actually use?

Globally, data centers used roughly 448 TWh in 2025 — more than the entire country of Saudi Arabia — and that's projected to nearly double to about 945 TWh by 2030, according to 2026 reporting, with AI the fastest-growing contributor.

Is AI raising my electricity bill?

It can, indirectly. A large data center can draw as much power as a small city, which can raise wholesale prices and strain local grids. In 2026, many US states introduced legislation on data-center energy and water use precisely because of bill and supply pressure on residents.

Is AI's energy use getting better or worse?

Both. Total usage is climbing as more people use AI, but the energy per query is falling thanks to more efficient chips, better cooling, and 2026 research on model efficiency that cuts compute dramatically. The open question is whether efficiency gains can outpace the growth in total demand.