A technician walks a dim server corridor lined with blinking racks, illustrating the vast physical infrastructure powering every AI interaction.

The Hidden Cost of AI: Energy, Water, and Smarter Business Use

by

You glance at the credit card statement. There it is — $20 a month for AI. Maybe $30. Maybe more if you’re running it across a team. Seems reasonable for everything it does.

What you don’t see on that statement is the other bill. The one nobody sends you. The one paid in electricity, fresh water, and carbon — by the infrastructure that processes every single query you send.

That’s what this article is about. Not to scare you off AI — it’s a genuinely useful tool. But if you’re using it in your business and you don’t understand what’s running under the hood, you’re flying blind on a cost that’s growing fast.

Imagine you’ve had a productive morning. You translated three supplier documents, generated a batch of product images for social media, had a long back-and-forth with an AI tool to refine a strategic report, and left a reasoning model running to analyze a competitor’s pricing.

It felt efficient. It felt cheap.

Then your accountant calls. Utility bills for the office are up again. And someone on the team mentions that their home electricity bill has jumped, too. Nobody connects the dots. But the dots are connected.

The Bill You Never See

Every AI query you send travels to a data center — a large physical facility packed with specialized computer hardware, cooling systems, and power infrastructure. These aren’t regular servers. They run on GPUs (graphics processing units), chips designed to perform the massive parallel calculations AI requires. GPUs consume significantly more electricity than the standard chips in everyday computers.

According to the International Energy Agency’s 2025 “Energy and AI” report, data centers globally consumed approximately 415 terawatt-hours (TWh) of electricity in 2024. That’s roughly 1.5% of all electricity used on the planet. The IEA projects that figure will more than double — to around 945 TWh — by 2030, with AI workloads driving the majority of that growth.

In the United States alone, data centers consumed about 183 TWh in 2024 — around 4% of all U.S. electricity. That’s more than enough to power a country the size of Pakistan for a year.

And that demand is landing on your utility bill.

Your Electricity Bill Is Involved

This isn’t abstract. The Pew Research Center analyzed U.S. energy market data and found that in the PJM electricity market — covering 13 states from Illinois to North Carolina — data center growth added an estimated $9.3 billion to capacity costs for 2025–26. The result: average residential bills in western Maryland rose by about $18 a month. In Ohio, about $16.

Carnegie Mellon University researchers estimate that data centers and cryptocurrency mining combined could push the average U.S. electricity bill up by 8% by 2030. In the highest-demand markets — central and northern Virginia, where data centers now account for roughly 40% of the state’s total electricity consumption — that increase could exceed 25%.

The average U.S. household was already paying about $142 a month for electricity in late 2025, up 25% over the prior decade. Electricity prices are expected to keep outpacing inflation at least through 2026, according to the U.S. Energy Information Administration.

Note: The sharpest electricity bill impacts are concentrated in states with heavy data center activity — Virginia, Texas, Ohio, and western Maryland. If your business operates in those areas, you’re likely already feeling this.

Training vs. Using: Where the Cost Lives

Most people assume the environmental cost of AI is in building the model — that one-time, energy-intensive process called training. Training is genuinely expensive. Building a large AI model can consume more energy than dozens of round-trip transatlantic flights.

But training happens once. Or occasionally, when a model is updated.

Inference — the process of running a model to generate a response when you send a query — happens billions of times a day, every day, forever. That’s where the real ongoing cost lives.

Researchers estimate that inference now accounts for 80 to 90% of total AI energy consumption in aggregate. The cumulative energy from daily queries rapidly outpaces the one-time training cost. One estimate found that 121 days of serving GPT-4 queries generates the same carbon emissions as training the model once — and as usage grows, that break-even point arrives faster.

In other words: every time you and your team open a chat window, you’re adding to that total.

Not All AI Costs the Same

Here’s where it gets practical — and where your choices actually matter. The energy cost of an AI query varies enormously depending on what you’re asking it to do and which mode you’re running it in.

Standard Text Queries

A typical ChatGPT text query uses approximately 0.34 watt-hours of electricity, according to estimates from Epoch AI and OpenAI’s own disclosures. Google’s Gemini comes in slightly lower at around 0.24 watt-hours per median prompt, based on Google’s August 2025 environmental disclosure — the most comprehensive per-query data any major AI provider has published.

For reference, a standard Google search uses about 0.03 watt-hours. A standard AI text query uses roughly 8 to 10 times more than that. Still relatively small for a single interaction — but the scale of daily business use adds up fast.

Reasoning Mode: A Different Animal

Some AI tools now offer a “reasoning” or “extended thinking” mode — you’ll see this in OpenAI’s o1 and o3 models, and in similar features from other providers. These models don’t generate a single response. They first produce a lengthy internal chain-of-thought — working through the problem step by step, generating hundreds or thousands of hidden calculations — before delivering their visible answer.

That process is dramatically more energy-intensive. A 2025 benchmarking study of 30 AI models found that o3 consumed over 33 watt-hours per long prompt — more than 70 times the energy of smaller standard models doing the same task. Separate research from the Hugging Face AI Energy Score project found reasoning models consume approximately 100 times more energy per 1,000 prompts than standard models.

Even for simple questions, reasoning models generated an average of 543 additional hidden tokens per query. Standard models averaged about 38. That invisible computation is real electricity, even when you never see it.

Image Generation

Generating an image with AI uses approximately 1.2 watt-hours — roughly 30 to 50 times the energy of a standard text query. A 2024 study by researchers at Hugging Face and Carnegie Mellon University found image generation consumed a median of about 1.35 kWh per 1,000 inferences, compared to 0.042 kWh for text generation. That’s a 30x gap.

The difference is in how the models work. Image generation processes millions of pixels repeatedly, iterating across the entire image with each pass. It’s computationally intensive in a way that text is not.

Video Generation

If image generation is expensive, video generation is in another category entirely. A five-second AI-generated video clip can require anywhere from 30 to 950 watt-hours, depending on resolution and model — potentially thousands of times more than a standard text query.

This doesn’t mean don’t use it. It means use it deliberately.

Water: The Hidden Resource

Energy isn’t the only resource data centers consume. Cooling those facilities requires enormous quantities of fresh water — primarily through evaporative cooling systems that draw water in, absorb heat, and release it as vapor.

Google disclosed that its global data center operations consumed approximately 8.1 billion gallons of water in 2024, with about 95% used at data centers. That was up 8% from 2023, continuing a multi-year trend Google directly attributes to AI workload growth.

Per individual query, on-site water use is small. Google disclosed that Gemini uses approximately 0.26 milliliters of water per median text prompt. But multiply that by billions of daily queries — across ChatGPT, Claude, Gemini, Copilot, and every other AI tool running simultaneously — and the aggregate is significant.

Morgan Stanley projects that AI data centers’ global water consumption will surpass 1 trillion liters annually by 2028 — about eleven times 2024 levels. About two-thirds of data centers built since 2022 are already located in regions experiencing water stress. The water isn’t being drawn from surplus.

Image generation uses considerably more water per task than text — estimated at roughly 23 milliliters per generated image once energy consumption is factored into water equivalents. That’s another reason to treat visual AI generation as an intentional action rather than a casual one.

Using AI Smarter: Eight Practical Tips

None of this means you should use AI less. It means you should use it better. The same habits that reduce energy consumption also reduce cost, improve output quality, and make your team more intentional about how they work.

1. Ask Whether AI Is the Right Tool

Before opening a chat window, ask whether a regular web search, a spreadsheet formula, or an existing template would do the job just as well. AI tools are genuinely powerful — but using a large language model to find a phone number or check a weather forecast wastes computing power that wasn’t needed.

Computer scientist Günter Klambauer of Johannes Kepler University puts it bluntly: asking a chatbot basic questions is like taking a Concorde to travel to your supermarket. The tool is capable of far more, and burning that capacity on low-value tasks is inefficient.

For web searches where you don’t need AI synthesis, turning off AI-generated overview features — like Google’s AI Overviews or Bing’s Copilot Search — cuts the energy cost of that search significantly. Most browsers let you select “Web results only.”

2. Match the Model to the Task

Not every AI task needs a large general-purpose model. For repetitive, defined tasks — translating documents, summarizing text, answering structured questions — smaller specialized models use a fraction of the energy with equal or better accuracy on those specific jobs.

A 2025 study by researchers at University College London, published through UNESCO, tested this directly. Switching from a large general-purpose model to smaller, task-specific models reduced energy use by up to 90% for summarization and translation tasks — without sacrificing performance. The smaller models used 15 times less energy for summarization, 35 times less for translation, and 50 times less for question-answering.

IBM makes the same recommendation for business deployments: use the smallest capable model for each defined use case. The largest models are built for open-ended complexity. If your task isn’t complex and open-ended, you don’t need one.

3. Control How Long the Response Is

This is the single most impactful thing most users can do right now. The energy cost of an AI response is driven primarily by output length — not by the length of your prompt.

The UCL/UNESCO research found that instructing a model to halve its output reduced energy consumption by approximately 50%. Shortening the prompt itself? Only about a 5% reduction. The output is what costs.

Build output constraints into your prompts as standard practice:

  • Respond in three bullet points
  • Keep this to 150 words
  • Give me one paragraph
  • Five bullets max

You’ll also get cleaner, more usable responses. Constraints improve AI output quality — they force the model to prioritize. This is a win from every angle.

And if you see what you need before the model finishes generating, hit Stop. Every token it generates beyond what you actually read is energy spent for nothing.

4. Turn Off Reasoning Mode for Simple Tasks

Reasoning mode is a genuinely useful capability for genuinely complex problems — multi-step financial analysis, legal reasoning, debugging code logic, building a strategic framework. Use it for those.

Don’t use it for writing an email, summarizing a document, translating a phrase, or answering a factual question. For those tasks, reasoning mode uses 10 to 100 times more energy than standard mode for no meaningful improvement in output quality.

Check which mode your AI tool defaults to. Some have reasoning or “extended thinking” on by default, and many users don’t realize it’s running. Switch it off for routine work.

5. Be Intentional About Images and Video

Image generation uses roughly 30 to 50 times more energy per task than a standard text query. Video generation can use hundreds to thousands of times more. These are not comparable operations.

That doesn’t mean stop generating images. It means treat it as the resource-intensive action it is. UCL researcher Ivana Drobnjak recommends:

  • Generate images only when genuinely needed — not as a casual experiment
  • Start at low resolution and escalate only if the direction is right
  • Edit an existing image rather than generating a new one from scratch — always lighter computationally
  • Batch multiple image requests into a single session rather than separate requests

For video, the same principle applies with even more weight. AI-generated video is a powerful tool. It’s also the most resource-intensive thing you can ask an AI to do. Use it with intention.

6. Stay in One Conversation Per Task

Every time you start a fresh chat session on a topic you’ve already been working on, the AI rebuilds context from zero. That repetition adds to cumulative energy use.

For an ongoing project — drafting a proposal, refining a campaign, working through an analysis — stay in one conversation per task and carry the context forward. Your prompts will be shorter, your results will be better, and the overhead will be lower.

7. Write Intentional Prompts

A vague prompt generates a vague response, which generates a correction request, which generates another response, which generates another correction. Each exchange in that loop costs energy.

A well-crafted initial prompt — specific about goal, format, constraints, and context — typically produces a usable result in one or two exchanges instead of six. That’s a meaningful difference at scale across a team.

Practical habits:

  • State the goal and format in the first message
  • Add constraints upfront — length, tone, structure
  • Give relevant context once rather than feeding it in pieces
  • Ask the AI to classify or prioritize rather than “analyze” — less intensive computation

8. Build an Internal AI Use Policy

This one is for managers and business owners with teams. Individual AI habits multiply. What one person does casually, ten people do at scale.

A simple internal AI use policy doesn’t need to be complicated. It can cover:

  • Which tools are approved for which tasks
  • Guidance on when to use reasoning mode (and when not to)
  • Output length expectations and prompt standards
  • When to use AI vs. a regular web search or existing template

According to PwC’s 2025 Responsible AI Survey, only about 61% of organizations have AI integrated into their operations in any kind of strategic way. Most are still using AI ad hoc, without guidelines. A basic policy pays for itself quickly — in reduced API costs if you’re billed per token, in better output quality, and in lower overhead across the board.

Why Model Choice Matters

Not all AI providers are equally efficient, and those differences are meaningful at scale. Google is currently the most transparent about per-query environmental data, having published comprehensive figures for Gemini in 2025.

A 2025 benchmarking study of 30 models rated them on “eco-efficiency” — balancing performance against energy cost. Anthropic’s Claude 3.7 Sonnet ranked highest in eco-efficiency. OpenAI’s o4-mini and o3-mini also ranked well. DeepSeek models ranked lowest, partly because they run on China’s carbon-intensive power grid in addition to being computationally heavier.

These rankings will shift as models evolve. But the principle is durable: when you’re choosing between capable tools for a defined task, efficiency is a legitimate selection criterion — not just for environmental reasons, but because more efficient models typically cost less to run via API, and their output arrives faster.

The Rebound Effect — and Why Habits Still Matter

You might be thinking: AI models are getting more efficient all the time, so does any of this really matter?

Yes. And here’s why.

Google reported a 33-fold improvement in Gemini’s energy efficiency over a single year. That’s real progress. But total AI energy demand is still rising sharply — because the number of queries, the complexity of tasks, and the number of people using AI are all growing faster than the efficiency gains.

Economists call this the rebound effect: when a tool becomes cheaper to use, people use it more, and total consumption increases even as unit cost falls. It’s happened with fuel-efficient cars, with LED lighting, and now with AI.

Individual habits don’t offset infrastructure-level decisions. But individual habits multiplied across millions of business users are not nothing. And the habits that reduce energy use — intentional prompts, appropriate model selection, controlled output length — are also the habits of a more effective AI user. There’s no trade-off here between responsibility and productivity. They point in the same direction.

Frequently Asked Questions

Q: Does my individual AI use actually make a meaningful difference?

Answer: Per interaction, any one person’s AI use is a small fraction of their overall footprint. The real impact is cumulative — across a team, a company, an industry. The main argument for responsible use isn’t environmental guilt.

It’s that the same habits that reduce energy use also reduce cost and improve output quality. They’re worth building regardless.

Q: Is ChatGPT worse for the environment than Google or Claude?

Answer: Not significantly, for standard text queries. ChatGPT (GPT-4o) estimates run around 0.34 watt-hours per query; Google Gemini clocks in at about 0.24 watt-hours per median prompt per Google’s own disclosure. Both are in the same range.

The bigger differences are between task types — text versus image versus video — and between standard versus reasoning modes. Those gaps are far larger than the differences between providers doing the same kind of task.

Transparency varies: Google currently publishes the most detailed per-query environmental data of any major provider.

Q: Why does asking AI for a shorter answer actually matter?

Answer: Because the output is what costs. UCL researchers found that instructing a model to cut its response length in half reduced energy consumption by approximately 50%. Shortening your prompt only reduced it by about 5%.

The model’s computation is primarily driven by how much it generates, not how much you write. Short, constrained prompts produce shorter responses — which is both more efficient and usually more useful.

Q: What is a reasoning model and when should I actually use one?

Answer: A reasoning model (like OpenAI’s o1 or o3) generates a long internal chain-of-thought before producing the answer you see. That hidden deliberation consumes 10 to 100 times more energy than a standard model for the same task — even for simple questions.

Use reasoning mode for genuinely complex, multi-step analytical work: legal analysis, financial modeling, debugging, strategic planning. For writing assistance, summarization, translation, or basic Q&A, standard mode produces equivalent results at a fraction of the cost.

Q: How much more energy does AI image generation use compared to text?

Answer: Roughly 30 to 50 times more per task, based on research from Hugging Face and Carnegie Mellon University. A single text query uses around 0.34 watt-hours. A single generated image uses approximately 1.2 watt-hours.

AI-generated video is far heavier still — a five-second clip can require 30 to 950 watt-hours depending on resolution and the model used. The math matters when your team is generating images or video at volume.

Q: Is the water consumption from AI data centers actually a real problem?

Answer: It depends on geography. Data centers use water for evaporative cooling — water that’s consumed, not returned to the source. About two-thirds of data centers built since 2022 are in regions already experiencing water stress.

Per individual query, on-site water use is small. At aggregate scale — billions of daily queries — it’s significant. Morgan Stanley projects AI data centers’ global water consumption will exceed 1 trillion liters annually by 2028, roughly eleven times 2024 levels. The concern isn’t any one query. It’s where the infrastructure is being built and what local water systems can sustain.

Q: Why is my electricity bill going up — is AI really causing it?

Answer:  It’s a contributing factor. Data center growth is a meaningful driver of rising electricity costs in the United States, particularly in regions with high data center density. In the PJM electricity market — covering 13 states from Illinois to North Carolina — data center demand added approximately $9.3 billion to capacity costs for 2025–26.

Carnegie Mellon University projects an average 8% U.S. electricity bill increase by 2030 from data centers and crypto mining combined. It’s not the only factor — grid infrastructure upgrades and extreme weather events also play a role — but data centers are, in the words of one University of Pennsylvania energy expert, “pretty much the whole boat when it comes to increases in electricity demand.”

Q: What can I do right now that will have the most impact?

Answer: Two things, starting today. First, add output constraints to every AI prompt — tell it how long to be. The research is clear: shorter outputs are the single biggest lever available to an end user.

Second, audit whether your team is using reasoning mode by default. If it’s on for routine tasks, turn it off. Those two habits alone will reduce both your energy footprint and your AI costs, with no loss in output quality for everyday work.

Looking at Your Next Login Differently

That productive morning — the translations, the images, the reasoning model running in the background — wasn’t wrong. It was useful work. The point isn’t to do less of it.

The point is to do it with your eyes open.

Every output constraint you add, every time you match the model to the task, every image you generate at the right resolution instead of the highest available — those are small decisions that add up across a team, a year, a business.

And here’s the thing: none of these habits cost you anything. They don’t slow you down. They actually make your AI use sharper, faster, and more focused. The same discipline that makes you a better AI user also makes you a more responsible one.

Tomorrow morning, when you open that chat window, what’s the first prompt you’re going to write differently?

Sources: