Short answer: AI uses a lot of energy in aggregate, because large models run in data centers packed with power-hungry chips, but the AI energy use behind one small business's everyday work is usually small next to its lighting, cooling, vehicles and shipping. You don't need to treat it as a crisis of your own making. Treat it like any other utility: use what the job needs, don't waste it, and pick the efficient option when it costs no more.

That's the calm version. This post explains where the energy actually goes, why the figures you read in headlines disagree so much, when your share starts to matter, and the practical choices that cut both your AI bill and its footprint. Conveniently, they are mostly the same choices.

Where does AI's energy actually go?

There are two very different phases, and it helps to keep them apart.

Training is building the model. A company feeds enormous amounts of text, images and code through thousands of specialized chips for weeks or months. Those chips are usually GPUs (graphics processing units), originally designed for video games and now the workhorse of AI. Training is a large energy cost, but it's paid by the model maker, mostly once per model.

Inference is using the model. Every time someone asks a question, the model runs a calculation to produce the answer, piece by piece. A single request is small. But billions of requests happen every day across the world, so inference adds up, and it grows with every new user and every new automated workflow.

Around both sits the facility: power for the building, and cooling to stop the chips overheating. Depending on the design and the climate, cooling can use a meaningful amount of electricity and, in some designs, water.

An everyday analogy: training is building a factory, inference is running the production line. The factory build is huge but happens once. The line runs every day, and your business only ever touches the line.

Why do the numbers you read vary so wildly?

You'll see headlines comparing one AI question to a light bulb, a phone charge, a kettle or a web search. Read them with caution. The honest answer is that per-query figures vary by a factor of many, for good reasons:

  • Model size. The largest models need far more computation per answer than small ones. "AI" covers both.
  • Length and type of output. A one-line reply is cheap. A long report, an image, or a video clip is much more expensive.
  • "Thinking" modes. Some models work through many hidden steps before answering. That improves hard answers and multiplies the computation.
  • Hardware and building efficiency. Newer chips and better-designed data centers do more work per unit of power.
  • The local power grid. The same electricity is cleaner or dirtier depending on how it's generated where the data center sits.
  • Limited disclosure. Providers publish little detail, so many estimates are careful guesses built on assumptions.

So we won't give you a number per question, and you should be skeptical of anyone who quotes one with precision. What is reliable is the direction: bigger models, longer outputs, images and video, and heavy reasoning use more. Smaller models and short, focused tasks use less.

Should a small business worry about its AI energy use?

Worry in proportion. For a team using a chat assistant to draft emails and summarize documents, AI is almost certainly a minor line in your overall footprint. Your premises, travel, deliveries and the hardware you buy matter more.

It becomes worth real attention in a few situations:

  • High-volume automation, such as running thousands of documents a day through a large model.
  • Routine image or video generation, where each item costs far more than a text reply.
  • Sustainability reporting. Large customers increasingly send suppliers emissions questionnaires, and cloud and AI services may show up as part of your indirect emissions. What you must report depends on your jurisdiction, your size and your customers' requirements, so check the rules that apply to you; this isn't legal or compliance advice.
  • Your brand promise. If sustainability is part of why customers buy from you, your AI choices should be consistent with it.

Here's a rough way to place your own usage:

How you use AIRelative energyWorth managing?
Staff drafting and summarizing with a chat assistantLowRarely; focus on using it well
Short automated text tasks (sorting emails, tagging tickets)Low to moderateYes, by using a small model
Long documents through a large model, at volumeModerate to highYes: right-size, trim and cache
Generating images or video regularlyHigh per itemYes: generate deliberately, not in bulk
Training or fine-tuning your own modelHigh upfrontRarely needed by small businesses

How does right-sizing a model cut energy and cost at once?

Most AI services charge by tokens, the small chunks of text (roughly pieces of words) a model reads and writes. More tokens through a bigger model means more computation, which means more money and more energy. Your AI bill is a decent proxy for your AI energy use, which makes this easy: the levers that lower one lower the other.

  1. Use a small model for simple steps. Classifying an email, pulling an invoice number out of a PDF, or routing a ticket rarely needs the biggest model available. Our guide to small language models for business covers when a smaller model is enough.
  2. Send only what's needed. Pasting a 200-page manual into every request is wasteful. Retrieving the three relevant pages does the same job with a fraction of the work.
  3. Cache repeat answers. If a hundred customers ask the same question, the answer doesn't need computing a hundred times.
  4. Batch work that isn't urgent. Many providers offer slower, cheaper batch processing for overnight jobs.
  5. Fix runaway loops. A badly designed AI agent that retries a task twenty times wastes twenty times the energy. Good monitoring catches this.
  6. Switch off heavy reasoning for easy tasks. Save the "think hard" mode for questions that need it.

We go deeper on these in how to control a growing AI bill. The pleasant surprise for most businesses is that efficiency here isn't a sacrifice. It's usually faster, cheaper and just as accurate.

Does choosing a "greener" AI provider make a difference?

Somewhat. Providers differ in how efficient their data centers are and where their electricity comes from. If it matters to you, ask:

  • Which regions your workloads run in, and whether you can choose.
  • Whether they publish a sustainability or environmental report, and what it actually measures.
  • Whether their renewable energy claims are based on matching consumption or on buying certificates elsewhere.

Be wary of vague claims like "carbon-neutral AI" with nothing behind them.

What about running a model on your own server instead? It isn't automatically greener. Large data centers are often more efficient per calculation than a server in an office that sits idle most of the day. Running your own model can make sense for privacy, control or cost at volume, which we weigh up in open-source models vs API models, but energy alone is rarely a strong reason.

What about water, power grids and the bigger debate?

The larger debate is real. Data centers use substantial electricity, some use water for cooling, and their growth is putting pressure on power grids in some regions. Utilities, governments and the large technology companies are the ones who will settle questions about new power plants and grid capacity.

A small business's lever on that debate is small but not zero. You can buy sensibly, ask providers questions, avoid generating content nobody needs, and prefer efficient designs. That's a proportionate response. Neither guilt nor dismissal is useful here.

It's also fair to count the other side of the ledger, carefully. If an AI workflow genuinely removes repeat site visits, printing, couriered documents or wasted delivery miles, that has an environmental value too. Measure it before you claim it.

How should you talk about AI and sustainability with customers?

Accurately and modestly. Statements like "we use right-sized models and prefer providers that publish environmental data" are honest and checkable. Statements like "our AI is carbon neutral" usually aren't, and environmental marketing claims are regulated in a growing number of places.

If you need a starting point for your own records, a short checklist works:

  1. List every place your business uses AI, including tools staff signed up for themselves.
  2. Identify the one or two highest-volume workflows. That's where nearly all your AI energy use sits.
  3. Check which model size each of those uses, and test a smaller one on a sample of real tasks.
  4. Note each provider's published sustainability information, if any, for your reporting file.
  5. Review once or twice a year. Models and data centers get more efficient, so yesterday's sensible choice can become today's waste.

For the full cost picture beyond energy, see what an AI system really costs after it's built. For where AI fits in your business at all, start with our complete guide to AI for business.

Frequently Asked Questions

Does using a chat assistant waste a lot of energy?

For ordinary text tasks, one person's daily use is small. It adds up across millions of users, which is why the industry figure is large, but cutting your team's email drafting won't move your footprint much. Wasteful patterns, such as generating dozens of images to pick one, matter more.

Is training AI worse for the environment than using it?

Training a large model uses a lot of energy in one concentrated burst, but use across the world happens constantly and keeps growing. Over a model's life, everyday use can account for a large share of the total. For your business, use is the only part you influence.

Will AI energy use make AI more expensive?

Energy is one of the costs providers pass on, but chips and data centers also keep getting more efficient. Prices have moved in both directions for different products, so don't plan around a prediction. Plan around efficient design, which protects you either way.

Is running AI on my own computer more environmentally friendly?

Not necessarily. A large, well-run data center usually does more work per unit of power than a single office machine. Local models make sense for privacy or offline use; they're not a reliable green choice on their own.

Do I need to report AI energy use?

It depends on where you operate, your size, and what your customers require in their supplier questionnaires. AI often falls under purchased cloud services rather than a separate category. Check the current rules with your accountant or sustainability adviser; this post isn't legal or compliance advice.

Want AI that does the job without the excess?

If you're running AI at volume and suspect you're using a sledgehammer for jobs that need a screwdriver, tell us about your workflows. We design systems that route each task to the smallest model that does it well, which is kinder to your budget and the grid alike. See how we approach agentic AI development.