Why does AI use so much electricity? In plain words

You type a question, press enter, and a few seconds later an answer appears. It feels weightless. Somewhere, a building full of machines just turned electricity into heat to make that happen.

This piece explains why, in plain words, and why the total keeps climbing even as each answer gets more efficient.

What actually happens when you ask

An AI model is a very large set of numbers. When you send it a question, those numbers do not sit still. A specialised chip multiplies your words, turned into numbers, against the model's numbers, layer after layer, to work out the most likely next word. Then it does the same again for the word after that, and again, until the answer is finished.

So a single reply is not one calculation. It is a long chain of them, one round for every word the model writes. A short answer means a short chain. A long report, a block of code or a summary of a fifty-page document means a much longer one.

Pictures, audio and video push this further. A model that reads an image or watches a clip has far more numbers to process than one reading a short sentence, so the same question asked about a photo usually costs more work than the question on its own.

Every one of those calculations moves electrical charge through billions of tiny switches. Moving charge takes energy. That energy ends up as heat, which is why AI chips run hot and why the buildings that hold them need serious cooling.

Three things that make the meter spin

The size of the model. Bigger models have more numbers to multiply for each word. They can give better answers, but each word costs more work.

The length of the conversation. Models re-read what came before to decide what comes next. A long chat history or a large attached file means more reading on every step.

Keeping the machines ready. A chip that is switched on and waiting for the next request still draws power. So do the fans, the pumps and the power equipment around it. Capacity that sits half-used still has to be paid for in electricity.

None of these is wasteful in itself. They are simply what it takes to produce an answer on demand.

Why the total keeps growing

Engineers keep making each answer cheaper in energy terms. Chips get better. Software gets smarter about skipping work it does not need to do. Smaller models get good enough for everyday tasks.

Yet total electricity use for AI keeps rising, for a simple reason: people use it far more as it gets better and cheaper. A tool that once answered a few questions a day now drafts emails, reviews contracts, watches camera feeds and runs inside other software all day long. Each new use adds requests, and every request runs that chain of calculations again.

There is also a shift in what the work is. For the first years of the boom, the headline energy story was training: teaching a new model, a huge job done once. Today most of the work is answering, called inference, and it never stops. It runs every hour of every day, in every time zone, for as long as people keep asking. (We look at the difference in training and answering are different jobs.)

Why this matters to anyone buying AI

You might think electricity is the data center's problem. In practice, it shapes three things a buyer cares about.

Price. Power is a major running cost behind any AI service. When it is expensive or scarce, that shows up in what you pay, one way or another.

Availability. New large data centers often wait a long time for a grid connection. If capacity can only be added where big new power lines are built, it arrives slowly, and buyers compete for what exists.

Location. Because power is the constraint, it increasingly decides where AI capacity can be built. That in turn decides how far your requests travel and where your data is processed. For a worked look at what a unit of power buys, see what a megawatt buys you in AI data center power.

A few plain questions to ask a provider

If you are choosing where your AI work runs, it is reasonable to ask:

• Where is the capacity that serves my requests, and how is it powered?
• Is the model I am using bigger than the task needs?
• Am I paying for capacity that sits idle?
• Can you show me what was actually delivered, and where?

You do not need to understand chip design to ask these. They are the same questions you would ask about any utility you depend on.

The direction this points

If power is the scarce ingredient, the sensible move is to put AI capacity where power is already available, close to the people and machines that use it, and to keep that capacity busy rather than idle. That means many smaller sites alongside the giant ones.

That is the idea behind Starfield. We're building Starfield Fabric to deploy standardized compute close to power, connectivity and demand, then connect it through one programmable network for distributed AI infrastructure. It is designed so workloads are placed by policy, on latency, data residency, model and load, and every workload can generate an auditable record of commitment, placement and delivery.

Electricity is where AI meets the physical world. The more of the world uses AI, the more that meeting point matters.