Every AI answer is computed somewhere, on chips in a data centre, using electricity. You do not see that when you type a question. You do see it indirectly, in the price of a plan, a delay in a reply, and in news about data centres.
This post explains, as of November 2026, what is known about the energy and compute behind AI, what drives the cost, and what a business can do to use less. It leans on published research and the numbers companies have chosen to release. Where the evidence is thin or comes from the companies themselves, we say so.
Training and inference: two different bills
Training is the one-time, very large effort of building a model: running huge amounts of data through thousands of chips for weeks or months. Only a few organizations do it for the biggest models.
Inference is using the finished model: every question answered, every email drafted, every record classified. Each use is small. Together they add up, because there are many uses and they keep happening.
For a business that buys AI as a service, training is somebody else's cost. Your cost is inference, and it grows with how much you use, how long your prompts are, and how large a model you call.
The big picture, with sources
The International Energy Agency's 2025 report *Energy and AI* estimated that data centres used around 415 terawatt-hours (TWh) of electricity in 2024, about 1.5% of global electricity consumption. In its base case, that rises to around 945 TWh by 2030, just under 3% of the total, growing about 15% a year. It said servers built for AI, "accelerated" servers, are growing faster, at about 30% a year, and account for almost half of the net increase. These are projections, and the IEA presents them as scenarios, not forecasts of a fixed future.
So the footprint is real and growing, and it is also a small share of global electricity. Both statements are true, and a headline that says only one of them is incomplete.
What one prompt uses
Companies rarely publish this, which is why one disclosure drew attention. In August 2025, Google published a paper and blog post estimating that a median text prompt to its Gemini app used 0.24 watt-hours of energy, produced 0.03 grams of carbon dioxide equivalent and used about 0.26 millilitres of water. It said this counted the AI chips, the host machines, idle capacity kept ready, and data-centre overhead, and did not include training, network or your own device. It also reported that energy per median prompt fell 33 times over twelve months, through software improvements and cleaner energy. Google noted these figures were not independently verified.
Take that as one company's number for one product's typical prompt, not a rule for every model or every task. Long documents, images, video and multi-step agents use more. We will not turn it into a per-employee estimate, because it would be a guess.
What drives cost and energy
- Model size and type. A 2023 study by Sasha Luccioni, Yacine Jernite and Emma Strubell ("Power Hungry Processing") measured the energy and carbon needed for 1,000 inferences across tasks. It found that multi-purpose, generative models were orders of magnitude more expensive than task-specific systems for a variety of tasks, even when controlling for parameter count. A single-purpose tool can be far cheaper than asking a general model to do the same thing.
- How much text goes in and out. Longer prompts and longer answers cost more.
- How many steps an agent takes. An agent that re-reads your whole file before each step, or loops until it is satisfied, multiplies the usage.
- Where and how it runs. A 2021 paper led by David Patterson (Google and UC Berkeley), "Carbon Emissions and Large Neural Network Training", estimated that the choice of model design, data centre, processor and location each change the footprint, with the combination making up to a 100 to 1,000 times difference for training. The same idea applies broadly: who hosts the model and on what hardware matters.
- Idle capacity. Systems kept ready for traffic that might arrive use energy even when quiet. This is part of why the Google figure counts idle machines.
Price is not the same as energy, and a provider's price reflects many things. But the factors that drive one tend to push the other, so lower usage generally helps both.
Four ways to use AI efficiently
1. Right-size the model
Use the smallest model that passes your tests for each task. Sorting emails into five categories rarely needs the most capable model. Reserve a larger one for hard drafting and judgment calls. The same testing set described in our post on choosing a model tells you where the line is.
2. Cache what repeats
If every request starts with the same long instructions, price list or policy text, you are paying to process it every time. Some providers offer prompt caching, where a repeated beginning is stored briefly and re-used. Anthropic's documentation, for example, lists cache reads at a fraction of the normal input price (0.1 times for the standard rate, lower for some models) with a default cache lifetime of five minutes. Check your provider's current terms, because they change.
3. Batch what can wait
Not everything needs an instant answer. Month-end coding, catalogue clean-up and overnight reports can be queued and processed together. Anthropic's batch interface, for example, advertises a 50 percent cost reduction with most batches finishing in under an hour. Other providers have similar options. Batching trades speed for cost.
4. Skip the model when code will do
A formula, a database query or a rule such as "invoice total equals the sum of its lines" is faster, cheaper and exactly repeatable. Use the model for the part that needs language or judgment, and plain code for the rest. This is also good for reliability.
A few smaller habits help too: keep prompts short, send only the records the task needs, stop agents that loop without progress, and measure usage per finished task so you notice when it creeps up.
What this means for your business
- Expect usage to be a real line item. Treat AI like any metered service: know what you use and what a finished task costs.
- Ask vendors how usage is measured and priced. Is it per seat, per volume, or per result? What happens at the limit?
- Efficiency and quality often agree. A shorter, cleaner prompt and a smaller, tested model are usually cheaper and easier to check.
- Be cautious with claims in either direction. "AI will drain the grid" and "a prompt is basically free" both leave out context. Look for the source, the date and what is counted.
- If you have sustainability reporting duties, ask your provider what data they can give you, and note that published figures are often averages from the provider, not measurements of your usage.
How this shows up at Humonix
AI use is included in each plan at the listed volumes. Beyond them, it is billed at the provider's list price with no markup, or runs on your own model key. That means the incentive is the same as yours: work that needs fewer steps costs less, so we keep routine checks as code and use models for language and judgment. See Hue and Bring your own.
We have not measured the energy use of Humonix itself and do not publish a figure for it. If we do, we will say how it was measured.
Sources
- International Energy Agency, "Energy and AI" (2025): energy demand from AI
- Elsworth et al., "Measuring the environmental impact of delivering AI at Google Scale" (arXiv:2508.15734, August 2025)
- Luccioni, Jernite, Strubell, "Power Hungry Processing: Watts Driving the Cost of AI Deployment?" (arXiv:2311.16863; ACM FAccT 2024)
- Patterson et al., "Carbon Emissions and Large Neural Network Training" (arXiv:2104.10350, 2021)
- Anthropic documentation: prompt caching
- Anthropic documentation: batch processing
This post is general information, not technical or legal advice.