Choosing AI Responsibly: The Resource Costs Behind Each Prompt

Key Takeaways

• AI's footprint depends on the model, task, and resources included in the estimate.

• Training matters, even when a per-prompt figure excludes it.

• More capable models and agents can add computation that routine tasks do not need.

• Choose the simplest tool that does the job reliably.

When I give trainings on AI, concerns about electricity, water, and other resources come up often, especially among younger participants. These are legitimate concerns. The convenience of an AI response does not make the resources behind it disappear.

I take those questions seriously because I use AI in my own work. Understanding its costs helps us decide when it is useful enough to justify them—and when a simpler tool would do.

AI's Resource Costs Are Real

The International Energy Agency's 2026 analysis estimates global data-center electricity use at 485 terawatt-hours in 2025 and projects about 950 terawatt-hours in 2030, roughly 3% of global electricity demand. These totals cover all data-center activities, not AI alone.

In the United States, Berkeley Lab's June 2026 update projects that data centers could account for 11.8% of electricity use by 2030, with scenarios ranging from 9.5% to 15.3%. Those are modeled possibilities, not measured future consumption.

Water also matters. The 2024 Berkeley Lab report estimated that U.S. data centers consumed about 66 billion liters directly for cooling in 2023, plus nearly 800 billion liters indirectly through electricity generation. The indirect estimate uses regional electricity mixes rather than individual facilities' power contracts or onsite generation. Neither figure is an AI-only total.

Local effects depend on where and how facilities operate. The IEA's October 2026 report notes that data centers can strain grids, that effects on other customers' bills depend on how infrastructure costs are allocated, and that cooling systems can trade lower water use for higher electricity use.

Why One Prompt Has No Fixed Footprint

A short email and a lengthy research synthesis are both prompted requests, but they require different amounts of work. Electricity use varies with model choice, input and output length, reasoning, hardware, and supporting infrastructure. Carbon and water impacts also depend on the electricity source, location, and cooling system.

A Google technical report estimated that the median Gemini Apps text prompt in May 2025 used 0.24 watt-hours of electricity and 0.26 milliliters of cooling water. Its measurement included active computing, host systems, idle capacity, and data-center overhead.

That is useful evidence about one product and period. It excludes model training, data storage, end-user devices, external networking energy, and water used to generate electricity. It should not become a universal conversion factor for AI prompts—or a complete account of AI's environmental footprint.

Training Also Belongs in the Picture

AI uses electricity before a customer asks a question. Pre-training develops a model's underlying capabilities, and further training can refine its responses.

Some researchers allocate training impacts across later use. A 2026 life-cycle assessment, for example, showed how allocating training impacts to an AI radio application depends on model lifetime and usage assumptions. This spreads an earlier cost across later services; sending a prompt does not repeat the training.

Users have little direct control over training practices. Providers and institutional purchasers have more influence. Either way, a per-prompt estimate should say whether it includes training.

Why an Agent May Use More Than a Simple Chat

The discussion changes as we move from chatting with AI to asking agents to carry out assignments. A chat might revise one paragraph; an agent might search sources, draft, check, and revise before returning a result. One request can therefore involve many model calls. I've explored agents in my earlier post on moving from chatbots to agents; the point here is that those extra steps can add resource use.

A 2026 Joule modeling study estimated roughly 13 times more electricity for its long-reasoning scenario than its standard-query scenario. That is not a universal agent multiplier. An initial study of web agents also found that greater energy use did not necessarily produce better performance. The added computation needs to provide useful value.

So What Can We Do?

I do not believe AI is going away, and I find it too useful to see avoiding it altogether as a realistic answer for my own work. But using AI does not mean ignoring its resource costs. We can make deliberate choices about which tools we use, how much capability we ask for, and whether the result is worth the resources involved.

Choose the Model That Fits the Task

You do not need a Ferrari or a Humvee for every commute. A Prius or Mini Cooper may get you where you need to go just as safely and on time. The same idea applies to AI: a routine email usually does not need the model or reasoning setting you would choose for a difficult research problem.

Think of smaller options such as OpenAI's GPT-5.4 mini or Claude Haiku as the Prius or Mini Cooper in this analogy. Try an efficient model for polishing an email, shortening a paragraph, or drafting an agenda. Google's Gemma 3n E2B and E4B and Gemini Flash-Lite options offer other examples of efficiency-oriented design, with access differing across apps, developer services, and local installations.

More powerful models, such as Astra used with a high reasoning setting, belong toward the Ferrari or Humvee end of the analogy: their added capability may be useful for difficult analysis, but you do not need it for every trip. Comparing conflicting studies or examining a complicated research design may justify more computation than turning three bullet points into a polite email.

This is an analogy about matching capability to purpose, not a measured electricity ranking. “Instant,” “light,” and “mini” are not environmental certifications, and older does not automatically mean more efficient. Start with an efficient option, check the result, and increase capability when the task requires it.

Practical Choices for Everyday Work

• Use conventional tools for fixed operations. Calculators, spreadsheets, and database filters can handle arithmetic, sorting, and exact matching without language-model reasoning.

• Keep requests focused. Provide relevant material and a clear stopping point. Longer reports, more revisions, and repeated regeneration can add computation.

• Reuse verified work. Securely save useful summaries, templates, and decisions instead of recreating them each time.

• Reserve agents for assignments that benefit from their extra steps. Consider quality, time saved, review effort, and whether the result will be reused.

• Ask vendors for meaningful disclosure. Look for measurement boundaries and task-specific estimates covering electricity, water, and carbon.

Efficiency gains do not guarantee lower total demand: more users and more intensive tasks can outweigh savings per response. Individual choices matter, alongside provider practices, infrastructure decisions, and institutional purchasing.

Final Takeaways

• Treat resource use as part of professional judgment, alongside accuracy, privacy, equity, and usefulness.

• Check what an environmental estimate includes before comparing products or tasks.

• Match the tool and model to the assignment, and consider the work needed to reach a usable result.

• Use more intensive AI when its added capability provides a clear benefit.

AI Disclosure: Generative AI was used to assist with drafting and editing this post and to create the accompanying image. I reviewed and revised the content and take responsibility for its accuracy and final form.