AI projects · Understand · weekend
How much energy does one AI query cost?
These projects are not aligned to any state standards or curriculum. They are generally about learning, understanding, using, and evaluating AI.
Ages 11-14
Overview
AI tools use electricity in data centers and on your device. Exact "one ChatGPT query costs X Wh" headlines disagree because methods differ. In this project you build a transparent estimate from numbers you can show, compare units carefully, and separate what you measured from what you only read.
Age band
Ages 11–14 (middle school). Comfortable with multiplying decimals and converting units (Wh ↔ kWh).
Time needed
Weekend (about 3–5 hours).
AI goal
Understand — estimate energy with clear assumptions, and refuse fake precision.
Materials
Free tools only; no paid accounts required.
- A calculator or spreadsheet.
- Your device's battery stats if available (laptop battery report, phone battery screen) — optional.
- Free public sources you can open without an account. Prefer primary documents. Examples of kinds of sources (verify the live URL yourself; mark any link you cannot open as UNVERIFIED):
- An electricity bill or utility rate card for your area (kWh price) — optional, for cost conversion.
- A public tech-company or research note that states a method for estimating AI inference energy. If you cannot verify a URL, write
UNVERIFIEDnext to it and do not invent a number.
- Notebook for assumptions list.
You will not invent a universal Wh-per-query figure. If a blog states one without a method, treat it as a claim to check, not a fact.
Steps
- Define the unit. Pick one: watt-hours (Wh) per request, or joules per request (1 Wh = 3600 J). Stick to it.
- Device-side lower bound (optional but concrete). If you can, note your laptop power draw while idle vs while a chat page is open (some laptops show watt estimates in OS battery settings; if not available, skip and say so). Multiply watts × seconds of a single short query ÷ 3600 to get Wh. Label this "device only."
- Server-side estimate — method, not myth. Find one publicly verifiable source that explains how they estimate model inference energy. Copy their stated figure only if you can quote the document and date. If you cannot verify the page, write the citation line as
UNVERIFIEDand leave the number blank. - Build a sandwich estimate. Write:
- Low estimate = device-only (or "unknown")
- Mid estimate = device + cited server figure (if verified)
- High estimate = mid × 3 as a caution multiplier for networking and overhead you did not measure — label it as a guess multiplier, not a measurement
- Compare to something familiar. Convert your mid estimate to "seconds of a 10 W LED bulb" using: seconds = (Wh_estimate / 10 W) × 3600. Show the algebra.
- Write the honesty paragraph. List every assumption. State: "I did not measure the data center myself." State what would be needed for a better number (vendor transparency, measurement tools).
Ethics and safety note
Do not scrape or attack any service to "measure" it. Do not present a single viral number as settled science. Energy and climate claims deserve careful units; wrong units mislead people.
What done looks like
| Criterion | Not yet | Done |
|---|---|---|
| Units | Mixes W, Wh, and kWh without conversion | One primary unit + shown conversions |
| Sources | Uncited round number | Verified cite or explicitly UNVERIFIED |
| Assumptions | Hidden | Numbered list of assumptions |
| Comparison | None | Bulb-seconds (or similar) with algebra shown |
Adaptations
Classroom
Provide a shared "assumption card" so groups differ only in the cited source they check.
Homeschool
Use a physical 10 W bulb and a stopwatch analogy for the comparison step.
Weekend
Skip device measurement; focus on source checking + unit conversion + honesty paragraph.
Research
Compare two verified sources' methods side by side; explain why their numbers disagree without picking a "winner" unless methods are comparable.
Science fair
Poster section titled "Numbers I refuse to invent" next to the estimate sandwich.