AI and the planet

We care about what AI costs the planet. Here is what we know, and what we are doing about it.

AI and energy are a big problem in the world and a big question for anyone building with it. This page is our answer: the real numbers, what we do today, and what we are building toward.

The short version

Thirty seconds, if that is all you have.

1One question is small.

About the energy of a 10-watt bulb for 90 seconds, and five drops of water. The bigger issue is how fast AI use is growing worldwide.

2We keep it light.

We never train models. Each Classmate runs on the lightest model that does its job, and heavy “thinking” modes stay off by default.

3We are moving to clean power.

As AI providers run on renewable energy, we shift our systems onto them, and we publish our footprint on this page.

4You can help.

Pick the right Classmate, ask one clear question instead of five vague ones, and skip AI for things you can just look up.

Want the numbers, the sources and the commitments? Keep reading.

The honest numbers

One question, measured.

0.24 WhElectricity for one typical text prompt (Google, 2025). About a 10-watt LED bulb running for 90 seconds. OpenAI’s estimate is similar.
5 dropsWater used at the data centre per prompt. Counting the water behind the electricity too, an honest range is 1 to 5 mL.
1.5%Share of the world’s electricity used by all data centres in 2024 (IEA). AI is a growing slice, and the IEA expects about 3% by 2030.

So one question is small. The real costs are training (a one-off job the model makers pay for, spread across billions of answers) and growth (data-centre electricity is rising about 15% a year, faster than any other sector). Neither is a reason to pretend AI is free. Both are a reason to use it well.

Where we stand

Our goal is to stay ahead of the curve, and to grow as a corporation that is equitable, sustainable and careful with everything we build.

We understand the impact. Every answer a Classmate gives runs on electricity in a data centre somewhere, and as more students use CMPUS, that adds up. We do not think anyone building with AI gets a pass on this. We think the right time to make good choices is early, before they are expensive to change, and that is where we are.

As we grow, our goal is to find and work with the companies providing clean, renewable energy to power AI models. There are many of them, and they are advancing fast. As they do, we will shift our systems to run on them, so that CMPUS can offer AI in a way that is careful with the planet, not just careful with your grades.

Until then, we do the part that is fully in our hands: we ask for the least compute that does the job, we say what we know, and we do not claim more than we have done.

What CMPUS does about it

Our lane is demand: ask for the least compute that does the job.

We train nothingCMPUS runs on models from the major AI labs. Our footprint is only the electricity it takes to answer your question.
Every Classmate is right-sizedModel choice is the biggest lever in AI energy. Each Classmate runs on the lightest model that does its job well; only deep-judgment work goes to the largest models.
Built to answer onceA Classmate asks what you actually need, then answers that. Fewer wasted answers, less wasted energy.
Nothing runs while you sleepNo background agents, no always-on jobs. Energy is spent only when a student is asking a question.

Where the electricity comes from

The models behind CMPUS run in data centres operated by Google, Microsoft and Amazon, which publish renewable-energy and water targets every year. We do not control those grids, and we will not call their goals ours. What we control today is how much compute we ask for. What we are working toward is running on clean, renewable-powered providers as those options mature, and this page will say when we make that move.

The future of sustainable AI

The technologies that make AI cleaner are already being built. We are building toward them.

Five shifts are changing what a single AI answer costs the planet. As each one matures, our systems move onto it.

Renewable-powered data centres

The largest AI providers are matching data centres to wind, solar and hydro, and moving from yearly matching to hour-by-hour clean power. This is the shift we are tracking most closely.

Carbon-aware computing

Non-urgent AI work can be scheduled for the hours and regions where the grid is cleanest. The same answer, powered by wind at night instead of gas at noon.

Smaller, smarter models

Google measured a 33-times drop in energy per median prompt in a single year, from better models and better chips. The lightest model that does the job keeps getting lighter.

Cooling without waste

Closed-loop and liquid cooling cut the water a data centre evaporates, and operators are committing to replenish more water than they consume in the regions they run in.

Measured, then published

The providers now publish per-prompt energy, water and carbon. When every layer is measured, every layer can be held to account. That is what makes our own footprint estimate possible.

What we are working on

Four commitments.

  1. A published footprint estimate. Our interactions multiplied by the published per-prompt figures, posted here after each full term.
  2. No “thinking mode” by default. Extended-reasoning modes can use several times the energy of a normal answer. Classmates use them only for work that needs them.
  3. Clean-powered compute as we grow. We track which AI providers run on renewable energy and move our systems toward them as they mature. Progress, including the honest parts, gets reported here.
  4. This page stays current. When the sources change, this page changes with them. Last updated 11 September 2026.
What you can do

Use it like it costs something. It does.

  • Pick the right Classmate. A specialist answers in one pass. A generalist needs three.
  • One good question beats five vague ones. Say what the assignment is, what you have, and what you are stuck on.
  • Look it up when looking it up is enough. A date, a definition, a formula: your notes or a search engine are faster and lighter than a conversation.
  • Do not regenerate for a nicer word. Edit it yourself. That is also the part of the work that is yours.
Straight answers

Questions students actually ask.

Is one prompt really that small?
For a typical text question, yes: the two largest AI companies both measure it at a quarter to a third of a watt-hour. Long documents, images and extended-reasoning modes cost more, sometimes many times more. That is exactly why CMPUS keeps those modes off by default.
Isn’t training the real problem?
Training is the big one-off cost, and it is real. CMPUS does not train models, so we do not add to it. What we can influence is how much compute is spent answering students, which is the part that grows every time a new user signs up.
Why does CMPUS care about this?
Because the habits are set now. A platform built for students will be used by a lot of students, and the choices about which models run, when the heavy modes switch on, and where the compute comes from are much cheaper to get right early than to fix later. We would rather build the right habits in than make excuses later.
Does CMPUS buy carbon offsets?
No. We would rather reduce the demand than pay to feel better about it. If that changes, it will be written here first.
Which model does each Classmate use?
We do not publish per-Classmate pairings because they change as models improve. The rule does not change: the lightest model that does the job well, with the largest models reserved for deep-judgment work.
Where did these numbers come from?
From the sources listed below: Google’s 2025 measurement of its own fleet, OpenAI’s published estimate, the International Energy Agency’s 2025 Energy and AI report, and a peer-reviewed benchmark of inference energy. Nothing on this page is our own measurement yet; the first one is commitment number one above.

Sources

Comparison: a 10-watt LED bulb uses 0.25 Wh in 90 seconds. Water figures for the electricity supply use the 1.8 to 4.3 litres per kWh range reported for large data-centre operators. Questions or corrections: hello@cmpus.ca.

Use AI well. That includes using less of it.

Sixteen Classmates, each on the lightest model that does its job. Five interactions a day are free, no sign-up.

Enter the CMPUS Meet the Classmates
5 free interactions a dayNo sign-up, no card
Enter the CMPUS