The choices that actually change the number

Where the leverage is

The previous lesson placed the big decisions with utilities and planning authorities. This one is about the decisions you make if you build or operate anything, where the difference between a thoughtless and a considered choice is often a factor of ten or a hundred — much larger than anything you achieve by using the tool less.

Model size dominates everything

The energy of an inference scales roughly with the number of parameters activated. A frontier model may use hundreds of times the energy of a small one for a single response.

So the highest-leverage question in any deployment is: is this the smallest model that does the job? Classification, extraction, routing, tagging, simple rewriting and structured output are handled well by models in the one to eight billion parameter range, and a great many production systems use a frontier model for all of it because that is what was easiest to wire up.

A related point: for many narrow tasks, a model is the wrong tool entirely. A regular expression, a lookup table or a small trained classifier from a decade-old technique may be more accurate, effectively free, and deterministic. "Could this be a rule?" is worth asking before every deployment and is asked almost never.

Modality dominates too

Text is cheap. Image generation has been measured at roughly the energy of a phone charge per image in some studies — several hundred times a text query. Video is far more again. If a task can be done in text, doing it in video is not a stylistic choice, it is an energy decision with three orders of magnitude in it.

The other four levers

Cache. A large share of production queries are repeats or near-repeats. Caching answers to identical requests is free efficiency, and semantic caching for near-matches goes further. Many systems recompute the same answer thousands of times a day.

Batch what is not urgent. Grouping requests uses hardware far more efficiently than processing one at a time. If a job can wait an hour, it should.

Schedule for the grid. Carbon intensity varies enormously by hour and region — a factor of five or more across a day is common. Moving flexible workloads to low-carbon hours is called carbon-aware computing, and free data is available from Electricity Maps' public tier and WattTime. This is one of the rare interventions that costs nothing and works.

Choose the region. The same computation in a hydro or nuclear-heavy region and a coal-heavy one differ by roughly an order of magnitude in emissions. Most cloud providers publish per-region carbon information.

Measuring it, with free tools

You cannot manage what you do not measure, and the tools are open source.

• CodeCarbon — a Python package that estimates emissions of a training or inference run from hardware usage and local grid intensity. A few lines to integrate.
• The ML CO2 Impact calculator — a web tool for estimating training emissions from hardware, hours and region. Useful for a paper or a proposal.
• Hugging Face's AI Energy Score — measured energy figures for models on standardised tasks, which is the closest thing to a comparable label across models.
• Cloud provider carbon dashboards — imperfect, methodologically inconsistent between vendors, and better than nothing.

For reporting, the Green Software Foundation's Software Carbon Intensity specification gives a defined way to express emissions per unit of useful work, which is the number that resists gaming better than a raw total.

What to be sceptical of

Annual offset claims. "100% renewable" usually means renewable energy purchased over a year, not power drawn from clean sources at the hour of use. Hourly matching is the stronger claim and far fewer organisations make it.

Per-query figures without a model named. They vary by orders of magnitude between models and tasks, so a single number quoted without saying which model and which task is not informative.

Efficiency framed as reduction. Covered in the previous lesson: per-unit improvement and total growth are compatible.

The proportionate conclusion

For an individual, this is a small part of your footprint and you should not organise guilt around it. For anyone operating a system at volume, the choices above routinely change consumption by a factor of ten or more, and they are usually also the choices that reduce cost and latency — which is the rare case where the responsible option is the one the finance department also wants.