Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 5 min read

Fact-check AI energy claims before they steer your architecture

You will hear AI called an environmental disaster. Many of the numbers behind that claim do not survive a calculator, a units conversion, or a question about what you are actually comparing. Before you throttle features,

You will hear AI called an environmental disaster. Many of the numbers behind that claim do not survive a calculator, a units conversion, or a question about what you are actually comparing. Before you throttle features, pick a β€œgreen” model path, or repeat a stat in a design doc, I'd run the claim through four checks drawn from the fact-checking work of Andy Masley and the infrastructure framing of Jeetu Patel. For more on power and data centers, the AI energy and data centers page collects the full conversations.

Does the number pass a units and scale sanity check?

Andy Masley, who taught physics for seven years and fact-checks AI environmental claims, describes the largest error he has found personally as a water figure in Empire of AI that is off by a factor of about 4,500. One root cause: a source mixed up liters and cubic meters. A cubic meter is a thousand liters. The published figure implied that each resident of an 88,000-person community outside Santiago was using about a tenth of a liter of water a day, for everything. Masley argues that is a number anyone could have sanity-checked, and for six months nobody did.

The number that she gave for this community basically implied that the average person there was consuming about point one liters a day, which is something like a fifth of a single bottle of water

Andy Masley, Effective Altruism DC, on Chain of Thought ep 45

The same chapter, Masley says, claimed AI would consume half as much water as all of Britain by 2027, while the cited study puts the figure at roughly 5%. An error that size changes the policy conversation.

When you see megawatt-hours, liters, or β€œequivalent to X cities,” stop at the noun: consumption versus withdrawal, municipal versus regional basin, inference versus training amortized over model lifetime. If the implied per-person daily use would kill people or violate physics, look for the error in the unit or the denominator.

A sketch of a units-trap check (mine, not from any guest):

# Sketch: reject claims whose implied per-capita rate is absurd
claimed_liters_per_person_per_day = reported_total_L / population / days
if claimed_liters_per_person_per_day < 1.0:
    flag("Likely unit mix-up (e.g. mΒ³ vs L) or wrong denominator")

Am I looking at one prompt or the whole grid?

These are different questions. Masley's party-level comparisons treat a single prompt as tiny relative to a normal day: about what printing a fifth of a page emits, running a space heater for half a second, or driving four feet. On water, his estimate is roughly one eight-hundred-thousandth of daily use. Anything you do on a computer carries a small climate cost relative to almost everything else in a day, which is why β€œI ran a thousand searches today” was never a confession.

Sending a prompt probably emits as much as printing a fifth of a single page of a book, or it's like running a space heater for about half a second, or driving your car for about four feet or something like that.

Andy Masley, Effective Altruism DC, on Chain of Thought ep 45

Aggregate impact is different. The bite is how much power the data-center buildout draws and where that power comes from. Utilities and siting decisions move that far more than how often any one person opens a chatbot. Don't design guilt into user flows based on per-prompt estimates while ignoring where the power for new capacity comes from.

What is the comparison baseline, and who moved the goalposts?

Bad comparisons usually smuggle in a baseline that sounds alarming but measures the wrong thing. Masley notes that ~90% of water in one cited framing is taken for power plants off-site and returned to the source, while β€œconsumption” in water-management terms is narrower. Quoting the big headline without that distinction mis-trains readers.

On the Santiago data center, Masley's corrected estimate puts the facility at about 25% of the water in one section of Santiago, but about 3% of the municipal supply. Activist pressure built on the bad figure, he hedges, may have pushed Google toward air cooling instead of water cooling for a proposed center that has not been built. Air cooling takes considerably more energy, so he suspects the redesign would mean higher emissions in exchange for saving water. As of September 2026, Google says it has not decided whether to go ahead with the redesigned, air-cooled project. Environmental policy is trade-offs. You cannot price a trade-off with a number that is wrong by three orders of magnitude.

When someone compares AI to β€œall of Britain” or β€œa thousand towns,” ask: same year, same definition of consume, same geographic boundary, same inclusion of training and embodied hardware? If the answer is fuzzy, treat the ratio as unsupported.

Why is the grid the constraint that actually binds?

Jeetu Patel names infrastructure as the first hard constraint on AI: not enough power, compute, network bandwidth, or data-center capacity to meet demand. His framing is that infrastructure is the input and intelligence is the output, and the goal is generating tokens at the lowest kilowatt and the lowest dollar. That problem is expensive, contested, and unresolved. It is also a different problem from whether you should feel bad for sending a prompt.

There is just not enough power, compute, network bandwidth, and data center capacity to go satiate the needs of AI. And the constraint for infrastructure is directly correlated to a company and a country's ability to to have intelligence. Infrastructure is the input. Intelligence is the output.

Jeetu Patel, Cisco, on Chain of Thought ep 44

For developers, that means the decisions that matter for energy are often upstream of your API call: where capacity is built, what fuel mix serves it, whether cooling choices shift burden from water to electricity, and whether networking losses force you to restart million-dollar training runs. Patel also describes data centers being built where power availability is highest, and scale-across networking when one site cannot host enough GPUs. Your latency and cost optimizations are real, but they sit inside a grid and siting story.

Checklist

  • Convert units and divide by population and time; if implied daily use per person is absurd, fix the math before the moral.
  • Label whether the claim is per query, per user-day, or whole-industry buildout; do not use one to justify policy for the other.
  • Match the baseline (consumption vs withdrawal, local vs municipal vs national, inference vs training amortization).
  • Ask what fuel and cooling trade-off the grid faces; Patel's constraint is power and capacity.
  • If you cannot verify all four, do not put the number in an architecture decision. Say β€œunknown” and point to primary sources.

The longer explainer, with the episode clips, is on Chain of Thought. It draws on this episode.

Subscribe to the Chain of Thought newsletter for new episodes and write-ups like this one.

Drafted with AI assistance from the episode transcripts.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.