Ask around about what one AI query costs and you'll get a dozen confident answers: "a bottle of water," "a phone charge," "a minute of a hair dryer," "basically nothing." Someone, somewhere, did the math behind each of these numbers, and each is true for its own narrow case. None of them works as a universal answer. The honest answer is less convenient: almost nobody actually knows the precise figure, and the companies that could tell you mostly don't.

This isn't a conspiracy. It's just how the market is built. A model's energy footprint is as much a trade secret as its architecture, because a competitor could use it to back out the cost per query and the margin behind it. So nearly every number you've seen in a headline is a third-party reconstruction, pieced together from chip spec sheets, data-center efficiency ratios (PUE), and guesses about how long an average conversation runs. "Roughly" and "on the order of" in these estimates aren't polite hedging — they're honest admissions of the error bars involved.

Why the Numbers Disagree by an Order of Magnitude

Different estimates often measure entirely different things. Some count only the electricity the chip draws during the computation itself. Others add in the data center's overhead — cooling, wiring losses, backup systems — which can easily double the total. Some fairly spread the cost of training the model across every query it will ever serve; training a frontier model burns energy on the scale of what several thousand homes use in a year, but that cost gets divided across billions of future queries and ends up as a sliver per request. And some estimates are still quoting measurements from three-year-old models that have since gone through several generations of optimization.

One of the rare recent data points came from Google itself: in the summer of 2025 the company published its own methodology and an estimate of roughly 0.24 watt-hours and about a quarter of a milliliter of water per text prompt to its Gemini model. That's noticeably lower than the viral "a glass of water per question" figures that circulated a year earlier, based on older, less efficient models. Independent researchers — including AI-energy specialists like Sasha Luccioni at Hugging Face — have published estimates in a similar range for comparable models, while cautioning that a number self-reported by a company with every incentive to look frugal deserves independent auditing before it's taken at face value.

What's Reasonably Well Established

The order of magnitude looks like this: a simple text query to a current, efficient model costs somewhere in the single digits to low tens of watt-hours. Roughly the range between a few seconds of a microwave running and a couple of minutes of an ordinary lightbulb. Real, but not the environmental catastrophe some viral posts imply — nor the "practically free" story the industry prefers. The water used to cool data centers is a separate and equally real story, one we've covered on its own: how AI companies are buying up entire power plants to feed their clusters. What matters here is a different axis — not where the power comes from, but how much of it goes into the specific action you take dozens of times a day without thinking about it.

Not All Queries Are Equal

This is where the averaged number lies hardest. A text answer to a simple question and an image generation aren't in the same order of magnitude: producing a single image can take tens of times more compute than a text exchange of comparable length. Video generation adds another order or two on top — a few seconds of synthesized footage can burn more energy than hundreds of text conversations combined. And "reasoning" models — the ones that think through a problem internally before producing a short answer — silently generate thousands of extra tokens the user never sees. The smarter a model looks from the outside, the more it may be burning on the inside, and interfaces almost never show you that number.

There's a flip side companies talk about even less: routing. A simple question increasingly gets handed not to a company's flagship model but to a small, cheap one that handles it faster and with less draw; the heavy model only gets escalated the queries the small one can't handle. This is sound engineering — but it also means your personal footprint depends on an invisible system decision you don't control and will never see.

The Efficiency Trap

Here's the most uncomfortable fact in the whole story. Each new generation of models really is more efficient per unit of computation than the last — that's not a myth, it's genuine engineering progress. And yet total energy consumption across the AI industry isn't falling. It's climbing, sharply. By the International Energy Agency's estimate, data centers already account for something on the order of one and a half to two percent of global electricity use, and that share is projected to roughly double within a few years — driven in large part by AI workloads specifically.

Economists have had a name for this since the 19th century. William Stanley Jevons first described it studying coal: the more efficiently a steam engine burns coal, the cheaper coal-powered work becomes — and the more coal gets burned in total, because cheapness creates demand that didn't exist before. The same thing is happening with AI. A falling cost per query doesn't lower the total bill; it lowers the barrier to entry. Cheap AI gets bolted onto everything — the search box, the email autocomplete, the online store's chat widget — and the aggregate footprint grows faster than the per-unit price falls.

Our record: Thoth is the god of writing and precise accounting, the one who weighs and records without distortion. The Shadow Thoth isn't a direct lie — it's control over what gets counted out loud and what stays off the ledger. When a company honestly tells you the energy cost of one of its queries but stays quiet about the sum of a billion such queries a day, and how fast that sum is growing, there's no deception in the narrow sense. But there's no complete truth either. It's the missing part that decides what you end up believing about the technology.

What This Means for You

Knowing the real cost of one query is neither a reason to panic nor a reason to shrug. One question to a chatbot isn't drowning the planet; but the "weightless cloud" story isn't an honest account either — it's a comfortable anesthetic someone chose on your behalf. There's a third, adult position between those two extremes: count honestly. Don't buy the viral figure claiming a bottle of water per casual question, but don't buy the reassuring PR about "basically nothing" without asking the follow-up question — what's the total, and who's paying for its growth? Demanding open, auditable numbers instead of convenient rounding is what Maat looks like applied to technology: not moralizing, but engineering an honest ledger. Whoever counts correctly is the one who gets to decide what to do with the result.