How much has Google cut the cost of its AI?
Google reported that the energy cost of a median Gemini text prompt fell 33x in 12 months (May 2024 to May 2025), and its total carbon footprint fell 44x. The median prompt now uses 0.24 watt-hours, 0.03 gCO2e, and 0.26 mL of water. This is Google's own self-reported figure, not independently audited.
Why — the first-principles explanation
The headline number people are looking for is Google's August 2025 technical report, and it's important to be precise about what it measures. It is not the price of the Gemini API, and it is not Google's total AI spending. It's the resource cost of serving one median text prompt, and the reduction is 33x for energy and 44x for total carbon footprint over twelve months.
Why would a per-query cost collapse by that much in one year? Because the cost of serving an AI model is mostly the cost of the arithmetic, and almost every lever attacks the arithmetic. Google's own breakdown attributes roughly 23x to model improvements — smaller and better-distilled models, mixture-of-experts routing that only activates a fraction of the network per token, and quantization that does the math in lower precision. Another 1.4x came from better machine utilization, meaning fewer idle accelerators per unit of work. Then the carbon number gets an extra boost that has nothing to do with efficiency: cleaner electricity and smarter workload placement, worth about 1.4x on emissions intensity.
The multiplication matters. 33x on energy times cleaner power gets you to 44x on carbon. These are compounding, not additive, and that's the whole story of AI cost curves — several independent multipliers stacking on top of each other. It's why per-token API prices have fallen so fast across the industry while total AI spending has exploded. Cheaper per unit, vastly more units.
Now the caveats, because this is a contested area. Google measured Google. There is no independent audit. The report uses a comprehensive boundary — it includes idle machines, CPU and RAM, cooling overhead, and embodied emissions from manufacturing — which is more honest than many published estimates that count only active chip power. But the choice of the median text prompt as the unit does a lot of work: it excludes image and video generation, and long reasoning chains, which are far more expensive. Also, the water figure of 0.26 mL is on-site consumption; it doesn't include the water used to generate the electricity off-site, which is a real and much larger number that different studies handle differently. That boundary choice is the single biggest reason published estimates of AI's footprint disagree with each other by orders of magnitude.
So: 33x is a real, specific, usefully-documented claim about one company's per-prompt energy — and it's also a number chosen and published by the company it flatters.
An example that makes it click
Imagine a pizza shop that used to fire up a whole industrial oven for every single order, even a slice. Then they do three things: they build a smaller oven for small orders, they stop letting the ovens idle hot between orders, and they batch orders that are going to the same neighborhood.
The energy per slice drops enormously — not because they invented a new kind of fire, but because they stopped wasting the fire they had. Then they switch to a power company running on wind, and the carbon per slice drops even further than the energy did. That's the 33x versus 44x. And notice what the number doesn't say: the shop is now selling ten thousand times more pizza, so the total electric bill went up.
Key facts
- Google reported a 33x reduction in energy consumption per median Gemini Apps text prompt over 12 months, from May 2024 to May 2025.
- Google reported a 44x reduction in total carbon footprint per median text prompt over the same period.
- The median Gemini Apps text prompt consumes 0.24 watt-hours of electricity, emits 0.03 gCO2e, and consumes 0.26 mL of water (about five drops).
- Google attributes about 23x of the energy reduction to model improvements and about 1.4x to improved machine utilization.
- An additional ~1.4x reduction in the carbon intensity of Google's data center electricity came from workload location and clean energy procurement.
- The 0.26 mL water figure is on-site consumption at the data center; it does not include off-site water used in electricity generation — a boundary choice that explains much of the disagreement between published AI water estimates.
▶ The 60-second explainer (script)
Google says it cut the energy cost of an AI prompt by thirty-three times in one year. Here's what that actually means. In August 2025, Google published a technical report measuring one median Gemini text prompt. Result: zero point two four watt-hours of electricity, zero point zero three grams of CO2 equivalent, and zero point two six milliliters of water — about five drops. And compared to twelve months earlier, energy per prompt was down thirty-three times, carbon down forty-four times. Why so fast? Because serving AI is mostly the cost of arithmetic, and every lever attacks the arithmetic. Google credits about twenty-three times to model improvements — smaller distilled models, routing that only wakes part of the network, lower-precision math. Another one point four times from not letting machines sit idle. Then cleaner electricity adds another one point four times on carbon, which is why carbon fell further than energy. It's like a pizza shop that stops firing a whole industrial oven for one slice, stops idling the ovens, then switches to wind power. Now the honest part. Google measured Google — no independent audit. The number is for a median *text* prompt, which excludes image and video generation and long reasoning chains. And that water figure is on-site only; it excludes the water used to make the electricity. That boundary choice is why published estimates of AI's footprint differ wildly. Real number. Also a number Google chose.
What authoritative sources say
People also ask
Is 33x about money or energy?
Energy. It's the electricity used to serve one median Gemini text prompt. It is not the Gemini API price and not Google's AI budget, both of which move for different reasons.
Does this mean Google's AI uses less electricity overall?
No, and Google doesn't claim that. Per-prompt cost falling while usage explodes can still mean total consumption rises. Efficiency per unit and total footprint are separate questions.
Is the 0.24 Wh figure trustworthy?
It's the most detailed disclosure any major lab has made, and it uses a broad boundary including idle machines and cooling overhead. But it's self-reported and unaudited, and it covers median text prompts only.
Why do other sources say AI uses way more water?
Boundary choices. Google's 0.26 mL counts on-site data center water. Estimates that also count water consumed generating the off-site electricity land much higher. Neither is lying; they're measuring different things.
Do image and video prompts cost the same?
No — they cost substantially more, and Google's headline figure excludes them. The report's unit is the median text prompt.