Google’s “Five‑Drop” Claim for a Gemini Query Is Technically Correct, but the Full Water Story Is Larger
Google says a median Gemini Apps text prompt consumes 0.24 Wh of electricity, emits 0.03 g CO₂‑e and uses 0.26 mL of water – roughly the volume of five drops. The figures come from an August 2025 technical paper and are presented alongside a claim that the energy and carbon impact per prompt have improved by 33‑fold and 44‑fold respectively over the previous year.
Key Takeaways
- Google’s median Gemini prompt uses 0.24 Wh, 0.03 g CO₂‑e and 0.26 mL of water – about five drops.
- The water figure covers only on‑site cooling; off‑site power‑plant water use is omitted.
- Researchers argue location‑based emissions and full water accounting would raise the footprint substantially.
- Gemini reached 1 billion monthly active users in August 2026, magnifying any per‑query impact.
- Google’s total water consumption rose 34 % to 10.9 billion gallons in 2025.
| Metric | Value | Why it matters |
|---|---|---|
| Energy per median Gemini prompt | 0.24 Wh | Sets a baseline for per‑query power consumption. |
| Water per median Gemini prompt | 0.26 mL | Basis for the “five drops” claim. |
| Gemini monthly active users | 1 billion (August 2026) | Scale amplifies cumulative resource use. |
| Google total water consumption 2025 | 10.9 billion gallons | Indicates overall corporate water demand. |
What the “Five‑Drop” Figure Represents
The August 2025 paper defines a “median Gemini Apps text prompt” as the typical request sent to the model. According to Google, such a prompt draws 0.24 Wh of electricity, releases 0.03 g CO₂‑e and requires 0.26 mL of water for on‑site data‑center cooling. Google likens the energy use to watching television for less than nine seconds. Over a twelve‑month window, the company reports that the same prompt’s energy and carbon footprints have shrunk by factors of 33 and 44 respectively, reflecting hardware and software efficiency gains.
How Google Calculates the Footprint
Google’s methodology is broader than many industry estimates. It adds the power draw of AI accelerators (TPUs or GPUs), host CPUs, memory, idle machines kept on standby, and data‑center overhead such as cooling systems. In this accounting, AI accelerators contribute 58 % of the energy, host CPUs and memory 24 %, and spare capacity plus overhead the remaining 18 %.
If the calculation were limited to active TPUs/GPUs alone, the energy would drop to 0.10 Wh and water use to 0.12 mL per prompt – numbers Google labels an “underestimate.” The company therefore argues its published figure is a conservative, inclusive estimate.
The Water Accounting Gap
Google’s 0.26 mL water figure accounts only for the water used on‑site to cool its data‑center fleet, based on a 2024 average. The paper notes that this data has not been verified by an independent third party. Power plants that generate the electricity also consume water for cooling and steam production, a component omitted from Google’s calculation.
“They're just hiding the critical information,” Shaolei Ren told an interviewer.
UC Riverside associate professor Shaolei Ren, whose research the paper cites, emphasizes that off‑site water use is a substantial portion of the total footprint. Ren’s own 2023 study estimated that GPT‑3 required a 500 mL bottle of water for every 10 to 50 medium‑length responses, a figure that already includes power‑plant water consumption.
A 2024 Lawrence Berkeley National Laboratory report supports Ren’s point, finding that water used to generate electricity for U.S. data centers in 2023 was roughly twelve times the amount used for direct cooling. Meta, the only hyperscaler that publicly discloses indirect water use, reports a ratio exceeding 20 : 1.
Scale Amplifies the Impact
Gemini’s user base crossed the 1 billion monthly active users threshold in August 2026. Even a modest per‑query water figure, when multiplied by billions of requests, can translate into a sizable aggregate demand. Google’s own environmental report shows total water consumption rose 34 % to 10.9 billion gallons in 2025, indicating that overall corporate water use is climbing despite per‑prompt efficiency gains.
Ren and Digiconomist founder Alex de Vries‑Gao argue that Google should report location‑based emissions, which reflect the actual grid mix of each data‑center region, rather than market‑based figures that credit the company’s clean‑energy purchases. Such granularity would likely raise the reported carbon intensity of Gemini queries.
Why It Matters
Understanding the full resource profile of AI queries is essential for policymakers, enterprises, and sustainability advocates. The “five‑drop” headline captures a technically accurate slice of the story, but omitting off‑site water use and location‑specific emissions can understate the true environmental cost. As AI workloads scale to billions of users, transparent accounting becomes a prerequisite for meaningful mitigation strategies.
Frequently asked questions
What does the “five drops of water” claim actually measure?
It refers to the 0.26 mL of water used for on‑site cooling of Google’s data centers for a median Gemini Apps text prompt, as reported in Google’s August 2025 technical paper.
How does Google’s methodology differ from other AI footprint estimates?
Google includes AI accelerators, host CPUs, memory, idle standby machines, and data‑center overhead in its calculations, whereas many estimates focus only on the active GPUs/TPUs.
Why do researchers criticize Google’s water accounting?
Critics point out that Google’s figure excludes water used by power plants to generate electricity, which can be an order of magnitude larger than on‑site cooling water.
How does Gemini’s user scale affect its overall environmental impact?
With 1 billion monthly active users (as of August 2026), even a small per‑query footprint aggregates to a substantial total water and energy demand, especially as Google’s total water consumption rose to 10.9 billion gallons in 2025.
What reporting changes do experts recommend?
Experts suggest using location‑based emissions data and accounting for both on‑site and off‑site water use to provide a more complete picture of AI’s environmental impact.
Bottom Line
Google’s claim that a typical Gemini query consumes “five drops of water” is accurate within the narrow scope of on‑site cooling water, but it omits the far larger water footprint associated with electricity generation and regional grid mixes. As Gemini’s usage scales to billions of users, transparent, comprehensive accounting will be crucial for assessing and reducing AI’s true environmental cost.
Future disclosures that integrate full‑life‑cycle water use and location‑specific emissions will enable stakeholders to gauge progress beyond headline‑grabbing per‑query numbers.
Reporting based on: original source.
Comments
Post a Comment