Google Gemini study claims tiny water and energy use per prompt, but experts say the picture is incomplete

TL;DR

• Google’s Gemini study estimates a median text prompt uses about five drops of water, or 0.26 milliliters, and about 0.24 watt-hours of electricity, producing roughly 0.03 grams of CO2 per prompt. The Verge AI coverage.
• The company says it has achieved substantial efficiency gains, including a 33× reduction in electricity per prompt from May 2024 to May 2025 and a 44× drop in the median carbon footprint over the same period. The Verge AI coverage.
• Critics argue the study omits indirect water use (such as water used in cooling data centers) and relies on a market-based carbon metric, which may understate local impacts. They also call for location-based emissions metrics per Greenhouse Gas Protocol standards and peer review.
• Experts warn that efficiency gains can mask greater overall resource use (the Jevons paradox) and that Google’s broader sustainability reporting shows rising ambitions-based emissions despite per-prompt improvements. The Verge AI coverage.
• The study has not yet undergone peer review; Google says it is open to it in the future, and the company notes it factored energy used by idling machines and supporting infrastructure. The Verge AI coverage.

Context and background

The debate over the environmental toll of artificial intelligence has intensified as models grow larger and deploy across services. Data centers powering AI workloads consume significant electricity and require cooling, which in turn uses water in many regions. Reports and academic work have highlighted both direct water use for cooling and broader electricity-driven water footprints tied to power generation. The Verge covered Google’s Gemini study as part of this ongoing conversation, highlighting both claimed efficiency gains and several notable omissions cited by external researchers and analysts. The Verge AI coverage. Researchers such as Shaolei Ren and Alex de Vries-Gao have argued that evaluating environmental impact requires accounting for indirect water use from data-center cooling and using location-based carbon emissions in addition to market-based totals. They also note that comparing a median prompt to other datasets can complicate apples-to-oranges interpretations if underlying methodologies differ. The Verge article discusses these points in detail and frames Google’s disclosures within the broader context of standardized environmental accounting in tech infrastructure. The Verge AI coverage. Google published its paper and accompanying blogs to promote transparency around Gemini’s water, energy, and carbon metrics and to outline efficiency improvements that Google attributes to internal optimizations. The company asserts that its approach goes beyond prior studies by including energy used by idling hardware and supporting data-center infrastructure. The Verge AI coverage.

What’s new

Google’s latest disclosures center on a few quantified claims for a “median” Gemini text prompt:

• Water: about 5 drops of water, i.e., 0.26 ml.
• Electricity: about the same energy as watching TV for less than nine seconds, roughly 0.24 Wh.
• Carbon: about 0.03 g of CO2 emitted per prompt.
• The company asserts these estimates are lower than earlier research for data-center footprints and notes efficiency gains between May 2024 and May 2025, including a 33× drop in energy per prompt and a 44× drop in the median carbon footprint. The Verge summarizes that Google claims these improvements reflect efficiency gains across Gemini’s operation, including energy used by idling machines and other infrastructure. However, the article also highlights ongoing questions from independent experts about the completeness of the accounting, particularly around water use and emissions accounting methods. The Verge AI coverage. A key point in the discourse is the distinction between market-based and location-based carbon accounting. Google’s paper emphasizes a market-based measure, which factors in corporate commitments to support renewables. Critics argue that location-based accounting—considering the local energy mix where a data center operates—can yield higher, and perhaps more policy-relevant, emission figures. The Verge outlines these considerations and cites Greenhouse Gas Protocol standards as a reference for more holistic accounting. The Verge AI coverage. The report notes that Google did not share certain numerical details, such as word counts or token counts per prompt, which some researchers say would aid in independent replication or benchmarking. Google said it would consider peer review in the future. The Verge AI coverage.

Why it matters (impact for developers/enterprises)

For developers and enterprises deploying Gemini or similar AI services, the reported metrics touch on several practical concerns:

• Metric transparency: A single number per prompt (water, electricity, carbon) can be informative but may obscure the full footprints if indirect water use or local energy mix isn’t included.
• Data-center planning: If projections rely primarily on direct water use and market-based emissions, operators may underestimate cooling-water needs or local grid impacts in drought-prone regions and in areas with constrained water resources.
• Policy and procurement: Enterprises seeking to align with sustainability goals may need to advocate for location-based metrics and standardized reporting to compare services consistently across providers.
• Jevons paradox risk: The article notes that efficiency improvements can encourage higher usage or larger-scale deployments, potentially increasing overall resource use despite per-prompt gains. This nuance is important for long-term planning and policy. These considerations matter for organizations building or procuring AI capabilities, and they reflect broader industry debates about how best to quantify and communicate environmental performance. The Verge coverage highlights that Google is positioning its disclosures as a step toward more standardized measurement, while critics urge fuller accounting and peer review. The Verge AI coverage.

Technical details or Implementation

The core figures Google publicized center on a median prompt basis. Specifics include:

• Water per prompt: 0.26 ml (about five drops).
• Electricity per prompt: 0.24 Wh, roughly the energy equivalent of a short television viewing session.
• CO2 per prompt: about 0.03 g.
• Timeframe for gains: May 2024 to May 2025, with a 33× reduction in electricity per prompt and a 44× reduction in the median carbon footprint. Google states that its estimates include energ y used by idling machines and other data-center infrastructure, which it argues contributes to a more complete picture than studies that only model active compute. The article notes that the study presents a “market-based” carbon measure, which accounts for corporate commitments to renewable energy, rather than the fully localized emissions profile that would arise from the local electricity mix. The Verge AI coverage. Research peers, however, have pointed out several limitations:
• Indirect water use: The study excludes a portion of a data center’s water footprint, notably water used for cooling systems in supporting infrastructure. This omission is viewed as a significant gap by researchers who study water usage in AI data centers. The Verge AI coverage.
• Comparison basis: Google emphasizes a median prompt to avoid skewing results by outliers, but researchers argue that the absence of token counts or prompt lengths limits replication and cross-comparison. Some describe the comparison as apples-to-oranges with prior work that used averages. The Verge AI coverage.
• Peer review status: The paper has not yet been submitted to peer review, though Google signaled openness to this step in the future. The Verge AI coverage. The Verge also notes that some researchers question the broader sustainability context in Google’s reporting, including the company’s own sustainability metrics that show rising “ambitions-based” emissions in recent years despite per-prompt improvements. This framing suggests a need to contextualize per-prompt efficiency with overall demand growth and infrastructure expansion. The Verge AI coverage.

Key takeaways

• The Gemini study provides median-per-prompt figures for water, energy, and CO2, highlighting relatively low per-prompt footprints in isolation.
• The methodology emphasizes a median prompt and includes idling energy, aiming for a more comprehensive accounting than some earlier work.
• Critics stress that indirect water use and location-based emissions are not included, which could understate real-world impacts in certain regions or under different energy mixes.
• The discussion illustrates tensions between market-based and location-based carbon accounting and underscores the importance of standardized, peer-reviewed methods for comparability.
• Even with efficiency gains, broader demand and infrastructure growth can influence total resource use, a phenomenon some researchers label as Jevons paradox.

References

• The Verge AI coverage