The Environmental Cost of AI: Google’s Gemini Efficiency Claims Face Intense Scrutiny

Asking Gemini to rewrite your email costs about five drops of water, according to recent disclosures by Google. In an August 2025 technical paper, the tech giant attempted to quantify the environmental footprint of its generative AI operations, estimating that a median text prompt processed by Gemini Apps consumes approximately 0.24 Wh of electricity, emits 0.03 grams of carbon dioxide equivalent (CO2e), and utilizes 0.26 mL of water. While these figures are presented as evidence of significant progress in AI efficiency, the methodology behind these numbers has ignited a firestorm of debate among environmental scientists and climate researchers, who argue that the data provides an incomplete picture of the industry’s true ecological toll.
The core of Google’s argument rests on a year-over-year improvement analysis. The company claims that, over a 12-month period, the energy and carbon footprint of a median prompt have plummeted by 33x and 44x, respectively. To put this in perspective, Google likened the energy consumption of a single AI prompt to the power required to watch television for less than nine seconds. However, these metrics are viewed by critics not as a definitive solution to AI’s sustainability problem, but as a narrow snapshot that omits the most significant portions of the resource-intensive supply chain.
Understanding the Methodology: What is Included?
To its credit, Google’s latest report is more granular than many previous industry disclosures. The company’s calculation accounts for the full operational stack of a prompt: the high-performance AI accelerators (TPUs and GPUs), host CPUs, system memory, and the "idle" power consumption of machines kept on standby. Crucially, it also incorporates data center overhead, such as the electricity required for cooling and power distribution.
According to Google’s internal breakdown, AI accelerators are the primary drivers of energy usage, accounting for 58% of the power consumed by a single prompt. Host CPUs and memory modules contribute another 24%, while the remaining 18% is attributed to standby capacity and facility-wide cooling infrastructure. When stripping away the broader data center overhead to look only at the energy used by active chips, the figures drop to 0.10 Wh and 0.12 mL—a figure Google openly acknowledges is an underestimate of the true operational cost.
The Water Consumption Controversy
The most contentious aspect of the report is the calculation of water usage. Google’s estimate of 0.26 mL per prompt focuses exclusively on direct water consumption—the water used on-site for evaporative cooling systems within its data centers. This methodology, while transparent in its scope, is fundamentally at odds with how environmental experts define "water footprint."

Shaolei Ren, an associate professor at the University of California, Riverside, and a leading expert on the environmental impact of AI, has been a vocal critic of the industry’s reporting standards. Ren argues that by focusing solely on on-site cooling, companies like Google are "hiding the critical information." The missing piece of the puzzle is "indirect" water usage: the water consumed by power plants to generate the electricity that fuels the data centers.
Power plants require vast amounts of water for cooling and for generating the steam that drives their turbines. When this off-site consumption is factored in, the environmental cost of AI is significantly higher than what a direct-usage metric suggests. A 2024 report by the Lawrence Berkeley National Laboratory found that the water required to generate electricity for U.S. data centers is approximately 12 times greater than the water used for direct cooling. Meta, which maintains more rigorous disclosure standards than many of its peers, has reported that its indirect water consumption is more than 20 times its direct usage.
A Chronology of AI Sustainability Concerns
The conversation regarding AI’s environmental impact has evolved rapidly since the widespread adoption of Large Language Models (LLMs).
- 2023: Early research, including a landmark study by Ren and his colleagues, estimated that a conversation with an early version of ChatGPT (GPT-3) could consume a 500 mL bottle of water for every 10 to 50 queries, depending on the server load and location.
- 2024: The Lawrence Berkeley National Laboratory published comprehensive findings on the energy-water nexus in data centers, highlighting the massive disparity between direct and indirect water use.
- August 2025: Google releases its technical paper, claiming significant efficiency gains in Gemini Apps and citing a five-drop-per-prompt figure.
- Late 2025 – Present: Independent researchers and industry analysts begin challenging the validity of Google’s comparison, arguing that the metrics rely on market-based accounting rather than location-based realities.
The Discrepancy in Accounting: Market-Based vs. Location-Based
A major point of contention between Google and independent researchers is the use of "market-based" emissions accounting. Google’s reporting frequently utilizes this method, which allows companies to credit their environmental footprint against the renewable energy credits they purchase.
Critics like Alex de Vries-Gao, founder of the research firm Digiconomist, argue that this approach obscures the actual environmental burden of AI. They contend that companies should report "location-based" emissions, which measure the carbon intensity of the power grid where the data center is physically located. By using market-based figures, a company might claim a low-carbon operation while its physical servers are still drawing power from a grid heavily reliant on coal or natural gas.
By ignoring the local impact of energy generation and focusing on corporate carbon offsets, critics argue that the tech industry is effectively "greenwashing" the intense physical demands of its infrastructure.

The Scaling Problem: Why Accuracy Matters
The urgency of this debate is compounded by the sheer scale of modern AI deployment. In August 2026, the Gemini app surpassed 1 billion monthly active users. When individual efficiencies are multiplied by billions of daily interactions, the cumulative impact becomes staggering.
Google’s latest environmental report underscores this tension. Despite the efficiency gains per prompt, the company’s total water consumption rose by 34% in 2025, reaching 10.9 billion gallons. This creates a paradox: while the per-task efficiency is improving, the total environmental cost is rising as the volume of AI tasks grows exponentially. This phenomenon is known in economics as Jevons Paradox, where increased efficiency leads to higher overall resource consumption because the product becomes cheaper and more accessible, thus increasing demand.
Implications for the Future of AI Infrastructure
The dispute over "five drops of water" serves as a microcosm of a larger challenge facing the tech industry. As AI models become more complex and demand for compute power grows, the physical limits of the planet—specifically water scarcity and electrical grid capacity—are becoming hard constraints.
For policymakers and regulators, the takeaway is clear: voluntary, self-reported metrics from tech giants are insufficient for a comprehensive understanding of AI’s environmental impact. There is an increasing call for standardized, third-party verified reporting that mandates the disclosure of both direct and indirect water and energy usage.
Furthermore, the industry faces a difficult transition as it attempts to reconcile its growth with the reality of climate change. If the current trajectory continues, the energy and water requirements of the "AI revolution" could place unsustainable pressure on local municipal resources, particularly in regions prone to drought.
Ultimately, Google’s technical paper highlights a valid point: artificial intelligence is becoming more efficient at an impressive rate. However, by ignoring the broader systemic costs—the water used to generate electricity, the location-specific impact on grids, and the compounding effect of massive user growth—the narrative of "efficiency" risks becoming a dangerous oversimplification. The five drops of water may be a technically accurate figure for a single prompt, but in the context of a global, power-hungry infrastructure, it is a number that obscures as much as it reveals. As the industry matures, the debate will likely shift from how efficient an AI can be to whether the infrastructure required to sustain it can remain compatible with long-term environmental sustainability.







