Since the release of ChatGPT in late 2022, generative AI has been on the rise. In July 2025, OpenAI announced that ChatGPT was receiving 18 billion “prompts” (instructions written by users) per week, from 700 million users—representing 10 percent of the world’s population.
Today, the rush to develop these tools is global: all the major Big Tech players are now developing their own generative AI models, primarily in the United States and China. In Europe, the French company Mistral, which developed the Le Chat assistant, recently set records with a market capitalization of nearly 12 billion euros. Each of these models operates within a specific geopolitical context, sometimes involving different technological choices. But all have a significant environmental footprint that continues to grow exponentially, driven by the proliferation of uses. Some experts, including those at the specialized think tank The Shift Project, are sounding the alarm: this growth is unsustainable.
However, all stakeholders in this field—including consumers—are now well aware of the environmental cost associated with digital usage, though they may not necessarily be aware of the specific figures involved.
Driven by a variety of factors (regulatory requirements, marketing, and sometimes environmental awareness), several major tech companies have recently conducted life cycle assessments (LCA, a methodology for evaluating the overall environmental impact of a product or service) of their models.
In late August 2025, Google released its own report quantifying the impacts of its Gemini model. How valid are these estimates, and can we trust them?
A surprisingly low carbon footprint
For a generative AI model to function, it must first be “trained” using a large number of written examples. To measure the electricity consumed by a “prompt,” Google therefore focused on the usage phase—rather than the training phase—of its Gemini AI. Based on its own calculations, Google reports that a prompt consumes an average of just 0.24 watt-hours (Wh)—which is very low: roughly the amount of energy consumed by a standard 15-watt light bulb in one minute.
How did the authors arrive at this figure, which is significantly lower than in other studies already conducted on this topic, such as the one conducted by Mistral IA in July 2025?
The first reason has to do with what Google actually measures. For example, the report reveals that 58% of the electricity consumed by a prompt is used by specialized AI processors (the graphics processing unit, or GPU, and the Tensor Processing Unit, or TPU), 25% by conventional processors, about 10% by processors in standby mode, and the remaining 7% for server cooling and data storage.
In other words, Google only takes into account the electricity consumed by its own data centers, and not that consumed by users' devices and routers.
Furthermore, no information is provided on the number of users or the number of queries included in the study, which calls its credibility into question. Under these circumstances, it is impossible to know how user behavior might affect the model’s environmental impact.
In 2024, Google purchased the equivalent of Ireland's annual electricity production
The second reason has to do with how the electrical energy consumed is converted intoCO2 equivalents. It depends on the electricity mix in the location where the electricity is consumed, both at the data centers and at the users’ endpoints. As we’ve seen, Google is only concerned with its own data centers.
Google has long been committed to energy optimization, turning to carbon-free or renewable energy sources for its data centers located around the world. According to its latest environmental report, these efforts appear to be paying off, with a 12% reduction in emissions over the course of a year, even as demand rose by 27% during the same period. The needs are enormous: in 2024, Google consumed 32 terawatt-hours (TWh) of electricity for its computing infrastructure—equivalent to Ireland’s annual electricity production.
In fact, the company signed 60 exclusive long-term electricity supply contracts in 2024, bringing the total to 170 since 2010. Given the scale of Google’s operations, having exclusive long-term electricity contracts undermines decarbonization efforts in other sectors. For example, the low-emission electricity that powers data centers could be used for heating, a sector that still relies heavily on fossil fuels.
In some cases, these contracts involve the construction of new energy-generation infrastructure. However, even for carbon-free renewable energy production, their environmental footprint is not entirely neutral: for example, the impact associated with the manufacture of photovoltaic panels ranges from 14 gCO2eq to 73 gCO2eq/kWh, which Google does not factor into its calculations.
Finally, many Google services rely on server “colocation” in data centers that are not necessarily carbon-neutral, a factor that is also not taken into account in the study.
In other words, the methodological choices made for the study helped to minimize the magnitude of the figures.
Five drops of water per person, but 12,000 Olympic-sized swimming pools in total
Freshwater consumption is increasingly being addressed in environmental reports related to the digital sector. And for good reason: it is a precious resource and a key component of a planetary boundary that was recently crossed.
The Google study estimates that its water consumption for Gemini is 0.26 ml—or five drops of water—per prompt. This figure may seem negligible when considered on the scale of a single prompt, but small streams make big rivers: it must be viewed in the context of the explosive growth in AI usage.
Overall, Google consumed approximately 8,100 million gallons (about 30 million cubic meters, the equivalent of roughly 12,000 Olympic-sized swimming pools) in 2024, a 28% increase from 2023.
But here, too, the devil is in the details: Google’s report only accounts for the water used to cool servers (based on a principle very similar to how we cool off when sweat evaporates from our bodies). The report effectively excludes water consumption related to electricity generation and the manufacturing of servers and other computer components—even though these factors are factored into the calculation of its carbon footprint, as noted above. As a result, environmental impact metrics (carbon, water, etc.) do not all have the same scope, which complicates their interpretation.
Studies that are still too opaque
Like most studies on this topic, Google’s was conducted in-house. While we understand the need for trade secrets, such a lack of transparency and independent expertise raises questions about its legitimacy and, above all, its credibility. Nevertheless, it is possible to draw comparisons with other AI systems, for example, based on the data presented by Mistral AI in July 2025 regarding the environmental impacts associated with the life cycle of its Mistral Large 2 model—a first of its kind.
This study was conducted in collaboration with Carbone4, a recognized French leader in life cycle assessment (LCA), with support from the French Environment and Energy Management Agency (ADEME), which lends it credibility. The results are as follows.
Over the model’s total 18-month lifespan, approximately 20,000 metric tonsof CO2 equivalent were emitted, 281,000cubic metersof water were consumed, and 660 kilograms of antimony equivalent were used (an indicator that accounts for the depletion of metallic mineral raw materials).
Mistral points out that using the model (inference) has effects that they consider “marginal, ” when considering an average prompt using 400“tokens” (processing units correlated with the size of the output text): this prompt corresponds to the emission of 1.14 g ofCO2 equivalent, 50 ml of water, and 0.5 mg of antimony equivalent. These figures are higher than those put forward by Google, which, as we have seen, were obtained using a favorable methodology. Furthermore, Google based its study on a “median” prompt without providing further statistical details, which would have been welcome.
In reality, one of the main motivations—whether on the part of Google or Mistral—behind this type of study remains a marketing one: the goal is to reassure thepublicabout the environmental impact of AI (what could be called“greenwashing”) in order to drive consumption. Focusing solely on the impact stemming from user prompts also obscures the bigger picture when it comes to costs (for example, those associated with training the models).
Let’s acknowledge that the principle of conducting impact assessments is a positive one. But the lack of transparency surrounding these studies—even when they do exist—must be questioned. To date, neither Mistral nor Google has disclosed all the details of the methodologies used, as the studies were conducted internally. There needs to be a common framework that clarifies what must be taken into account in a comprehensive life-cycle assessment (LCA) of an AI model. This would allow for a meaningful comparison of results across different models and help limit marketing hype.
One limitation likely stems from the complexity of generative AI. How much of the environmental footprint can be attributed to the use of a smartphone or computer for the prompt? Models that enable fine-tuning to adapt to the user consume more energy?
Most studies on the environmental footprint of generative AI treat these systems as closed systems, which prevents them from addressing the crucial issue of unintended consequences resulting from these new technologies. This obscures the staggering increase in our use of AI, by reducing the problem to environmental cost of a single prompt.
