Billions Of AI Prompts, Billions In Energy Costs: The Planet Pays As Data Centres Heat Up
Google and OpenAI reveal the per-query energy and water costs of AI tools like Gemini and ChatGPT, highlighting growing environmental concerns over large-scale AI adoption.
Published : September 11, 2025 at 5:00 PM IST
Hyderabad: The title for the world's most powerful artificial intelligence model keeps changing hands as tech giants continue to roll out new and advanced models capable of understanding user needs better and performing tasks more efficiently—such is the nature of the AI race. The benefits of the technology can't be overlooked, as it has impacted almost every imaginable sector from healthcare to agriculture, education, governance, finance, and more, helping not only with increased productivity across industries but also playing an irreplaceable role in advancing scientific research. This has also led to a better-than-ever generative AI, allowing users to produce high-quality images and realistic videos with just a text prompt.
While debates around the ethical use and implementation of AI continue, an equally urgent concern is its environmental impact—particularly the rising electricity consumption required to power these systems and the high amount of water to keep them from overheating. Organisations like the United Nations and the International Energy Agency have documented the growing carbon footprint of tech giants, largely driven by the energy demands of data centres supporting AI development.
However, pinpointing the environmental cost of consumer-level generative AI applications remained difficult, leading to only speculations—until recently. We finally have an official estimate of energy and water consumption for an average AI query, thanks to numbers thrown by OpenAI and a detailed account shared by Google.
Environmental cost of ChatGPT and Gemini user queries
Back in June 2025, OpenAI CEO Sam Altman penned a blog revealing the energy consumption of a single ChatGPT query: 0.34 watt-hours of power (about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes) and 0.32176 millilitres of water (roughly one-fifteenth of a teaspoon). He did not specify which model the figures referred to, nor did he provide documentation explaining how the estimates were calculated.
Google, however, recently published a paper detailing how it measures the carbon emissions, energy usage, and water consumption of its AI models. It says that the per-prompt energy impact of Gemini is equivalent to watching TV for less than nine seconds.
According to the methodology, the median Gemini Apps text prompt uses 0.24 watt-hours (Wh) of energy, emits 0.03 grams of carbon dioxide equivalent (gCO2e), and consumes 0.26 millilitres (or about five drops) of water.
Google calls this a comprehensive estimate that accounts for all critical elements of serving AI globally, and is considerably higher than simplified estimates that measure just 0.10 Wh of energy, 0.02 gCO₂e, and 0.12 mL of water. It says that most current estimates only account for active machine usage, leading to overly optimistic figures, but its methodology reveals that the true energy, carbon, and water footprint of AI is significantly higher than simplified models suggest.
Google’s methodology considers not just active computation but the full system dynamics across hardware, data centres, and model operations, which include:
- Full system power usage, including actual chip utilisation
- Idle machine energy, required for reliability and scalability
- CPU and RAM consumption, beyond just TPUs and GPUs
- Data centre overhead, like cooling and power distribution (measured via PUE)
- Water usage, especially for cooling systems
While the simplified estimates for the median Gemini App text prompt are 0.10 Wh, 0.02 gCO₂e, and 0.12 mL water, Google’s comprehensive estimates put them at 0.24 Wh, 0.03 gCO₂e, and 0.26 mL water, accounting for all critical elements of serving AI globally.
AI efficiency is important. Today, Google is sharing a technical paper detailing our comprehensive methodology for measuring the environmental impact of Gemini inference. We estimate that the median Gemini Apps text prompt uses 0.24 watt-hours of energy (equivalent to watching an… pic.twitter.com/86v42LLkrW
— Jeff Dean (@JeffDean) August 21, 2025
In addition to providing Gemini’s energy usage, Google advocates for industry-wide consistency in measuring AI’s resource consumption to better reflect real-world operational efficiency. “By sharing our findings and methodology, we aim to drive industry-wide progress toward more efficient AI. This is essential for responsible AI development,” Google said.
The cause for concern remains
While industry experts and stakeholders praise Google’s move, calling it a step towards transparency, they also highlight that AI demands more resources than traditional computing, and the load caused by billions of user queries would be worlds apart when compared to a single prompt.
“While today’s usage might seem under control, the real challenge is what happens when billions of such queries are made daily. That’s when the impact becomes huge,” Manoj Dhanda, Founder and CEO, Utho Cloud, tells ETV Bharat.
Notably, the data obtained by Axios suggests that OpenAI’s ChatGPT alone sees more than 2.5 billion prompts every day, which is over 912.5 billion requests every year. If we bring in tools like Gemini, Grok, Copilot, Character AI, and others into the mix, the number of daily AI prompts worldwide will climb even higher.
The environmental cost of training these models is another story altogether.OpenAI in 2018 revealed that the amount of “compute” required to train its largest AI models was doubling every 3-4 months. Researchers at the time estimated that training GPT-3 (with 175 billion parameters) consumed 1,287 megawatt hours (MWh) of electricity and generated 552 tons of carbon dioxide, equivalent to the annual emissions of 123 gasoline-powered passenger vehicles. The emission is believed to be escalating since then, according to an Epoch report, as the training requirements for the more recent frontier AI models have increased by 4-5 times annually during the past four years. According to the National Institute of Communication Finance (NICF), data centres are on track to account for 14 per cent of all global emissions by 2040.
While Google showcased a lower emission level for a single text prompt, its overall carbon footprint has increased exponentially, primarily due to its AI efforts. Driven by the energy demands of power-intensive data centres, indirect carbon emissions from the operations of four leading AI-focused tech companies—Google, Microsoft, Meta, and Amazon—have increased by an average of 150 per cent between 2020 and 2023, according to Greening Digital Companies 2025 report, published by the International Telecommunication Union (the UN agency for digital technologies) and the World Benchmark Alliance.
Environmental impact of AI data centres
Data centres are used to train and run the large language models, multimodal models, and deep learning models that power scientific research, enterprise needs, and consumer-centric generative AI tools, allowing users to generate text, images, and videos. While they have been around for years to support the internet, cloud computing services, and other obligations, the AI race has prompted a sudden rise in their construction and operation. For instance, Meta plans to spend hundreds of billions of dollars to build several massive AI data centres for superintelligence, with its first multi-gigawatt data centre, dubbed Prometheus, expected to come online in 2026.
Talking to ETV Bharat, Zigment AI Founder and CEO Dikshant Dave says, “Artificial intelligence has the power to transform industries, but its environmental footprint is also significant and should not be ignored. Training and running AI models do consume significant electricity and water, since data centres need power to run servers and keep systems cool.”
According to IEA, while traditional data centres use between 10 and 25 megawatts (MW) of power, demand by hyperscale AI centres can exceed 100 MW–equivalent to the annual electricity consumption of 100,000 households. The largest data centre announced is set to consume as much electricity as 5 million households.
Abhishek Agarwal, President, Judge India & Global Delivery, The Judge Group, highlights the high environmental cost of training and running AI models in large-scale data centres and says, “This is a cause for concern, particularly when looking at areas that already face energy deficiencies or water scarcity.”
An assessment of 200 digital companies by ITU reveals that they consumed an estimated 581 terawatt-hours (TWh) of electricity in 2023 or 2.1 per cent of the global demand, with AI being one probable driver of this growth. It noted that the rapid expansion of AI is fuelling a sharp increase in global electricity consumption, with data centre usage rising by 12 per cent annually since 2017—four times faster than the overall growth in electricity demand.
According to the International Energy Agency, global data centres consumed 415 TWh of electricity in 2024 or 1.5 per cent of the global demand, with the United States holding the largest share (45 per cent), followed by China (25 per cent) and Europe (15 per cent). The report says that the consumption is projected to more than double to 945 TWh by 2030—reaching between 700 TWh and 1700 TWh by 2035, depending on adoption and efficiency.
Apart from putting pressure on existing energy infrastructure, the carbon emissions of AI data centres are also on a sharp trajectory, with top-emitting AI systems contributing up to 102.6 million tons of CO2 equivalent per year, according to a study published in Springer Nature.
“The environmental effect of training and running AI models is definitely a serious concern. Training bigger models means that thousands of GPUs need to run continuously, which means huge electricity bills. If the source of the energy is non-renewable, then carbon emissions skyrocket,” Manoj Dhandha tells ETV Bharat, adding that the use of conventional data centre cooling relies heavily on water, which adds another strain on natural resources.
Experts advocate for sustainable AI systems
While acknowledging the environmental toll of AI systems, experts echo the sentiment that the future could be greener if we make decisions rooted in responsibility, sustainability, and efficiency.
Manoj Dhanda of Utho Cloud says that AI’s initial financial and environmental costs are typical of new technologies, but innovation drives efficiency. “We can already see this happening with AI advancements, such as improved chip designs that consume less power, smaller yet more capable models that require fewer resources, and smarter cooling systems that reduce energy and water usage,” he says.
Advocating for thoughtful design and deployment, Dhanda calls for fueling data centres with renewable energy, using green cooling, employing transparent reporting, and continuously optimising model efficiency. “That is why I do not view AI as a continuing danger to the environment. Rather, it is a responsibility we need to deal with sensibly, so that AI will eventually become cleaner, more efficient, and sustainable,” he adds.
Dikshant Dave of Zigment AI says that while AI’s benefits are undeniable, ignoring its environmental toll risks long-term damage. “The path forward isn’t to slow down adoption, but to scale responsibly, ensuring that each breakthrough in AI is matched with a step forward in sustainability,” he says. “True innovation solves problems without creating new ones.”
Applauding Google’s transparency report and active investment in efficiency, he adds that hardware and infrastructure are evolving toward greater sustainability. “As models get smarter, we’ll likely see better optimisation, meaning AI will be able to deliver more output per unit of energy or water, continuing the trajectory towards cleaner, more efficient AI at scale,” he further says.
Abhishek Agarwal of Judge India shares a similar sentiment. Mentioning how energy-intensive computing devices have become smaller and more efficient, he says, “History suggests that new techniques often begin with heavy resource usage but become more efficient over time.”
Highlighting the positive impact of AI in climate research, energy management, and agricultural science, he says, “We should not turn our back on AI as it needs energy, but should improve its efficiency and sustainability.”
Nitin Lahoti of Mobisoft Infotech acknowledges AI’s energy and water demands as a concern, but remains optimistic about its future, expecting it to become cleaner and smarter in the long term.
“The industry is already moving toward smaller, more efficient models, and new hardware that uses less power. Data centres are experimenting with better cooling and renewable energy sources,” he says. “With this kind of innovation, the impact per AI query will keep going down.”
He also called for the need to explore innovative ways to repurpose the excess heat released by AI data centres, and now a new study by Rice University offers an “economically compelling” way to convert this wasted heat into usable electricity using a compact thermal-boosted power system with low-cost, rooftop flat-plate solar collectors. They validated its performance in two of America’s biggest data centre hubs and recovered 60-80 per cent more electricity annually from the same waste heat.
Tech giants gear up for efficient and sustainable AI systems
Tech giants are aware of the AI data centre emissions, and they are looking at ways to minimise their increased carbon footprint. While their ambitious targets have yet to be fully translated, giants like Meta, Amazon, Microsoft, and Google all highlight their work to achieve sustainable AI by investing in chip-level liquid cooling design and new carbon-free energy projects, including nuclear and renewable energy.
Last year, Microsoft signed a 20-year power purchase agreement (PPA) for the full 837MW power output of the Three Mile Island nuclear plant in Pennsylvania, USA, which is enough to power around 800,000 American households. Beginning in 2028, this clean energy will be used to support the company’s growing network of AI data centres.
Meta and Microsoft are adopting sustainable construction practices in their data centres by replacing high-emission materials like steel and concrete with engineered wood products. With this switch, Meta estimates a 41 per cent reduction in its carbon emissions, while Microsoft projects up to 65 per cent savings. Both companies are aligning these efforts with ambitious climate goals for 2030, where Meta targets net-zero emissions across its value chain and Microsoft aims to become carbon negative.
In 2023, Google’s data centres saw a 17 per cent rise in electricity use, leading to a 13 per cent increase in emissions. To address this, the company developed more efficient AI infrastructure, which includes its 67 per cent more energy-efficient 6th-gen tensor processing unit (TPU). In 2024, despite a 27 per cent increase in electricity consumption, the tech giant reported a 12 per cent reduction in data centre emissions, courtesy of clean energy procurement and improved hardware efficiency.
In its 2024 Environmental Report, Google said that its centres became 1.8 times more efficient than the industry average, with 64 per cent of their energy coming from carbon-free sources. The company also revealed its plan to run on 24/7 carbon-free energy by 2030.
In its recently released paper, Google detailed its research innovations across software, hardware, and infrastructure, which helped its AI systems become more efficient, minimising environmental impact. Over a recent 12-month period, the energy and total carbon footprint of the median Gemini Apps text prompt had reportedly decreased by 33-fold and 44-fold, respectively. Google is also investing in reducing the power provisioning costs and water required per prompt.
Outlining its efficiency strategy, the tech giant highlighted several initiatives: implementing smarter model architectures, optimising algorithms, leveraging custom hardware for inference, managing resources dynamically, deploying an advanced software stack, operating data centres powered by 24/7 carbon-free energy, aiming to replenish 120 per cent of the water they consume, and applying science-based assessments to guide cooling strategies (details below).
For a greener AI, the ITU urges companies to fully disclose the energy use and emissions from AI training and operations. It calls for the creation of AI-specific reporting standards and efficiency benchmarks, such as energy per computation or emissions per training cycle, to ensure transparency, comparability, and sustainability—similar to what Google has now revealed.
AI efforts in India
Led by public-sector leadership, India has started making strides in the AI space and semiconductor infrastructure. It now hosts the world’s fastest-growing developer base and ranks second in public generative AI projects on GitHub. The government’s flagship IndiaAI Mission, launched with over Rs 10,300 crore in funding, aims to democratise AI under the “AI for All” slogan—enhancing data quality, building indigenous models, supporting startups, and encouraging the use of AI to solve real-world problems in areas like healthcare, farming, energy, education, finance, and more.
A key pillar of IndiaAI is the creation of a massive compute facility with 18,693 GPUs, one of the largest globally—designed to power indigenous AI solutions. With 17,300 GPUs already deployed and pushing national capacity beyond 34,000 GPUs, it supports an open marketplace for compute, cloud, and AI services.
The mission also introduced AIKosha, a dataset platform for India-specific models, and the creation of AI Centres of Excellence for healthcare, agriculture, sustainable cities, and education.
India’s AI breakthroughs include BharatGen, the first government-funded multimodal LLM, and tools like Bhashini for real-time language translation, Sarvam-1 (2B parameters), Hanooman’s Everest 1.0, and Chitrelekha, an open-source video transcreation platform for Indic languages.
Parallelly, the India Semiconductor Mission is building domestic chip-making capacity, with Tata Electronics and PSMC constructing the first fab in Dholera, Gujarat, expected to deliver chips by late 2026. The recent inauguration of CG Semi’s OSAT facility and the unveiling of India’s first homegrown chips at SEMICON India 2025 mark major milestones in the country’s tech ambitions.
Data Centre infrastructure in India
The AI efforts in India have also resulted in the growth of the country’s data centre capacity. India ranks 14th globally in terms of the number of data centres and is the second-fastest growing market in the Asia-Pacific region, according to Avener Capital.
According to a Colliers report, the data centre capacity of the country has grown four times in the last 6-7 years and stands at 1,263 megawatts (MW) as of April 2025, driven by increasing adoption of AI, among other factors like a surge in demand for digital and cloud services. Mumbai hosts the majority of the data centre capacity with 41 per cent share, followed by Chennai (23 per cent) and Delhi NCR (14 per cent).
As per the Colliers report, the capacity is projected to cross 4,500 MW by 2030 with an estimated investment of around $20-25 billion in the next 5-6 years. It says that Mumbai will continue to dominate the overall market, while smaller growth markets like Hyderabad, Bengaluru, and Pune could see multifold growth in inventory levels, with Hyderabad emerging as a major hub.
The Avener Capital also presents similar estimates, suggesting that India’s data centre capacity could reach 5,000 MW by 2030. It values the current data centre market at $10 billion, and expects it to grow at a compound annual growth rate (CAGR) of 25 per cent.
The environment pays the price
The rapid growth of data centres in India poses sustainability challenges, especially as the country’s AI market continues to grow rapidly.
Globally, low-carbon energy sources like renewables and nuclear now supply 40.9 per cent of electricity, yet India still relies on coal for over 70 per cent of its power, with renewables contributing just 25 per cent. This heavy dependence on fossil fuels results in a high carbon intensity—averaging around 820 grams of CO₂ per kilowatt-hour (kWh), as per IIT Varanasi. This is nearly double the G20 average of 445 grams of CO₂ per kWh.
In 2023, India’s power sector emitted approximately 1.4 billion metric tons of CO₂ equivalent. According to CREA, about half of the total carbon emissions in India in FY2021-22 were estimated to have originated in the power sector.
Similarly, the country also needs to take into account water scarcity that already affects millions of Indians, with tech hubs in Maharashtra and Karnataka already stressed due to the existing data centre infrastructure.
Sustainable way forward for AI in India
As AI continues to reshape industries, experts warn that its environmental impact is already being felt. India, however, has a unique opportunity to lead in building climate-conscious AI infrastructure from the ground up. The experts outlined principles that the country must follow to make sure AI doesn’t become our next climate problem:
Dikshant Dave emphasises that while innovation is essential, it must not come at the cost of the planet. “Newer model architectures, energy-efficient chips, liquid cooling, and renewable-powered data centres are all moving the field toward sustainability. AI can become less harmful, but it requires intentional design choices and strong accountability.”
Abhishek Agarwal highlights the rapid adoption of AI by both government and businesses, stressing the need to balance innovation with long-term stability. “India should encourage collaboration between academia, industry, and government to explore green AI technologies,” Agarwal adds. “With the right plan, AI can drive both economic growth and environmental health.”
Manoj Dhanda and Nitin Lahoti echo the sentiment, calling it a chance for India to do things right from the start. If India aligns its tech growth with sustainability policies, it can become a global model for scaling AI responsibly without straining natural resources, they said.
The experts share several measures the government and companies can follow to make sure AI in India evolves as both powerful and planet-friendly:
India is well aware of its heavy dependence on coal for electricity and has been actively working to curb emissions in the power sector by ramping up its share of clean energy from renewable sources.
The country aims to achieve about 50 per cent of its installed capacity from non-fossil-fuel energy sources by 2030. However, as AI accelerates electricity demand, transitioning to a low-carbon system could become increasingly challenging. Without careful planning, this surge could become an obstacle to India’s long-term goal of achieving net-zero emissions by 2070.
