By skyforbes 
As a student, entrepreneur, or employee, on every task like researching, you wonder that giving a prompt to LLM (ChatGPT, Gemini, or Claude) is actually easy and time-efficient, instead of normally searching it on the web like Google. That’s becoming completely normal in the AI era. You might find it easy and think it’s saving my time, and hence my money, but that’s not actually the seen; every prompt you give to ChatGPT releases three times more CO2 emissions than a Google search.
Infrastructure of AI
Artificial Intelligence revolutionized the way we access information and automates tasks by powering the infrastructure of enormous computational resources. These data centers, filled with high-performance GPUs and specialized hardware, require not only significant electrical power but also substantial water for cooling. The cumulative environmental footprint of billions of AI prompts daily is larger than most realize, raising questions about the sustainability of AI’s rapid expansion.
The infrastructure of prompts in ChatGPT is complex and highly specialized; it encompasses powerful data centers filled with specialized hardware, optimized software layers, and extensive energy and water resources to sustain operations.
- Data Centres and Hardware: The backbone support of LLMs is its data centres, specially designed for AI workloads, performing the matrix and tensor calculations essential for training and inference. These data centres are run by thousands of GPUs (Graphical Processing Units) or TPUs (Tensor Processing Units), and thus the computational demand is intense it runs 24/7 to do operations.
Electricity Usage: On average, a modern data centre GPU consumes between 250 -700W, which is a significant increase compared to traditional data center CPUs, which typically run at 150–200 watts. An H100 GPU running at an average of 600W continuously for a year consumes approximately 5,256 kWh of electricity, not including cooling overhead.
Water Usage: To manage the heat generated by these GPUs, data centers rely heavily on cooling systems. Water-based cooling is common due to its efficiency, but it results in high water usage. Millions of liters of water can be consumed for cooling purposes annually in large-scale operations.
2. Training and Inference: The training is the initial phase where the LLM learns from vast databases, requiring immense computational cycles on multiple GPUs in parallel. Training also involves substantial water use for cooling during these extended periods.
Interference: When the model is trained on real-world data, it goes to the inference phase, where the model generates responses to user prompts, while less intensive than training. Each inference consumes a fraction of energy and water but scales up cumulatively.
How Much Electricity Does ChatGPT Consume?
ChatGPT uses approximately 0.34 watt-hours of electricity per query, according to OpenAI, Towards Data Science, though some researchers suggest the smartest models can consume over 20 watt-hours for complex queries. To put this in perspective, 0.3 watt-hours is less than what an LED lightbulb or laptop consumes in a few minutes Epoch AI.
However, standard Google search uses far less, approximately 0.0003kWh of energy per search query, which translates to 0.2 grams of CO2 emissions. To visualize this, enough energy to power a 60-watt light bulb for 17 seconds.
The individual impact may seem tiny, but the reality is different. OpenAI reports ChatGPT has 700 million weekly users and serves more than 2.5 billion queries per day. This amounts to 850 megawatt-hours of electricity daily and 226.8 gigawatt-hours of electricity yearly, enough to power approximately 21,602 US homes for a full year.
Compared to Google, which serves more people and uses less energy, with over 3.5 billion searches each day, amounting to 1.05 gigawatt-hours (GWh), equivalent to the energy consumption of approximately 30,000 American homes. (Best Brokers)
How Much Water Do Data Centers Need to Stay Cool?
The water usage problem represents a critical and often overlooked dimension of AI’s environmental impact. According to “Business Energy UK A single 100MW data center can demand approximately 1.1 million gallons of water daily — an amount equivalent to the daily water usage of a city housing 10,000 people. To look into more scale, “The average data center uses 300,000 gallons of water a day to keep cool, roughly equivalent to water use in 100,000 homes.”
Per Query Impact: GPT-3, an AI model, is estimated to consume 500 ml of water per 10–50 responses, Towards Data Science. Each 100-word AI prompt is estimated to use roughly one bottle of water (or 519 milliliters). Balkan Green Energy Use
Major Tech Companies Analysis
In 2023, Google operations worldwide consumed 6.4 billion gallons of water (24.2 billion liters), with 95%, 6.1 billion gallons (23.1 billion liters), used by data centers RW Digital. Google’s data center in Council Bluffs, Iowa, consumed 1 billion gallons of water (3.8 billion liters). Google’s data centre water consumption has increased by nearly 88% since 2019, according to Climateq.
Meta
In 2023, Meta consumed 813 million gallons of water globally (3.1 billion liters) — 95% of which, 776 million gallons (2.9 billion liters), was used by data centers Chipkin.
Future Projections:
By 2027, global AI demand is expected to account for 1.1 to 1.7 trillion gallons (4.2 to 6.6 billion cubic metres) of water withdrawal, more than 4–6 times the total annual water withdrawal of Denmark, Towards Data Science.
Conclusion
The environmental costs of AI are real, measurable, and growing. However, there is hope: advances in energy-efficient AI models, specialized hardware, renewable-powered data centers, and innovative cooling technologies can significantly reduce the footprint. Users, developers, and companies must recognize that the convenience of AI comes with tangible resource demands. Sustainable AI is achievable, but it requires transparency, innovation, and collective commitment to balance technological progress with planetary health.