Teaching with Generative AI

How does GenAI impact the environment? 

The 'engines' of GenAI, data centres require considerable raw material extraction. Far from existing in some nebulous cloud, data centres house computer servers, network infrastructure, cooling systems and other equipment. The electronics alone take an impactful toll on the environment, given that 800kg of raw materials is needed to manufacture just one 2kg computer. With some estimates putting the life-span of GPUs - the circuits used in GenAI applications - at between one and three years, the extractive costs of data centres are ongoing.

Speaking of these short lifespans, data centres contribute to the growing problem of ewaste. A recent study suggests that LLMs alone could contribute up to 5.0 tons of additional ewaste by 2030. The United Nations Institute for Training and Research expects overall worldwide ewaste to grow 33% by 2030. In 2022, less than a quarter of ewaste was documented as having been properly recycled. Improperly processed computer and network components exposes the environment to toxins like mercury and lead.

Water consumption, both for power generation and cooling data centre servers, is another area of environmental impact. While hard data is sparce, researchers from the Lawrence Berkeley National Laboratory (LBL) see hyperscale data centres alone consuming up to 124 billion litres of water in the United States alone by 2028. The same report pegged indirect water consumption due to data centres' use of electricity at 800 billion litres annually.

GenAI relies on considerable electricity consumption. Electricity is used to power servers, the cooling systems and the networks that support GenAI. The heightened power density of these systems means that, for example, a GenAI training cluster's power consumption might exceed typical computing workloads by up to eight times. Again, reliable data on actual electricity use is difficult to find, but the upward trend is unmistakable. For example, electricity consumption by Google's data centres doubled between 2020 and 2024, due mostly to AI development at the end of that period. Data from the IEA suggest that, while current data centre electricity consumption may be relatively modest compared to other industries, the projected scale of investment in GenAI points to ever-increasing impact. Data centres will account for between six and twelve percent of the United States' total energy use by 2028, depending on the demand and efficiency of GenAI systems.

GenAI's carbon footprint is complicated to calculate. For data centres alone, it depends on various factors such as the efficiency and size of the LLM, the nature of the prompt, efficiency of the centre's architecture and servers, whether or not the power source is renewable, and even the time of day. Of course, manufacturing and transportation of equipment and buildings have to be accounted for as well. That said, we know that the cumulative carbon footprint must be considerable. A 2021 paper put out by UC Berkely and Google estimated that training ChatGPT3 (that is, not considering deployment) created 552 tons of CO2 equivalent. For context, just one ton of CO2 is equal to driving a gasoline car 8047km.

Researchers from MIT describe the current rate of GenAI growth as unsustainable in environmental terms.        

How impactful is my use of GenAI in terms of energy and water consumption? 

It depends whom you believe.

• A oft-quoted estimate from early in the GenAI explosion pegged the electricity consumption in a typical ChatGPT prompt at 2.9 watt-hours, which is about ten times the figure of a Google search (roughly 0.3 watt-hours).
• In a June 2025 blog post, OpenAI's Sam Altman stated that the electricity used for a typical ChatGPT is 0.34 watt-hours, "about what an oven would use in a little over one second", and that water consumption was about 0.00032176 litres, "roughly one fifteenth of a teaspoon." For context, the company also recently revealed elsewhere that ChatGPT receives 2.5 billion queries per day.
• An August 2025 blog post released by Google pegged the electricity used in its Gemini app's median text prompt at 0.24 watt-hours (Wh), and the water consumed at 0.26 milliliters. There are no figures for Gemini daily use but Gemini is increasingly integrated into Google Search, which is projected to receive 5 trillion queries in 2025. 

At present, anyone interested in measurements of GenAI's impact on the environment can consult either estimates that have been derived from complex calculations and/or lab research, or the figures released by GenAI companies. While Google/Alphabet has a respectable record in this area, requiring all technology companies to be more transparent would be an important step in the accurate assessment of GenAI's impact on the environment and identifying green solutions. 

Teaching with Generative AI LibGuide by BCIT Library Services is licensed CC BY-NC, meaning it can be used for non-commercial purposes if attribution is provided. Learn more about Creative Commons licenses on the BCIT Open Education LibGuide.

.