AI Carbon Footprint: Less Than Streaming, More From Data Centres

Concerns about artificial intelligence’s growing environmental impact are widespread, yet an hour of Netflix viewing can generate 500 times more carbon dioxide than sending two text prompts to AI models like Gemini or ChatGPT. A new report by TRG Datacenters indicates that many other digital activities contribute significantly more to carbon emissions.

Credit: GETTY

Data centres, encompassing AI and cryptocurrency operations, used an estimated 460 terawatt-hours of electricity in 2022, a figure projected to exceed 1,000 terawatt-hours by 2026, according to the International Energy Agency. In 2023, data centres consumed approximately 4.4% of all United States electricity, with this expected to rise to between 6.7% and 12% by 2028, as reported by the U.S. Department of Energy.

Blaming AI solely for increasing energy demand may be too simplistic, the TRG Datacenters report suggests. Many common digital activities beloved by users potentially have a greater environmental cost.

For example, one hour of high-definition YouTube or Netflix streaming consumes approximately 0.12 kilowatt-hours, releasing 42 grams of carbon dioxide. This ties for the highest single activity measured in the study.

An hour-long Zoom call is more efficient, using approximately 0.0486 kilowatt-hours and releasing 17 grams of carbon dioxide. Generating a six-to-ten-second text-to-video clip produces about 17.5 grams of carbon dioxide, similar to the environmental impact of a Zoom conference.

Lesser impacts include sending a short email without an attachment, which releases 4.7 grams of carbon dioxide. An AI image generation produces 1 gram of carbon dioxide per image, while a voice assistant query generates 0.175 grams of carbon dioxide. A Google search or an AI chatbot prompt releases 0.105 grams of carbon dioxide.

Two Gemini prompts combined produce approximately 0.084 grams of carbon dioxide, highlighting the relatively low impact of simple AI queries. However, creating text-to-video content is an exception, requiring significant data centre energy and representing a heavy computational load. The study did not examine the energy-intensive process of training AI models, which can involve hundreds of thousands of high-end graphics processing units over months.

The tech sector added roughly 900 million tons of carbon dioxide to the atmosphere in the last year, a volume comparable to Germany’s annual emissions. This figure is expected to exceed 1.2 billion tons by the end of 2025.

The issue lies not in technology use, which is inevitable, but in its power source. Only about 30% of data centre energy currently originates from renewables. Increasing this to 80% or 90% could reduce the carbon footprint of every digital activity by more than half without requiring behavioural changes.

A Goldman Sachs report states that all forms of AI will drive a 165% increase in data centre electricity usage by 2030, underscoring the urgency of transitioning to renewable energy. Apple currently powers all its corporate operations, including offices, retail stores, and data centres, with 100% renewable electricity.

The company also has an Apple 2030 goal to make its entire business, including manufacturing, supply chain, and customer use, carbon neutral within five years. Google’s seventh-generation Ironwood AI chips are designed to be both more powerful and energy-efficient, aiming to reduce power demand while meeting increasing digital appetites.

---

• Streaming one hour of Netflix can generate 500 times more carbon dioxide than sending two AI text prompts.

• Data centres, including AI operations, are projected to increase electricity consumption significantly by 2026.

• The overall tech sector emitted approximately 900 million tons of carbon dioxide last year.

Source: FORBES