The carbon impact of AI and video

Carbon impact of AI

The carbon footprint of AI primarily arises from the training and usage of machine learning models:

• Training AI Models: Training AI models, especially large ones like GPT, is highly energy-intensive and can generate several tons of CO₂, depending on the model size, the number of iterations, and the energy efficiency of the data centers used.
• AI in Production: Once an AI model is trained, its production use (responding to queries) is less energy-intensive, but the impact can accumulate with a large number of queries. For example, a simple search query or interaction with a virtual assistant consumes less energy than streaming a video, but at scale, this usage becomes significant.
• Generative AI’s Impact: Generative AI severely challenges the environmental goals of tech giants. The construction of new data centers (Scope 3 emissions) increases their carbon footprint—29% for Microsoft between 2020 and 2023, and 48% for Google since 2019. Amazon plans to invest $150 billion in data centers over the next 15 years. Microsoft, in partnership with OpenAI, is investing $100 billion in the development of the Stargate supercomputer, expected to launch in 2028 and consume up to 5 gigawatts of electricity at full capacity—equivalent to one-twelfth of France’s nuclear capacity.

Carbon impact of Video

Video streaming is energy-intensive due to several factors:

• Data Transmission: Videos require the transmission of large amounts of data over the internet. Video streaming accounts for about 60% of global internet traffic, and this proportion is increasing. This data transmission consumes energy for the servers hosting or delivering the videos and the network infrastructure. For example, one hour of HD video streaming can generate around 300 g of CO₂, depending on the quality and type of network connection.
• Device Energy Consumption: Video playback on user devices (computers, smartphones, TVs) consumes energy for decoding and display. High-resolution screens and powerful devices amplify this consumption.
• Sustainable Streaming: There’s a growing body of work on measuring greenhouse gas emissions related to streaming, promoting digital sobriety, and advocating for responsible streaming. An entire blog category is dedicated to this topic.

Comparison between AI and Video

It’s challenging to directly compare the impact of large-scale data center construction (Scope 3) for growing computational needs with the electricity consumption (Scope 2) of streaming videos.

• Per-Action Carbon Impact: If we consider the carbon impact per action (e.g., an AI query vs. an hour of video streaming), video generally has a higher impact, especially in high quality.
• AI vs. Streaming: AI can have a massive impact at scale during the training of large models, but for daily tasks, its carbon footprint is relatively low compared to continuous video streaming.
• Usage Patterns: Video streaming is often a prolonged activity, whereas using AI for specific tasks (like one-time queries) is more transient.
• Overall Impact: In general, for one-time usage, streaming video has a higher carbon impact than an AI query. However, training large AI models is extremely energy-intensive and can, at scale, surpass the impact of video streaming.

Conclusion

Reasonable practices and digital sobriety are the best approaches to minimizing the overall carbon footprint of these technologies. Streamlike suggests combining AI and video thoughtfully for specific justified uses.