The demand for AI, especially generative AI, has been increasing exponentially in the past two years. With this, comes a staggering increase in the energy intensity, and, by extension, carbon emissions intensity of AI-related activities. The Government should look into regulating the carbon footprint of the use and development of AI.
To put things into perspective, just one query with ChatGPT uses up to 10 times as much energy as a Google search query, enough to power a light bulb for 20 minutes. (NPR) Generating a single image using an AI model can take as much energy as fully charging a smartphone, thousands of times more energy-intensive than generating text. (MIT)
However, not all AI models are the same. Using large, general-purpose models, such as large-language models like ChatGPT, can consume up to 30 times more energy than a smaller model created specifically for the task. (MIT) The issue arises when companies of different sizes rush to integrate general-purpose AI into their products and services, leading to such general-purpose AI models being deployed in an indiscriminate manner millions if not billions of times a day, for tasks that users do not require the use of general-purpose AI, thereby causing emissions to add up quickly.
I would like to ask the PMO if there is an estimate of how Singapore’s energy demand, especially for industries, may increase with the push for widespread adoption of AI. As Singapore is positioning itself to become an AI hub as well as a data centre hub, has this been taken into consideration in the 2035 Nationally Determined Contribution (NDC) commitments?
The Government should start to look into developing voluntary guidelines, and eventually mandating, the reporting of energy consumption by AI models. This has already been recommended by the United Nations Environmental Programme (UNEP), and governments, including those in the United States and the European Union, are actively discussing similar regulations.
This would encourage developers to explore and adopt energy-efficient practices to reduce the carbon intensity of their products. AI developers would be pushed to use smaller, task-specific AI models instead of large, general-purpose models.
Such regulations would also have ripple effects within the AI ecosystem. Like the sustainability reporting landscape, there will be increased demand for products and services that help support such disclosures. Data centres in Singapore may start looking at developing hardware that can more easily show AI developers their energy consumption (MIT). There would be greater innovation (MIT). Downstream businesses would also be more discerning about the AI models they adopt, taking into account their energy consumption and emissions.
By being an early adopter of such regulations, Singapore can enable local AI developers and related businesses to integrate carbon-conscious practices early and stay ahead as global demand for sustainable AI solutions grows.
