IT is responsible for about 2 - 4% of all global emissions, and electricity use at Microsoft and Google has climbed because of AI, according to their own reports. Sustainability ought to be a top-level architectural concern. Some recommendations are: use pre-built and smaller models, run compute where low-carbon electricity is abundant, and train on renewable energy.
The Growing Environmental Impact of AI IT is responsible for approximately 2-4% of all global emissions. Major tech companies like Microsoft and Google have reported significant increases in electricity consumption and carbon emissions, largely driven by the demands of AI development and data center expansion. Global data center electricity use alone is comparable to the entire electricity consumption of Brazil, contributing significantly to greenhouse gases and climate warming projections. Key Questions for Sustainable AI in Business To embed sustainability into AI initiatives, Charles Humble suggests asking critical questions: Is AI truly necessary for the task? Can an existing pre-built model be utilized instead of training a new one from scratch? If service level agreements (SLAs) can be flexible, can workloads be run in locations with greener electricity? Can demand shaping techniques be applied to optimize energy use? Strategies for Greener AI Model Training When custom model training is unavoidable, Humble advises minimizing the data collected, opting for smaller models, and using open-source datasets. Implementing demand shifting can help align data collection and training with times and locations where green energy is readily available. Training on 100% renewable energy can lead to substantial carbon emission reductions, ranging from 45% to 99%. Four Steps for Overall Greener Computing The majority of energy consumption related to AI occurs during model deployment and ongoing maintenance in production. To achieve greener computing, four steps are crucial: employing the smallest hardware configuration sufficient for the task, running compute workloads in areas with abundant low-carbon electricity and strong green grid plans, utilizing cloud providers that offer data centers in sustainable locations and tools for footprint reduction, and optimizing job execution times to further reduce energy usage. How does location matter when it comes to reducing carbon emissions? The carbon cost of a compute workload's energy varies greatly by geographic location and time, depending on the electricity grid's sources (e.g., hydro, solar, nuclear vs. coal, gas). Running workloads in regions with high renewable energy (like Norway or France) is significantly cleaner. Carbon-aware computing uses 'demand shifting'—either time-based (for latency-insensitive tasks like AI training) or location-based—to prioritize cleaner regions or periods, with advanced methods involving seasonal shifts to maximize solar power usage across hemispheres. How does demand shaping work? Demand shaping involves dynamically adjusting the workload or functionality of software based on the real-time cleanliness or constraints of the electricity supply. This can manifest as giving users choices (e.g., AI vs. human interaction), or implementing graceful degradation, such as serving lower-quality media, disabling autoplay, or reducing refresh rates during high-carbon periods. Using smaller AI models also falls under this strategy. While effective, it introduces design complexity and necessitates evaluating the acceptability of reduced service quality. What can we do to make model training greener? Beyond demand shifting for both data collection and training, the primary recommendation is to avoid training AI models from scratch whenever possible. Instead, leverage pre-trained models for fine-tuning, or employ Retrieval-Augmented Generation (RAG). Reducing model size through techniques like distillation, quantization, and pruning can enhance training efficiency. In some scenarios, on-device training via Federated learning can also be a greener alternative. It's also vital to consider the lifetime emissions from widely used models during inference, emphasizing the importance of serving efficiency.
The Growing Environmental Impact of AI
IT is responsible for approximately 2-4% of all global emissions. Major tech companies like Microsoft and Google have reported significant increases in electricity consumption and carbon emissions, largely driven by the demands of AI development and data center expansion. Global data center electricity use alone is comparable to the entire electricity consumption of Brazil, contributing significantly to greenhouse gases and climate warming projections.
Key Questions for Sustainable AI in Business
To embed sustainability into AI initiatives, Charles Humble suggests asking critical questions: Is AI truly necessary for the task? Can an existing pre-built model be utilized instead of training a new one from scratch? If service level agreements (SLAs) can be flexible, can workloads be run in locations with greener electricity? Can demand shaping techniques be applied to optimize energy use?
Strategies for Greener AI Model Training
When custom model training is unavoidable, Humble advises minimizing the data collected, opting for smaller models, and using open-source datasets. Implementing demand shifting can help align data collection and training with times and locations where green energy is readily available. Training on 100% renewable energy can lead to substantial carbon emission reductions, ranging from 45% to 99%.
Four Steps for Overall Greener Computing
The majority of energy consumption related to AI occurs during model deployment and ongoing maintenance in production. To achieve greener computing, four steps are crucial: employing the smallest hardware configuration sufficient for the task, running compute workloads in areas with abundant low-carbon electricity and strong green grid plans, utilizing cloud providers that offer data centers in sustainable locations and tools for footprint reduction, and optimizing job execution times to further reduce energy usage.
How does location matter when it comes to reducing carbon emissions?
The carbon cost of a compute workload's energy varies greatly by geographic location and time, depending on the electricity grid's sources (e.g., hydro, solar, nuclear vs. coal, gas). Running workloads in regions with high renewable energy (like Norway or France) is significantly cleaner. Carbon-aware computing uses 'demand shifting'—either time-based (for latency-insensitive tasks like AI training) or location-based—to prioritize cleaner regions or periods, with advanced methods involving seasonal shifts to maximize solar power usage across hemispheres.
How does demand shaping work?
Demand shaping involves dynamically adjusting the workload or functionality of software based on the real-time cleanliness or constraints of the electricity supply. This can manifest as giving users choices (e.g., AI vs. human interaction), or implementing graceful degradation, such as serving lower-quality media, disabling autoplay, or reducing refresh rates during high-carbon periods. Using smaller AI models also falls under this strategy. While effective, it introduces design complexity and necessitates evaluating the acceptability of reduced service quality.
What can we do to make model training greener?
Beyond demand shifting for both data collection and training, the primary recommendation is to avoid training AI models from scratch whenever possible. Instead, leverage pre-trained models for fine-tuning, or employ Retrieval-Augmented Generation (RAG). Reducing model size through techniques like distillation, quantization, and pruning can enhance training efficiency. In some scenarios, on-device training via Federated learning can also be a greener alternative. It's also vital to consider the lifetime emissions from widely used models during inference, emphasizing the importance of serving efficiency.
