IT accounts for about 2% to 4% of global emissions, and public reports from Microsoft and Google show their electricity use continuing to rise under the influence of artificial intelligence. Speaking at GOTO Copenhagen, Charles Humble argued that sustainability should be a top-level architecture concern. His suggestions include using prebuilt and smaller models, running compute where low-carbon electricity is plentiful, and training with renewable energy.
What the reported energy and emissions figures show
Humble cited International Energy Agency (IEA) data that the combined electricity use of Amazon, Microsoft, Google, and Meta more than doubled between 2017 and 2021, reaching about 72 billion kilowatt-hours in 2021.
Microsoft’s 2024 environmental report said its total carbon emissions had risen 30% since 2020, mainly because of new data centers built to meet AI demand. According to Google’s own report, its emissions had surged by about 50% since 2019 and rose 13% year over year in 2023 alone, driven mainly by AI. These reports are primarily about carbon dioxide.
Humble said greenhouse-gas concentrations in the atmosphere have been increasing since the Industrial Revolution. The aim is to limit warming to 1.5 degrees by 2100, but projections indicate warming of about 2.5 to 3 degrees.
He explained that electricity use is one reason IT is a significant contributor. The IEA estimates that global data centers used about 240 to 340 terawatt-hours in 2022, or 1% to 1.3% of global electricity demand. Communications networks added another 260 to 360 terawatt-hours. Together, that is about 700 terawatt-hours, roughly comparable to Brazil’s national electricity use of 726 terawatt-hours over the same period.
Humble put IT emissions at about 2% to 4% of the global total, compared with 3% for aviation. He said aviation is actively cutting emissions while the technology industry is moving in the opposite direction.
Questions to ask before adding AI
Before introducing AI in a business setting, Humble suggested asking:
- Do you really need it?
- Do you need to train a new model from scratch, or can you use a prebuilt model?
- If some service-level agreements can be relaxed, can the workload run in a lower-carbon region?
- Can you use demand shaping?
If a custom model is required, he recommended first assessing how much data actually needs to be collected. He said smaller models often perform better in many situations. He also suggested using open-source datasets and, where possible, shifting data collection to times and places where green energy is available.
Humble said that, in theory, training could use 100% renewable energy throughout and cut carbon emissions by 45% to 99%.
He argued that most energy use occurs after a model is deployed and during ongoing operations. He listed four ways to make compute greener:
- Use the smallest hardware configuration that can safely run the job.
- Run workloads where low-carbon power is abundant and the grid has a clear plan to become cleaner.
- Choose cloud providers with data centers in greener regions and with emissions-reduction tools.
- Schedule jobs to further reduce the carbon footprint.
In summary, Humble said sustainability should sit alongside cost, performance, security, compliance, and reliability as a top-level design consideration for compute workloads, and that systems should be built in a more environmentally responsible way.
Why location changes the carbon cost
InfoQ interviewed Humble after the talk. On why geography matters, he said the total electricity a compute task consumes is fixed, but its carbon cost varies sharply by place and time according to how the data center is powered.
If a data center uses grid electricity, running a workload in a region that relies mainly on hydropower, solar, nuclear, and wind—such as Norway, Sweden, and France—can be many times cleaner than running it where coal or natural gas dominates. Carbon intensity also changes through the day and year with wind, solar, and demand.
Carbon-aware computing can shift demand by time or by location. Time-based shifting is especially suited to work that is not sensitive to latency, such as AI training. The simplest location-based approach is to place new latency-tolerant workloads in low-carbon regions by default. A more advanced option is to move work between hemispheres with the seasons to make greater use of sunlight and solar power.
Demand shaping and greener training
Humble described demand shaping as adjusting a software workload or the type of task according to how clean power is at the moment and how tight supply is. One implementation is to give users a choice: for an AI-driven service, they could speak with a person instead of AI at the start, or switch to a person during a conversation.
He also described graceful degradation when carbon intensity is higher, such as lower-quality video or images, turning off autoplay, skipping non-essential animation, reducing the refresh rate, or calling a smaller model. The tradeoff is greater system complexity, and the reduced service level still has to be acceptable.
On making training greener, Humble said that, beyond shifting training and data collection, teams should avoid training from scratch when possible. He favored fine-tuning a pretrained model or, where that meets the need, prompting a model directly or using retrieval-augmented generation (RAG).
Smaller models can train faster and use energy more efficiently. Distillation, quantization, and pruning can compress models. In some cases, on-device training based on federated learning can be lower-carbon. He noted that for large, widely used models, inference often produces more emissions over the full lifecycle, so inference efficiency matters as well.
He pointed readers who want more detail to Raghavendra Selvan’s book Sustainable AI.
Source: InfoQ China. The article points to the English original at InfoQ.