At GOTO Copenhagen, Charles Humble argued that environmental sustainability should sit alongside cost, performance, security, compliance and reliability as a top-level design concern for compute workloads. InfoQ reported the talk and a follow-up interview in which he outlined questions to ask before adopting AI and practical ways to cut the carbon cost of training and inference.<\/p>\n
Humble said the IT industry accounts for about 2% to 4% of global emissions. He noted that public reports from Microsoft and Google show electricity use continuing to rise under the influence of AI. The emissions discussed in those reports are mainly carbon dioxide.<\/p>\n
He 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.<\/p>\n
Humble said greenhouse-gas concentrations in the atmosphere have been increasing since the Industrial Revolution. The stated hope is to limit warming to 1.5 degrees by 2100, but projections indicate warming of about 2.5 to 3 degrees.<\/p>\n
On electricity, he cited an IEA estimate 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\u2019s national consumption of 726 terawatt-hours in the same period.<\/p>\n
He put IT-sector emissions at about 2% to 4% of the global total and aviation at 3%, and said aviation is actively cutting emissions while the technology industry is moving in the opposite direction.<\/p>\n
Humble suggested asking:<\/p>\n
If a custom model is required, he said teams should first assess how much data they actually need to collect. In many cases, he argued, smaller models perform better. He also suggested using open-source datasets and, where possible, shifting data collection to times and places where green energy is available.<\/p>\n
He said that, in theory, training can run entirely on renewable energy and that this can reduce carbon emissions by 45% to 99%.<\/p>\n
He argued that most energy use occurs after a model is deployed and during ongoing operations, and offered four ways to make compute greener:<\/p>\n
His closing point was that sustainability should be a top-level design consideration for compute workloads, together with cost, performance, security, compliance and reliability, and that teams should think about how to build systems in a more environmentally sound way.<\/p>\n
In the InfoQ interview after the talk, Humble said the total electricity a job consumes is fixed, but the carbon cost of that electricity varies sharply by place and time depending on how the data center is powered. If a data center draws from the grid, running a workload in a region that relies mainly on hydro, solar, nuclear and wind\u2014examples he gave were Norway, Sweden and France\u2014can be many times cleaner than running it where power comes mainly from coal or natural gas. Carbon intensity also changes through the day and the year as wind, solar and demand change.<\/p>\n
He said this is the basis for demand shifting in carbon-aware computing, done by time or by location. Time-based shifting is especially suited to latency-insensitive work such as AI training. The simplest location-based approach is to place new latency-tolerant loads in low-carbon regions by default. A more advanced approach is to move workloads between hemispheres with the seasons to make more use of sunlight and solar power.<\/p>\n
Humble described demand shaping as adjusting a software workload or the type of task according to how clean the electricity is at that moment and how tight supply is. One approach is to give users a choice: for an AI-driven service, they might talk to a person instead of an AI at the start, or switch to a person partway through a conversation.<\/p>\n
He also described graceful degradation when carbon intensity is higher, such as lower-quality video or images, turning off autoplay, skipping non-essential animation, lowering the refresh rate, or calling a smaller model. The tradeoff is higher system complexity, and the degraded service level still has to be acceptable.<\/p>\n
Beyond demand shifting for training and data collection, Humble said teams should avoid training from scratch when they can. He preferred fine-tuning a pretrained model or, when that meets the need, prompting an existing model or using retrieval-augmented generation (RAG).<\/p>\n
Smaller models can train faster and use energy more efficiently. Distillation, quantization and pruning can compress models. In some cases, on-device training with federated learning is lower carbon. He noted that for large models used widely, inference emissions over the full lifecycle are often higher than training emissions, so the energy efficiency of inference matters as well.<\/p>\n