AivexaNewsSearch
AI news for builders and product teamsChecked every hour

How to Make Artificial Intelligence More Environmentally Sustainable

Collected Oct 29, 2026

Charles Humble, speaking at GOTO Copenhagen and interviewed afterward by InfoQ's Ben Linders, made the case that environmental sustainability belongs alongside cost, performance, security, regulatory concerns and reliability as a top-level consideration when designing computing workloads, AI included. His argument rests on scale: IT accounts for roughly 2 to 4 percent of global emissions, he said, compared with about 3 percent for aviation, and he contrasted an aviation industry working hard to cut emissions with an IT industry he described as moving the other way.

The numbers he cited come from the companies' own reporting and from the International Energy Agency. IEA figures put combined electricity use by Amazon, Microsoft, Google and Meta at more than double its 2017 level by 2021, around 72 terawatt-hours. Microsoft's 2024 environmental report recorded total carbon emissions up 30 percent since 2020, attributed primarily to data center construction for AI demand. Google's emissions rose about 50 percent since 2019, including 13 percent year on year in 2023, largely driven by AI. The IEA also estimated global data center consumption at 240 to 340 TWh in 2022, 1 to 1.3 percent of global electricity demand, with networks adding 260 to 360 TWh; Humble put the combined figure near 700 TWh, comparable to Brazil's 726 TWh over the same period.

On climate context, Humble noted that the referenced emissions are mainly carbon dioxide, a greenhouse gas whose atmospheric concentration has risen since the Industrial Revolution, trapping more heat. The stated goal of holding warming to 1.5 degrees by 2100 looks unlikely on current projections, which he put at roughly 2.5 to 3 degrees.

His practical guidance starts with questioning whether AI is needed at all. Teams should ask whether a new model must be trained from scratch or a pre-built one will do, whether relaxing part of an SLA would allow workloads to run in a greener location, and whether demand shaping is feasible. When a custom model is genuinely required, he suggested scrutinizing how much data must be collected, noting that smaller models sometimes perform better, and favoring open-source data sets. Data collection itself can be shifted in time or place toward periods and regions with green energy available.

Training on renewable energy, he said, can in theory cut carbon emissions by somewhere between 45 and 99 percent. He also argued that much of the energy may be spent deploying and maintaining a model in production rather than training it, which shifts attention to inference efficiency.

Four steps he offered for greener computing in general: use the smallest hardware configuration that can safely run the job; run compute where low-carbon electricity is abundant and where credible plans exist to clean the grid further; choose cloud providers with data centers in green locations and good carbon tooling; and optimize job execution time to reduce the footprint further.

The location argument, as he explained it, is that a workload consumes a fixed amount of energy but the carbon cost of that energy varies enormously by place and time, depending on the generation mix powering the local grid. A region running mostly on hydro, solar, nuclear and wind, such as Norway, Sweden or France, can be many times cleaner than one dependent on coal or gas. Carbon intensity also fluctuates through the day and across the year as wind, solar and demand change. This is the basis of carbon-aware computing, where demand shifting moves work either in time or in location. Time-based shifting suits latency-insensitive work such as AI training; location-based shifting ranges from simply defaulting new latency-tolerant workloads to lower-carbon regions up to moving between hemispheres seasonally to chase sunlight.

Demand shaping, in his framing, means changing how much work software does, or what kind, in response to how clean or constrained the electricity supply is at that moment. One version hands users a choice: an AI-powered service might offer to connect them with a person instead, up front or at any point in the process. Another is graceful degradation tied to carbon intensity, such as serving lower-quality video or images, disabling autoplay, skipping non-essential animations, reducing refresh rates, or falling back to a smaller model. The trade-offs he names are added complexity and the need to judge whether the reduced service remains acceptable.

For training specifically, his advice is to avoid starting from scratch wherever possible: fine-tune a pre-trained model, or simply prompt one or use retrieval-augmented generation, the pattern where relevant documents are fetched at query time and supplied to the model as context, if that meets the need. Shrinking a model can speed training and improve efficiency, and he lists distillation, quantization and pruning as compression techniques. Quantization, for instance, stores weights at lower numeric precision to cut memory and compute. Federated learning, which trains on devices locally rather than centralizing data, can be greener in some cases. He pointed readers wanting depth to Raghavendra Selvan's book Sustainable AI.

A likely trade-off for teams is that carbon-aware scheduling adds operational complexity: you need carbon-intensity data for candidate regions, workload classification by latency tolerance, and fallback behavior when a greener region is unavailable. The latency-sensitive tier of most consumer products will be hard to shift at all, which is why training and batch jobs are the natural first targets. His point that inference over a widely used model's lifetime may outweigh training emissions suggests measurement, not assumption, should decide where effort goes.

Why it matters: Developers and product teams building AI features will increasingly be asked to justify model size, hosting region and training approach against carbon as well as cost and latency, and the hyperscalers' own reports supply the figures that make the question concrete. The suggested levers, pre-trained and smaller models, carbon-aware placement, and user-facing degradation or human-handoff options, are mostly architectural choices that are cheaper to make early. My inference is that pressure will concentrate first on large-scale training runs and inference serving, where the emissions and the flexibility are both largest.

Read at InfoQ · AI, ML & Data Engineering

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

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. By Ben Linders