This article first appeared in Forum, The Edge Malaysia Weekly on June 8, 2026 - June 14, 2026

Artificial intelligence (AI) has become one of the most powerful applications of cloud computing. From writing code to analysing documents, AI is now embedded in everyday business workflows. As adoption accelerates, so does the environmental footprint of the infrastructure required to run these systems.

A common misconception is that AI’s greatest environmental impact lies in training large language models. In reality, the bulk of emissions arise from inference — the process of using a trained model to generate outputs. Inference occurs every time a user submits a prompt. While model training is controlled by a small number of AI providers, inference is driven by how organisations use AI in day-to-day operations. This gives businesses direct influence over both the cost and environmental impact of AI usage.

The scale of this impact is significant. The International Energy Agency projects that global data-centre electricity demand will more than double by 2030, largely driven by generative AI (Gen AI). The World Economic Forum estimates that accelerated AI adoption could result in an additional 4.2 billion to 6.6 billion cu m of water withdrawal by 2027. Data centres consume substantial amounts of electricity and water, particularly for cooling, placing increasing strain on global energy and water supply.

In many regions, electricity grids are already under pressure as demand grows faster than new capacity can be built. Water scarcity is also becoming a material risk as data centre cooling competes with residential, agricultural and industrial needs. In response, markets such as Malaysia and Singapore have moved from unrestricted data centre expansion to stricter, sustainability-focused planning frameworks, with greater emphasis on energy efficiency, carbon intensity and water usage effectiveness.

Not all AI workloads are equal. Traditional AI typically involves machine learning models trained on structured data for tasks such as classification, regression, anomaly detection and recommendation. These models are generally smaller, often run on central processing units (CPUs) or modest graphics processing unit (GPU) configurations, and consume relatively little energy per prediction. Gen AI, by contrast, relies on large language models with billions of parameters to generate text, images, code and other content. Each request requires substantial GPU compute, making Gen AI significantly more energy-intensive on a per-use basis.

The environmental impact ultimately stems from the infrastructure activated by each AI request. Every prompt triggers coordinated computation across GPUs, CPUs, memory, networking and storage in a data centre. Gen AI usage is commonly measured in tokens, representing units of text, images, audio or other data processed by the model. A single prompt can generate hundreds or thousands of tokens once both inputs and outputs are included. Although tokens are not a direct proxy for energy consumption, higher token volumes generally correspond to greater compute demand, and therefore higher electricity and water use.

There are practical levers available to organisations. Reducing unnecessary token usage is among the most effective. Instead of submitting entire documents, users can limit inputs to relevant sections. Clear and well-structured prompts often produce better results while consuming fewer resources. Where real-time responses are not required, non-urgent workloads can also be batched for more efficient processing.

Model selection matters as well. Not every business task requires the most advanced model available. Smaller models with billions, rather than tens of billions, of parameters can often meet operational needs at a fraction of the environmental cost.

As AI becomes embedded in routine business activity, its environmental impact can no longer be treated as an abstract or upstream issue. Every prompt, automated workflow and deployed model contributes incrementally to energy use, water consumption and infrastructure demand. For organisations with environmental, social and governance (ESG) commitments, responsible AI usage is therefore a matter of governance, not just technology. Greater visibility into inference usage — and disciplined choices around prompts, workloads and model selection — can help ensure AI innovation progresses without undermining sustainability and carbon transition objectives.

Francis Xaviour Joe is client relations director at Sustainlaterre PLT, a sustainability-related software solutions provider

Save by subscribing to us for your print and/or digital copy.

P/S: The Edge is also available on Apple's App Store and Android's Google Play.