The Actual Environmental Cost of AI

Key Takeaways

• GPT‑3 training consumed ~5.4 million litres of water
• Inference dominates AI’s long‑term water and carbon footprint
• Data‑centre siting shifts to water‑stressed regions, raising policy concerns
• Scope 3 emissions from AI hardware construction are largely unreported
• Independent lifecycle assessments for frontier models remain unavailable

Pulse Analysis

The conversation around artificial intelligence’s climate impact has long been anchored to the one‑off energy and water use of model training. While those figures—such as the roughly 5.4 million litres of water needed for GPT‑3—are striking, they represent a static snapshot. In reality, the real environmental bill is generated every time a model processes a query, a process known as inference. With hundreds of millions of daily users, the cumulative water draw can reach tens of billions of litres per year, and the associated electricity consumption translates into thousands of tonnes of CO₂, dwarfing the training phase.

A second, often overlooked dimension is the geographic and supply‑chain context of AI infrastructure. Data‑centres are increasingly being built in regions already facing water scarcity because those locales offer grid capacity and cheaper land. This siting strategy externalizes environmental costs onto vulnerable communities, creating a policy blind spot. Moreover, corporate sustainability reports typically disclose only Scope 1 and Scope 2 emissions, leaving the much larger Scope 3 emissions—embodied carbon from hardware manufacturing and construction—unaccounted for. Microsoft’s recent report, for example, shows a 30.9% rise in Scope 3 emissions tied to datacentre expansion, yet it omits explicit attribution to AI workloads.

The absence of comprehensive, independent lifecycle assessments for today’s frontier models hampers meaningful debate and policy formulation. Without a full audit that tracks water, energy, and carbon from raw material extraction through end‑of‑life disposal, stakeholders are left comparing partial metrics that favor either critics or defenders. This knowledge gap fuels a fragmented discourse and delays actionable regulation. Industry players, regulators, and civil society must therefore push for transparent per‑query disclosures, standardized reporting frameworks, and funding for independent research to close the data gap and guide sustainable AI deployment.