By: Dimpho Sekhaolelo - Quantitative Analyst and ESG Committee Chair at Prescient Investment Management

Every time you ask a chatbot a question, somewhere in the world a server warms up. What feels like a simple digital interaction is, in fact, a physical process. Behind every prompt sits a data centre that consumes electricity, generates heat, and relies on water to keep systems cool and running.

 And that physical footprint is no longer marginal. AI, and increasingly agentic AI that can complete multi-step tasks autonomously, has shifted from a curiosity to a core accelerator of business operations. The scale of adoption is moving faster than most balance sheets have caught up with. Anthropic’s Claude alone recorded close to 290 million website visits in February 2026, up almost 31% month on month, with mobile usage growing nearly 50% over the same period. Multiply that pattern across OpenAI, Google, Meta, and the agentic AI tools now embedded in every major enterprise software suite, and demand on physical infrastructure becomes the headline story.

 Globally, data centres already consume more than 560 billion litres of water a year (Aquatech, 2023). For a country where the grid is fragile and entire suburbs already live in de facto Day Zero conditions, that reality is no longer abstract. South Africa is not a bystander to the global AI build-out; it is increasingly a destination for it. Microsoft has pledged a further ZAR 5.4 billion, roughly USD 297 million, to expand its cloud and AI infrastructure in South Africa by the end of 2027, on top of more than ZAR 20 billion already invested locally over the past three years. The latest commitment explicitly earmarks investment in power and water readiness. The water question, in other words, is no longer someone else’s problem.

Recent studies have attempted to quantify the water usage of AI systems, raising the question: how much water does a single AI prompt consume?

Estimates of water use per AI interaction span three orders of magnitude, depending on the model, the data centre, and what exactly is being measured.