THIS is with reference to the letter ‘AI and reality’ (Sept 3) and article ‘Dirty data centres’ (Aug 1) which rightly pointed out that consideration should be given to the challenges associated with building artificial intelligence (AI) data centres.
However, the emphasis on supposed excessive water usage in this regard is rather misplaced. Evidence that a handful of ChatGPT queries can consume 500ml of water is based on variable and often conflicting estimates. Most modern data centres, including the one inaugurated in Pakistan (Sky47), use direct-to-chip cooling systems that are closed loop cooling systems designed to significantly reduce water usage compared to evaporative cooling systems that were prevalent in the past.
Besides, coupling data centres with solar energy further increases their water efficiency. Recent analysis has shown that one of the largest data centres in the United States, approximately 50 times larger than Sky47, uses a mix of evaporative and direct-to-chip cooling, and consumes about the same amount of water annually as three fast-food restaurants. If there is a strain on water supply, it is likely caused by poor water management.
This is not to say the concerns are unfounded; only that fixating on water distracts from the more important elements. The strain on the electrical grid during peak hours is a real problem. So is the reliance on diesel generators to sustain 24-hour operations that carry their own environmental costs. More broadly, there is an important discussion that we should consider, which relates to the economic sustainability of AI-focused data centres. Graphics processing units (GPUs) depreciate very fast, depending on usage, and most front-running AI labs are yet to show a credible path to profitability, highlighting the risk of an AI bubble.
Worldwide, financial institutions are gradually realising the true holistic cost of AI usage and starting to pull back from over-investment. It is important that all those pushing for having data centres in Pakistan also look at the many long-term economic forecasts and demand patterns, instead of getting caught up in hype cycles.
Shareef Khalid
Lahore
Published in Dawn, October 4th, 2026
