Consulting firms warn that as AI adoption spreads and costs fall, power demand from AI data centers could continue to climb despite efficiency improvements.

AI costs are falling, and the pressure on demand is rising.

AI is getting cheaper to use, but as companies use it more, the electric grid faces mounting pressure. In two new reports published this week, McKinsey and Boston Consulting Group found that as companies move beyond experimenting with chatbots and deploy AI across more of their operations, cheaper models and falling token prices could justify more frequent, higher-volume AI use—a dynamic that threatens grid stability.

Companies typically pay AI providers based on tokens, the small units of text that models read and generate. In the near term, McKinsey said that cheaper tokens will make AI more widely used and increase electricity demand, even as companies work to improve the efficiency of models, chips, and data centers.

This matters because data centers already consume more power. McKinsey called data-center electricity demand "the fastest-growing load segment in OECD power markets," noting that in several markets, data centers are already the largest driver of expected electricity growth through 2030. McKinsey projects global data-center electricity demand will grow 24% annually through 2030, then slow to 5% annually between 2030 and 2040.

Yet companies deploying AI at scale are not simply allowing token spending to rise unchecked. Boston Consulting Group found that three-quarters of its "future-built" companies—the most AI-mature group in a survey of 1,300 C-suite and senior executives across 20-plus sectors—had established explicit targets for token spending and managed it to deliver specific returns. Half of these advanced companies actively encouraged employee use of paid AI tools to maximize adoption, compared with 25% of lagging companies. Meanwhile, 22% of future-built companies imposed limits or controls on usage to manage costs.

But there is a tradeoff. As companies make each AI task cheaper and less power-intensive, AI becomes useful in more applications, and total power demand rises. McKinsey said the broader picture for data-center growth after 2030 remains uncertain because companies are still determining whether AI delivers sufficient value. The reports suggest that cheaper AI will not automatically reduce energy use; instead, it may drive businesses to find many more ways to deploy it.