Power consumption has emerged as the starkest measure of artificial intelligence's explosive growth, and whether that trajectory is sustainable.
Google, Microsoft and Amazon all touted efficiency improvements in sustainability reports released recently. Yet those gains are being drowned out by sheer expansion. Google's electricity consumption jumped more than 140 percent between 2021 and 2025, already surpassing projections made in 2023 by researcher Alex de Vries-Gao of VU Amsterdam.
The scale is staggering across the sector. De Vries-Gao estimates that new electricity demand from Google, Microsoft, Amazon and Meta combined between 2022 and 2025 roughly equals New York City's annual power consumption. That growth has sparked a fundamental question: Are the benefits worth the environmental and financial costs?
Tech company sustainability leaders have been deliberately vague when pressed on the tradeoff. Kate Brandt, Google's chief sustainability officer, spoke of being "deeply committed to responsibly managing" operations but didn't address whether growth might slow. Melanie Nakagawa at Microsoft framed the issue as a matter of reducing intensity per unit rather than questioning expansion itself. "Future growth really does become increasingly decoupled from that future impact," she said, describing the goal.
The industry has long operated on a simple assumption: larger models yield better performance and ultimately higher profits. But there's a catch. Many applications touted by companies today rely on narrower models, not the massive frontier models driving data-center buildouts. While companies argue that advances in frontier models eventually unlock practical applications downstream, skeptics question whether the current investment wave justifies the immediate costs.
Boris Gamazaychikov, co-founder of Sustainable AI Group, has a specific fix in mind. AI models should disclose standardized energy-efficiency metrics, he argues, the way cars display fuel economy. "You don't need a Hummer to go to the grocery store," he said, capturing the inefficiency problem.
The real brake on AI expansion may come not from environmental concerns but from simple economics. Daron Acemoglu, an MIT economist and Nobel laureate, warns that if investment continues racing ahead of actual demand, the boom will eventually collapse on financial grounds alone. History suggests this pattern repeats: canals, railroads and the internet all sparked massive investment booms years before payoffs materialized. The current AI wave appears even more extreme than those episodes, according to recent analysis by international central banks.
Industry leaders counter that the next few years will prove transformative, that this technology is exceptional enough to break the pattern. Acemoglu remains unconvinced but stops short of ruling it out completely. "I find that not completely convincing," he said. "But it cannot be completely ruled out."
Author James Rodriguez: "The efficiency theater from Big Tech is nice, but it's not math that matters here,it's megawatts, and they're doubling faster than any company can spin an efficiency story."
Comments