On Steam Power

Surprising nobody, one topic of conversation following my post about power tools was the environmental impact of generative models, and “are these new tools worth the cost?” is an entirely valid follow-up question. Running these models is incredibly expensive, with energy and water consumption numbers that boggle the mind—datacenters belonging to Google, Microsoft, and Meta consumed approximately 2.2×10⁹ m³ of water in 2022¹!

That’s 0.5% of the 444.4×10⁹ m³ consumed by the United States alone in 2020², but people need food and don’t need, as one person put it, “bullshit factories”. Which is true! I don’t need “spicy autocomplete” to write code any more than I need a router to cut a rabbet, but both are useful and neither are going away. To continue with the metaphors: these new power tools are themselves in their steam power era.

Steam power is a nice label that helps historians³ avoid having to mention coal. The initial mechanization of all industries (including carpentry) came at such a high environmental cost that “a black form of the peppered moth rapidly took over in industrial parts of the UK during the 1800s, as soot blackened the tree trunks and walls of its habitat”⁴.

The early days of the Industrial Revolution were grim, but there’s not much coal left in today’s industrial power mix—the UK is even on track to shut down their last plant this year⁵. This has mostly been made possible by a combination of electrification and improved grid composition, and the impact has been significant. In 2020, 70.6% of domestically-produced steel in the United States was made by electric arc furnace⁶ and 59% of California’s electricity came from carbon-free sources⁷. Municipal utilities have started experimenting with providing high quality recycled water (or “purple” water), and renewable sources in Europe now regularly cause power production to exceed demand, resulting in negative electricity prices⁸.

There is cause for optimism, and it’s important that we let ourselves feel that. If anything, it’s inspiring that people’s access to information today has improved so much that the costs of datacenters—which do not directly affect most people!—are such common knowledge; it only adds to the incentives pressuring them to improve.