Matteo Wong writes:

The AI boom has showered some of the nation’s most prominent companies in market value. OpenAI and Anthropic are now the two most valuable private companies in the world. Google, Microsoft, and Nvidia have become larger than ever. But among the biggest winners has been Caterpillar, a purveyor of yellow trucks and cranes. Caterpillar’s stock has more than doubled in value over the past year, making the company worth six times as much as Nike. Its ascent has little to do with the construction vehicles (both real and toy) that it is known for. Instead, Caterpillar’s success comes from its giant gas-powered engines that are helping power the nation’s data-center build-out.

For all the attention given to improving model capabilities, the most important inputs to AI are not bits but atoms and electrons. To train and run its models, the AI industry needs to build. Data centers are incredibly complex technological and industrial operations—they require tech companies to erect power plants, build wastewater facilities, and set up all manner of electrical equipment, power lines, and advanced-cooling equipment. That calls for concrete, steel, silicon, glass, copper, and liquefied natural gas.

This data-center construction has long been frenzied, but now it is fast approaching an inflection point that is both awesome and alarming. The major players—Amazon, Google, Microsoft, Meta, and Oracle—are on track to spend more on data centers by the end of the year than they bring in from their operations. That means tech giants whose non-chatbot businesses bring in tens of billions of dollars every year will likely need to take on debt to keep paying for their AI ambitions. From the launch of ChatGPT, in late 2022, through the end of last year, these firms’ capital expenditures—most of which go toward data centers—well exceeded half a trillion dollars. The companies intend to spend a similar amount in 2026 alone. And next year, their AI investments will likely exceed $1.1 trillion, according to a recent report from J.P. Morgan.

Silicon Valley has long enjoyed a straightforward advantage: Its products and services cost relatively little to make but generate tremendous revenues. But operating a data center is in many ways more like running a steel mill than a smartphone app. In particular, data centers require stunning amounts of electricity, especially as AI models have gotten more powerful. Five years ago, a standard data center might require 10 or 50 megawatts of power, enough for tens of thousands of homes. A year ago, data centers demanding a gigawatt or two of power were considered gargantuan. This week, Meta announced that it is more than doubling the size of its flagship AI data center, which will increase its peak power needs to five gigawatts. A data center proposed in Utah, if completed, will demand nine gigawatts of power. That’s several large cities’ worth of energy pulsing through a few big warehouses, all dedicated to bots. [Continue reading…]