GOMMA's $2.5T AI Bet Is the Biggest Capital Expenditure in History

A LessWrong article by dan.parshall highlights that Amazon, Google, Meta, Microsoft, and Oracle (dubbed GOMMA) have collectively committed $2.5 trillion to AI infrastructure—datacenters, Nvidia chips, and the like—making it the largest capital expenditure in history. Half of that money has already been spent, and the hardware depreciates at 25% per year. By the time the buildout finishes, roughly $800B of the purchases will have expired, leaving a working pile of $1.7T that sheds $420B annually from depreciation alone. Shareholders expect a 15% return on that pile ($250B), meaning GOMMA needs to generate $670B in profit per year from AI services.

To achieve that profit, GOMMA is betting on AI replacing human labor at scale. They rely on OpenAI, Anthropic, and their own in-house models to sell subscriptions that save companies more in salary costs than the subscription price. The article calculates that even if every software engineer in the U.S. were replaced (total salaries ~$450B/year), GOMMA would only capture one-third of that—$150B—far short of the target. Thus, the bet implicitly assumes AI will replace a much broader swath of the economy. The author compares this to 19th-century railroad investments, but notes that GOMMA's spending ramped up in just four years, signaling an extreme conviction that AGI is imminent.