Financial Times columnist questions big spending on AI data centres

In short: Tech giants are pouring trillions into AI computing power, but a Financial Times opinion piece warns this may not fix AI accuracy and could create financial risk.

What's going on

US tech “hyperscalers” like Meta, Microsoft, Alphabet, and Amazon are expected to spend more than $5 trillion on “compute” by 2030, according to Gary Marcus writing in the Financial Times. “Compute” is a simple shorthand for the raw computer power needed to train and run today’s AI systems, which often means building more data centres and buying large numbers of specialised chips.

This spending supports large language models, the type of AI behind chatbots like ChatGPT, Gemini, and Claude. The industry idea has been that if you add more chips and more data, the AI will get better, like giving a student more books and more study time.

Marcus argues there are two problems with that approach. First, bigger models still make “hallucinations” (when an AI confidently makes something up) and can make basic reasoning mistakes. Second, if many companies build similar models in similar ways, it is hard for any one of them to stand out, which can lead to price competition while costs stay high.

What to watch

Marcus suggests the world could end up with more data centres than it needs, especially if newer, more efficient AI models reduce the need for so many expensive chips. He also raises a broader concern about what happens if the spending does not pay off, including potential knock-on effects for investors like pension funds, and even pressure for government support.

Source: Financial Times