I think that realistically, the thing that kills AI data centres isn't going to be cost, regulation, or public opposition. I think it's going to be the gen-AI equivalent of the WDM fiasco.
For context, back when commercial Internet service was just taking off and the dot-com bubble was at its peak, a bunch of companies put down vast quantities of fibre-optic cables in anticipation of future demand. The vast majority of that fibre subsequently ended up laying dormant, becoming what's known in the industry as dark fibre.
The dot-com bubble bursting is often cited as the reason why, but it can't account for all of the problem, or even most of it; demand for commercial bandwidth continued to grow in spite of the bust, albeit somewhat more slowly than predicted. The real culprit was the development of a technology called wavelength-division multiplexing, or WDM.
WDM tremendously increased the amount of data that could be transmitted using existing fibre-optic cables; post-WDM, many major urban centres found that they now had over one hundred times as much fibre as they needed to accommodate present and anticipated near-future demand for Internet bandwidth. Many of the corporations that had speculatively bankrolled that fibre installation ended up selling off their infrastructure for pennies on the dollar and subsequently went bankrupt.
Basically, I suspect that's what's going to happen here. The present rush to construct more and more AI data centres is being short-sightedly driven by predictions based on the first couple of generations of general-purpose generative AI technology. Any day now there's going to be some random unpredictable breakthrough that makes gen-AI models like a hundred times more efficient, and a lot of very powerful people are going to be left holding a lot of very expensive bags.
