Comment by Arn_Thor

There is one key way in which I believe the current AI bubble differs from the TMT bubble. As the author points out, much of the TMT bubble money was spent building infrastructure that benefited us many decades later.

But in the case of AI, that argument is much harder to make. The cost of compute hardware is astronomic relative to the pace of improvements. In other words, a million dollars of compute today will be technically obsolete (or surpassed on a performance/watt basis) much faster than the fiber optic cables laid by Global Crossing.

And the AI data centers specialized for Nvidia hardware today may not necessarily work with the Nvidia (or other) hardware five years from now—at least not without major, costly retrofits.

Arguably, any long-term power generation capacity put down for data centers of today would benefit data centers of tomorrow, but I'm not sure much such investment is really being made. There's talk of this and that project, but my hunch and impression is that much of it will end up being small-scale local power generation from gas turbines and the like, which is harmful for the local environment and would be quickly dismantled if the data center builders or operators hit the skids. In other words, if the bubble bursts I can't imagine who would be first in line to buy a half-built AI data center.

This leads me to believe this bubble has generated much less useful value to benefit us in future than the TMT bubble. The inference capacity we build today is too expensive and ages too fast. So the fall will be that much more painful for the hyperscalers.