How the AI Bubble Bursts

I have had this bouncing around my head for a few days and wanted to get it out before the new year. Everyone is doing predictions, and here is mine – before I start the focus on workplace / professional mental health. Everyone is predicting an AI bubble burst. The trick is going to be to know when it’s going to hit and timing your exit properly. Too early, you miss some opportunity. Too late and you lose your shirt. I look back at my own work history, and I think I can use that to deduce how and why the AI bubble will burst, not just exactly when. I’m hoping this gives you (and me) some ideas of what to watch for. We’ll have more warning track than we think. In the late 90’s and into late 2001, I worked for a small construction company. This company usually did work with oil patches, oil pipeline maintenance, and directional drilling. At some point before I started working there, someone figured out that the equipment and processes they used for that work were almost identical to the equipment and processes needed to put fiber optic cable in the ground. This was the heyday of the dotcom bubble, and Qwest, T-Cubed, SBC, AT&T, and anyone with a will to do so was putting as much fiber in the ground as possible. The big providers put cables in the ground as fast as they could, using money from the flourishing dotcoms. The dotcoms took investor dollars and invested in the big providers to build out the internet. It was a self-sustaining, self-reinforcing loop of apparently infinite wealth. I didn’t realize it at the time, but I had a front row seat to watch the beginning of the dotcom bubble burst. In some lab someplace, someone created Dense Wavelength Division Multiplexing (DWDM) transceivers. Why is this important? The fiber we’d been putting in the ground – and the outlays for the next 5 years – was put in with the understanding that transceivers could encode/decode 8 channels on a strand of fiber. With DWDM transceivers the floor became 96 channels, with the capability to extend to 128 or even 160 in some cases. That’s with the existing fiber in the ground – just replace the transceivers at the ends of the run and you had at least a 12-fold increase in capacity, with no new fiber needed. As early as May of 2001, there were demonstrations of 10Tbps using the new tech. And just like that, the capital investment to build out the infrastructure of the Internet went from needing billions for big fiber runs to being a few million to replace some transceivers. The funding loop of dotcoms funneling money to big providers and providers funneling money to dotcoms vanished – almost overnight. It wasn’t that there were too many dotcoms or that the infrastructure wasn’t ready (I have no doubt that some of the fiber we put in the ground is still sitting there dark thanks to DWDM’s extended capacity) – it was that the finance model got innovated out of existence. Fast forward to today. We have a bunch of AI providers out there. The AI providers invest in GPU’s, datacenters, and energy infrastructure. The GPU makers, datacenter / hosting platforms, and energy companies invest in AI providers. It’s a loop – just like in the fat times of the dotcoms. I suspect that somewhere in some lab, someone is working on a dedicated AI chip that will give 10X the AI output for one quarter the energy input. Maybe ratios are slightly more or less, but someone is going to do this, and when they do the loop will break. Instead of thousands of datacenters dotting the countryside consuming the energy of a small star, you will have broken leases, empty datacenters, and piles of old GPU’s headed to recycling centers. That is how I think the AI bubble will break. AI providers and investors are over-leveraging against a vision of tomorrow that is just a bigger louder version of today – and the future almost never plays out that way, at least not for long. When the bubble breaks, it will not be because we had 5,000 providers and that 5,001st was just too much. It will be because one of the legs of the finance stool gets kicked out from under it. Probably with a chip – possibly with a distributed computing model that uses idle time from millions of devices around the world, or something else. But sometime within the next five years it will likely burst. So I’m looking for innovation in those areas (chips and distributed computing). The first DWDM transceivers were demonstrated in 1998, before I even started at the company doing fiber work. By the time I left in August 2001, we were seeing job cancellations for 2002. All the signs were there; you just needed to know how to read them. I think the same holds true for today. A long post, but thanks for reading. I hope you have a good, productive, and healthy 2026.