On AI “Bubble”

Everybody likes to speculate how AI will change the world and the jobs. Reading Howard Marks’ memo on whether AI is a bubble, I got to think, what would happen if the AI bubble instead burst?

One of the key ideas from this memo is that technological bubbles often end up being “inflection bubbles”. That is, they create something: infrastructure, capabilities, etc. that changes the world, and the world never goes back to how it used to be. Examples he brings up include radio, aviation, and cars. Each technology survived a bubble, losing up to 97% of its stock value, and yet all these technologies were revolutionary and changed the world. Radio bubble, for example, created the infrastructure required for future radio and TV. Cars bubble created road networks and gas station infrastructure.

So, if the AI bubble were to burst, what would it leave behind? One of the largest expenses of today’s AI startups is data centers and GPUs. It stands to reason that if AI were to bust, then companies would be left with hundreds of thousands of unnecessary GPUs. Having this resource, they would likely try to rent or sell these capabilities to get some cash. Meaning, that just like the dot-com bubble created infrastructure for cheap, fast Internet, a potential AI bubble may create infrastructure, where

Running a multi-billion parameter model will cost an average Joe about 20 $/month

In the “AI-succeeds” world, that won’t happen. AI companies will hoard more and more GPUs, making ML and AI inaccessible to commoners. But in the “AI-busts” world, that’s suddenly a possibility.

So, if you are not in the AI industry and want to “short the AI bubble”, consider this. What would you do, which project would you build, if you knew that tomorrow you could get a rack of GPUs cheaper than your iPhone? Commodore has famously beaten the market and become a giant by betting that memory would become cheap in a few years¹. If you could make such a bet today, what would you build for your work, for your projects, for your industry?

Additional reading

• Is it a Bubble? - A well-balanced essay about whether AI is a bubble, and whether that would be a good or a bad thing. Written from an investor’s perspective, it touches many important subjects around the history of (tech) bubbles, “inflection bubbles” vs “mean-reversion bubbles”, loans vs VC, and how this all applies to AI today. Referenced above in “On AI “Bubble””.
• Ironies of Automation - A classic 1983 paper (reads like an essay) that points out that the more you automate, the more work you offload from workers… the higher the training, preparedness, and skill of the remaining workers should be and the more critical it becomes (think pilots who must take years of training to fly a mostly automated airplane + regular simulator re-trainings to keep their skills sharp facing the lack of practice).
• The optimal amount of fraud is non-zero - “You should welcome greater than zero fraud. You can think of it as a necessary expense. The reason is that policy choices impact the user experience of fraudsters and legitimate users alike. This argument generalizes, and it has some important ethical considerations.”