Of course, Google could be wrong about the economics of running old AI processors. Or maybe TPUs are special.
This is a perfect circle of reasoning. The bubble exists because people say the bubble exists and reject data centers, therefore creating the bubble?
This hasn't historically been true. The clusters have been getting bigger along with the models over time.
Not if the datacenters are in space, you can upgrade the models for a little while but once you need to upgrade the silicon you need to launch new hardware.
The companies that sold core infra during dot com mostly died, because they invested in fiber etc far too quickly. Likewise with server factories like Sun.
The B2B and B2C survivors were patchy. A few did extremely well but there were a lot of casualties.
The biggest winners offered B2G - business to government - disguised as search and social media. They were actually selling individual user profiling and targeted ads, and that turned out to be the biggest money maker.
The reason Anthropic and OpenAI appear wildly overvalued is because their offering isn't AI in some abstract sense, it's the next generation of surveillance and persuasion technology. But now with corporate surveillance in addition to the traditional profiling and targeting.
You can tell a lot about what someone is doing - and the industry is doing in general - by looking at everyone's code bases. You don't need to train on them or steal keys to extract immense value.
Potentially there's also the option to eventually make claims on corporate IP developed with AI tech.
Anyone who thinks the companies who pulped and torrented whole libraries of books for their training data would never, ever, pinky promise, try to extract value from the code they're generating is being very naive.
AI models can be deployed on the same datacenter as the old one and be either orders of magnitude cheaper to run, or orders of magnitude more capable.
The only thing that can pop this bubble is if we're not smart enough.