Big AI is a bubble but AI in general is not.

If anything, the DRAM shortages will apply pressure on researchers to come up with more efficient AI models rather than more efficient (normal) software overall.

I suspect that as more software gets AI-assisted development we’ll actually see less efficient software but eventually, more efficient as adoption of AI coding assist becomes more mature (and probably more formalized/automated).

I say this because of experience: If you ask an LLM to write something for you it often does a terrible job with efficiency. However, if you ask it to analyze an existing code base to make it more efficient, it often does a great job. The dichotomy is due to the nature of AI prompting: It works best if you only give it one thing to do at a time.

In theory, if AI code assist becomes more mature and formalized, the “optimize this” step will likely be built-in, rather than something the developer has to ask for after the fact.