How to use computing power faster: on the weird economics of semiconductors and GenAI
The author posits that increasing sales of ever more powerful processors requires a market need for more powerful processors. During Microsoft's heyday, that was each version of Windows or Office requiring more computing power (the author thinks mostly because of bloat). The proliferation of smartphones, then crypto, then AI have continued the trend. So the semiconductor industry has become dependent on AI.
Unless a new fad emerges, given the supposed unprofitability of AI, the author sees a few future possibilities:
- AI companies become much more efficient, using much less computing power. This would lead to a bust in the semiconductor industry.
- AI companies can't become efficient enough to become profitable, leading to an AI bust that also brings a bust in the semiconductor industry.
- AI increasing efficiency just enough to become profitable while not harming the semiconductor industry too much. This is the "inference inefficiency optimum."