- LLMs have broken economics because their token-based costs are metered and unpredictable, unlike typical software subscriptions, and users often pay for wasted tokens.
- AI companies like OpenAI and Anthropic are unprofitable; they subsidize consumer subscriptions by giving away far more tokens than the subscription cost, and enterprise token billing has led to budget overruns (e.g., Uber).
- Most AI services are not differentiated or reliably useful, leading to low adoption and revenue concentration at OpenAI and Anthropic, with many startups unprofitable.
- AI data center construction is extremely expensive and debt-financed, but demand is mostly from unprofitable AI companies, risking a collapse that could hit pension funds and private credit.
- The AI buildout has driven up hardware costs (e.g., doubled memory prices), which Apple has passed to consumers, but Apple itself has spent little on AI infrastructure and relies on renting Gemini from Google.
- Apple Intelligence was poorly received, leading Apple to slow its AI investments; if the bubble deflates, Apple will likely watch from the sidelines and may make selective acquisitions.
- Ed Zitron suggests Apple should focus on improving the Vision Pro rather than chasing AI hype, as the bubble stems from a lack of new interface ideas.