Clouded Judgement • 12 implied HN points • 13 Mar 26
- Model labs can reach high, sustainable gross margins as they scale because serving and architecture improvements, better GPU utilization, and product optimizations drive down inference cost per token.
- Training costs are likely paybackable within reasonable timeframes similar to CAC payback, and even though retraining is recurring, marginal gross profit after payback can make labs profitable.
- Platform lock-in and enterprise needs (fine-tuning, SLAs, tooling, context storage) raise switching costs, so open-source models won’t fully commoditize large customers and retention should stay high.