Tuhin Srivastava explains that over 95% of tokens served on Baseten come from custom, post-trained models, not vanilla open-source ones, highlighting the strategic link between inference and post-training. He advises customers to prove product-market fit with frontier models before investing in custom solutions. The discussion addresses severe GPU supply crunches, with Baseten operating across 18 clouds and 90 clusters, and notes that capacity now requires 3-5 year contracts with significant prepayment. Srivastava emphasizes that software, not hardware access, creates market stickiness, and predicts a future of inference-specific chips. He also shares operational lessons, including the importance of hiring leaders who own entire problems and maintaining a high-stakes on-call culture. The episode concludes with the Jevons paradox in AI, where lower inference costs drive increased demand, and a vision of personalized AI concierge services transforming industries.