Hello Betamax, With traditional SaaS, the cost of serving an extra user is predictable and easy to control. Companies mostly pay for additional server compute and storage costs, which scale incrementally. It's a different story with AI apps. One paying subscriber might barely touch their plan, leading to unused funds. Meanwhile, someone on the free tier could spend hours prompting the AI and burning through tokens, which still necessitates the company to pay large sums to the model provider. The only way to escape this trap, argues Bruce Yang, is to own the models yourself. That's what his company, Sapiens AI, has spent the past year building toward. Our top story today dives into how the Singapore startup trained a fleet of around 20 proprietary models to power its Agnes AI platform, and why Yang believes application-layer companies that rely entirely on third-party models are structurally disadvantaged. He's also making some bold bets: US$100 million in ARR and a potential Singapore Exchange listing by end-2026, all while the company is still unprofitable. Whether Sapiens can pull it off is an open question. But the logic behind its approach is worth understanding, especially as more AI startups in the region wrestle with the same cost problem. It's an issue they have to address if they want to hit profitability. As our data story shows, AI companies in the black are still a rare breed in Southeast Asia, making up only four out of the over 60 profitable tech firms we've tracked across the region. The startups that figure out how to control their unit economics early will likely be the ones that make it onto that list in the future. Glenn Kaonang, journalist | 
