Where AI money is made, and where SEA founders should actually compete

Over the past two years, revenue in AI server assembly and GPU rental grew by 106 to 346 percent year on year. Dell, Supermicro, and CoreWeave reported operating margins of 11 percent, 8 percent, and 6 percent, respectively. Growth and value capture are decoupled, with high margins at the scarce bottom layer and low margins in the commoditizing middle. The top layer remains an unproven but potentially valuable space.

Picture the AI stack as six layers: power and data centers, chips, cloud, models, execution and tooling, and applications. High value accumulates at the scarce bottom, declines in the middle, and rises again at the locked-in top. Semiconductors received 75 percent of the AI ecosystem's revenue, while energy constraints limit margins. Models exhibit a 100x price spread but deliver limited quality improvements. Application-layer margins are thin, ranging from 0 to 30 percent, against software margins of 80 to 90 percent.

The thesis suggests that value migration is currently occurring at the thinnest-margin, most concentrated, and least proven layer of the stack. Cloud's journey from infrastructure-heavy to application-heavy took about a decade, and AI in 2026 resembles that timeline. Massive capex, thin margins at the bottom, and an uncertain application layer characterize AI's current state.

Southeast Asia presents unique opportunities for AI founders. The bottom layer is closed to entrants due to capital intensity, while the middle is dominated by research-heavy laboratories. The top layer offers a rare chance for seed investments, as the region has leapedfrogged SaaS and possesses engineering talent in countries like Vietnam, Indonesia, and the Philippines.

However, founders should select application layers with strong moats and avoid building features similar to those offered by AWS or the next model release.