DBS Bank, Southeast Asia's largest bank by assets, is adopting a strategic approach to AI deployment, emphasizing cost management and efficiency. While the cost of individual AI tokens has reportedly halved between 2024 and 2025, overall AI spending by companies continues to climb due to the technology's widespread adoption. DBS runs its AI workloads on a private cloud with elastic capacity, managing token costs by holding memory at the application level and passing only incremental context to models. This strategy helped the bank unlock approximately S$1 billion in economic value during the year from over 2,000 models across more than 430 use cases, as stated in its 2025 annual report.
Nimish Panchmatia, DBS's chief data and transformation officer, dismisses "tokenomics" as overhyped, viewing tokens as another cost alongside traditional expenses like electricity and rent. To keep consumption in check, capabilities are rationed by role; for instance, video generation is limited to the marketing department, which manages it within a budget similar to advertising agency spending. Panchmatia noted that the bank's architecture was designed to be model-agnostic, anticipating the commoditization of large language models and focusing on the "harness" or software scaffolding above the model as the critical component.
DBS is cautious about fully autonomous AI agents, preferring a human-in-the-loop approach with checks between stages. Panchmatia stresses the need to understand AI operations, stating, "You've got to know what's going on." The bank also limits the number of steps in agent chains to prevent unexplainability and large "blast zones" in case of failure. This reflects a broader trend among Singaporean companies, which are moving beyond experimentation to demand clearer returns on investment amid rising AI bills, as they match models more carefully to workloads.
