At its 2025 annual general meeting of shareholders held today, China Merchants Bank Co.,Ltd. (CMB) addressed the costs and benefits associated with large language model programming. The bank's Chief Information Officer, Zhou Tianhong, discussed the widely noted case study of Uber, highlighting that the core issue it revealed was excessive, over-budget token consumption. He clarified that this high consumption was primarily in the area of using large models to generate code, rather than Uber's business units applying the models for operational management purposes.

Zhou stated that the emergence of large language models has transformed software development paradigms. While their use in programming currently plays a significant role, substantial work remains in this direction, and several fundamental challenges persist. Beyond cost concerns, he identified two key technical issues with code generation by large models: firstly, the architecture of large-scale software remains critically important and is currently a relative weakness; secondly, even when functional correctness rates are high, model-generated code can still introduce various technical problems. These include unreadable "spaghetti code," performance bottlenecks, and security vulnerabilities—issues that are not yet fully resolved by current large model programming capabilities.

Regarding these challenges, leading U.S. companies are adopting a more measured perspective. Zhou pointed out that besides Uber, major American firms like Microsoft and Amazon have internal policies aimed at maximizing token utilization, a concept often termed "Tokenmaxxing." However, he noted that these companies are now collectively re-evaluating this approach and beginning to make adjustments.

From CMB's own standpoint, Zhou explained that the bank is already employing large model programming technology for internal software development, adhering to a strategy of "active follow-up with prudent application." In terms of computing power allocation, the resources dedicated to large model programming currently constitute only about 5% of the bank's total computing capacity. The token usage figures disclosed during this year's earnings presentation primarily reflect consumption by business departments, with the daily average reaching 33 billion by the end of May.