CMB Details Large Model Ledger: Burns 33 Billion Tokens Daily, Emphasizes “Invest 20 to Earn 100”

```

On June 25, China Merchants Bank held a shareholders’ meeting. Amid the banking industry’s exploration of implementing artificial intelligence, relevant person in charge Zhou Tianhong answered questions at the meeting regarding the input and output, business empowerment, and strategic considerations of large model technology.

Zhou Tianhong regards this round of technological breakthroughs as “on the level of an industrial revolution,” but he also pointed out that, as with historical technological advances such as steam engines and internal combustion engines, there is a process from the emergence of the technology to actual societal changes.

Currently, the technology has already produced clear business effects within CMB:

Zhou Tianhong shared a set of data: the ratio of AI-handled operation hours to actual input by human employees at CMB has risen from 1:13 at the end of last year to nearly 1:9 by the end of May this year;

Among them, some application scenarios in the retail sector have shown particularly evident effects.

With the rapid development of foundation model and agent technologies this year, CMB is continuously analyzing the application potential of these technologies in various business fields;

Considering the banking industry’s strict regulation and high demands for rigor in operations, relevant applications are kept under ongoing scrutiny and improvement.

In a company with over 100,000 employees, preventing “high input, low output” is the top concern for management when promoting large models across the board.

Zhou Tianhong revealed that before proposing the “AI First” strategy, CMB had already established a relatively complete cost-benefit measurement system in its technology lines.

Under this system, the cost side is broken down into R&D staff investment and Token expenses; the revenue side has a measurement system with six dimensions.

According to Zhou Tianhong’s calculations, CMB’s current cost-to-benefit ratio in large model direction remains around 20%, meaning “an investment of 20 yuan can create a return of 100 yuan.”

Although cost data can be precisely calculated, measuring returns with precision remains a complex task in the banking industry with its messy business value chains.

Since the beginning of this year, CMB has been further scrutinizing algorithms, data sources, and data quality for measuring returns. Supporting this exploration is CMB’s annual technology investment of around 13 billion RMB. The share of computing power purchases is not high within the overall budget, so there is still room for additional investment in AI computing power going forward.

In terms of specific resource allocation, CMB exhibits a differentiated strategy, keeping a cautious approach regarding large model code writing.

Zhou Tianhong cited the recent much-discussed Uber incident, noting that over-budget Token consumption often occurs during code writing, rather than actual management application in business departments;

He believes that although large models have changed software development patterns, there are still fundamental technical obstacles at this stage: large models perform poorly when handling large software architectures and tend to generate hard-to-read “spaghetti code,” leading to unresolved performance and security issues.

Overseas tech companies like Microsoft and Amazon, which previously promoted internal “Tokenmaxxing” policies (maximizing Token use), are also currently reflecting and adjusting their strategies.

Based on this trend, Zhou Tianhong made clear that CMB adopts the strategy of “active follow-up, cautious application” in internal software development.

Currently, the computing power invested in programming large models at CMB accounts for only 5% of total computing power.

Compared to the restraint in code development, the application of large models on the business side is rapidly increasing. Zhou Tianhong revealed that by the end of May, Token daily consumption mainly from business departments had reached 33 billion.

Having clarified the boundaries of investment and focus of application, CMB’s long-term goals remain unchanged.

Zhou Tianhong stated that CMB has a clear vision, which is “to build an intelligent bank.” In achieving this goal, relevant supporting metrics and management systems are still being constantly refined and deepened.

Risk Warning and Disclaimer ClauseThe market involves risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions herein are suitable for their particular circumstances. Investment based on this is at one’s own risk. ```