The financial industry has always been a testing ground for new technologies. What are the latest advancements in large models and independent innovation this year?

The financially robust industry has always been a crucible for new technologies. Despite budget cuts at many financial institutions this year, there are still numerous new developments in large models and independent innovation.

01

Leading banks cautiously advancing large models

At the China International Finance Expo held in Beijing in July, many industry insiders came to explore and exchange ideas on large model applications. Interestingly, since May this year, more banks have launched large models, and employees have begun interacting with various assistants. However, large models are still in the early stages of experimentation in the financial industry, and their adoption speed and depth are even slower than those in manufacturing.

Industry insiders describe the situation in finance as conducting large-scale research, small-scale validation, and testing on large models. The primary applications are tool-like assistants rather than transaction production operations.

"They haven't been widely adopted yet," a leading financial application developer told数智前线, illustrating their standard with an example: "If someone says, 'I want to open an account,' it involves compliance oversight, data reporting, and multiple business departments and approval procedures in the process. These are implicit mechanisms. A truly comprehensive large model would need to break down all these requirements and business processes clearly, which is not yet feasible."

"It's not so much that it can't be done as that there's hesitancy. Currently, using AI for actual transaction operations is a bit daunting," a city commercial bank representative told数智前线.

In fact, leading banks began researching generative AI even before ChatGPT gained popularity. However, due to the accuracy issues and lack of interpretability of large models, the meticulous financial sector requires continuous experimentation and observation. Simultaneously, the financial industry has yet to identify sufficient pain points to achieve higher value, and the scale is insufficient.

Under these circumstances, leading banks exhibit a cautious and pragmatic attitude. "Many traditional methods, like keyword searches, are already highly efficient," an ICBC representative told数智前线. "We first need to address business issues. If the business doesn't require it, forcing it in can actually reduce efficiency." He emphasized that technology should not be pursued for technology's sake. ICBC's primary internal applications currently include code assistants, customer service, and risk control.

Bank of China uses large models for data analysis, among other things. A source told数智前线 that internal progress has not been swift, and risk control is not yet integrated.

A CCB representative said she saw the internal launch of a large model in May. She uses it to help write customized research reports for corporate banking services. "These reports are not yet highly targeted and require revisions to about half the content, but moving forward, this is a trend," said the young professional, who expressed avid interest in large models and had personally tried them out since the early days of ChatGPT and ERNIE Bot.

Currently, large models in the financial industry are primarily applied to code development, operations and maintenance, interactive data analysis, marketing tools, insurance quotes, etc. Due to limited application scenarios, the head of a large model manufacturer remarked, "I didn't expect code assistants to become so prominent in banks."

Apart from leading banks, some city commercial banks or joint-venture banks are moving faster than their larger counterparts in the financial system. "Investing a few million yuan can solve a practical problem."

Notably, the base models or industry-specific large models used by financial enterprises for training their large models are diverse. Commercial options include Huawei Pangu and Zhipu, while open-source alternatives like Alibaba Tongyi, Zhipu, Baichuan, and foreign players like LLMa are available. However, they admit that no industry-specific large model stands out. Additionally, some bank staff rely on word-of-mouth recommendations and use tools like Kimi as document assistants.

"The biggest problem for banks is that their data cannot be shared externally, limiting their ability to judge user behavior solely based on their own data," said a representative from Hangzhou Bank. Therefore, many banks invest heavily in data collaboration.

A representative from Hangzhou Bank said they partnered with Zhejiang University, focusing on foundational data training. "The model can be open-source, but data is challenging, costing nearly ten million yuan." Current applications concentrate on assistants like HangXiaoZhu, code assistants, and contract reviews.

Apart from data, most banks first purchase computing power. "Smart computing began around 2022," said several sources. Due to NVIDIA's intermittent bans on various products, supply instability prompted financial enterprises to build heterogeneous computing platforms in recent years.

Overall, the widespread adoption of large models in the financial industry is still some time away. Apart from the 21 leading state-owned and joint-stock banks, many city commercial banks and insurance companies are also experimenting. A city commercial bank representative told数智前线 that they are conducting research and preparing for a small-scale pilot program.

"The direction is clear," said a financial application developer. "It's just a matter of when large-scale adoption will occur, and everyone wants to know the answer."

02

Core system independent innovation in the crucial stage

Apart from large models, the technological strategies of financial enterprises have entered deeper waters, marked by the transformation and reshaping of traditional technological architectures into new systems vastly different from "IOE."

Overhauling the core system is no easy feat, involving full-stack capabilities from operating systems, databases, middleware to hardware: all technologies must meet performance and stability requirements; regardless of being open-source or closed-source, they must undergo large-scale commercial validation; and these technologies are not simple replacements but aim to fulfill the high demands of the new generation of businesses. Database replacement is particularly illustrative.

In this regard, leading banks are moving faster. Many banks started with peripheral systems and are now in the crucial stage of overhauling their core systems. "The database has not been fully replaced yet, and there will be a parallel running stage, so it may seem like 70% to 80% is done, but the cycle is still long," said an industry insider. Full core replacement might be achieved in the next one or two years, while joint-stock banks are slower than leading banks.

Apart from leading and joint-stock banks, city and rural commercial banks are more scattered, with some local city commercial banks progressing faster and having already revamped their core systems. "There's a trend towards deeper penetration, with smaller banks starting to overhaul their core systems."

During this process, the database market is also being reshaped. Although there are over 300 domestic database enterprises, "I believe the database market will further consolidate around a few vendors," predicted a database enterprise.

Regarding database replacement, some in the industry have raised concerns about the high replacement costs and usability issues of these new architecture-based databases. However, a city commercial bank representative told数智前线 categorically, "Our motivation for exploring replacement in 2015 was cost. Traditional centralized architectures are expensive in the long run, with high maintenance, hardware, software, and personnel costs. Replacement definitely saves costs; otherwise, we wouldn't have pursued it. I think we should consider costs in the long term."

From a technical complexity perspective, with the progress of domestic databases, the difficulty of platform migration is less significant. However, revamping existing bank applications poses challenges and complexities due to the need to ensure data consistency while maintaining high performance and multi-activity, which requires substantial capabilities.

In terms of usability, "I don't think there's much difference; it can ensure continuous and stable production operations," said a responsible person from a city commercial bank. The main differences lie in ecosystem and service. For instance, domestic database technology dissemination and application talent supply are insufficient. Furthermore, in terms of service systems, users of IBM, for example, can automate problem resolution with a single phone call, whereas domestic databases often require reaching out to multiple parties.

"If there's a clear indicator that we've caught up, it's when I can solve all problems with a phone call without meeting these database enterprise representatives," the aforementioned person told数智前线, emphasizing the need for a comprehensive system and knowledge base.

In the financial industry, database enterprises share the market, with many financial institutions mentioning Ant Group's OceanBase, Tencent's TDSQL, ZTE's GoldenDB, and Huawei's GaussDB, unlike the operating system sector, where Kylin dominates.

"These companies adopt vastly different strategies," said an industry insider. Ant Group's OceanBase entered the market early and holds a high market share in city commercial banks, rural commercial banks, insurance, and securities. Huawei aggressively targets the market with its full-stack products, including GaussDB. ZTE also offers a full stack, including operating systems, databases, and servers, focusing its efforts on bank core systems. Tencent's TDSQL pursues a comprehensive market strategy, offering both on-premises deployment and public cloud services to both leading banks and small-to-medium financial institutions.

Beyond software, core systems also require high-performance computing, networking, and storage, such as FC fibre channel storage. Currently, the industry is in an innovative research phase for fibre channel storage. Due to the full-stack capabilities required for core systems, companies like Huawei, ZTE, and H3C leverage their full-stack layouts to aggressively target the market.

03

Diverse hardware applications

Hardware innovation began with financial terminals. Interestingly, while terminals using the Kylin system previously dominated the market, HarmonyOS terminals have gained momentum since the second half of last year. At the China International Finance Expo, one could see HarmonyOS self-service devices piloted by Hangzhou Bank.

In computers, Lenovo holds about three-quarters of the financial market share, with others including Huawei and Tongfang. Lenovo showcased locally sourced computers, such as the high-end Lenovo Kaitian X1, featuring a carbon fiber body, weighing 990 grams, with an eight-hour battery life, tailored for bank management.

Bank customers have customized requirements for computers, including pre-configured usernames, passwords, and browser bookmarks for lightweight customization. Deeper customization involves compatibility with peripherals, pre-optimization and testing of Linux systems based on customer scenarios.

An industry insider said banks have invested significantly in resolving various issues during computer application. "We address hundreds of software, peripheral, and other issues every quarter, resolving them within days," he said. "The number of issues will逐渐减少, and other enterprises can reuse the solutions." A new round of computer procurement is approaching in the second half of the year.

Regarding smart computing, due to the accelerated development of AI, leading financial enterprises invested in smart computing infrastructure around 2022.

According to government requirements, recently, the National Development and Reform Commission, Ministry of Industry and Information Technology, National Energy Administration, and National Data Administration jointly issued the "Special Action Plan for Green and Low-Carbon Development of Data Centers," mandating that by the end of 2025, the Power Usage Effectiveness (PUE) of newly built and expanded large and super-large data centers should be reduced to below 1.25, and national hub node data center projects' PUE must not exceed 1.2. Shanghai even has regulations for real-time PUE monitoring of data centers.

Policies are forcing the upgrading of existing data centers. According to relevant statistics, the PUE of advanced green data centers in the industry has dropped to around 1.1. Liquid cooling products are becoming a trend, with urgent market demand, prompting suppliers to launch flexible upgrade solutions. For instance, by adding complementary liquid cooling equipment to an existing 3-4 kW cabinet under unchanged floor space and load-bearing conditions, the capacity can be increased to 20 kW.

From a hardware platform perspective, the market presents a diverse landscape, generally meeting market demands. The transformation of application systems exhibits a trend of old and new iterations.