Data Matters

Fifteen years ago, OceanBase had just been born and proven its amazing ability during the 11.11 Global Shopping Festival. In those days, we spent every waking moment researching emerging technologies and envisioning an exciting new future.

Then, a particular quote left a deep impression on me. Marc Andreessen, pioneer in the IT industry and co-founder of Netscape, wrote in The Wall Street Journal that "financial services have been remade by software, and modern financial companies are software companies."

These past years feel like a stage being progressively illuminated. During the mobile internet era, we encountered massive, concurrent data for the first time and learned to record it with high fidelity. Later, leveraging this data alongside various big data technologies, we deepened our business analytics and witnessed the genuine unleashing of productivity from data. Today, in the era of artificial intelligence, multimodal data enables us to construct a digital mirror of the physical world, allowing us to examine the underlying logic of business within a much broader context. Finance has now evolved into an industry that is both data-intensive and model-driven.

Mission-Critical Modernization

Banking was traditionally anchored by physical branch networks and over-the-counter services characterized by high transaction values, low operational costs, and low latency sensitivity.

But with the mobile internet era, "online" became a non-negotiable imperative. The shift toward micro-transactions eroded margins per transaction, while soaring transaction volumes and stringent latency requirements drove costs higher. Classical architectures could not adequately address these challenges.

When scaling was needed, classical architectures could only scale vertically by purchasing more advanced mainframes rather than horizontally stacking commodity servers. Consequently, unexpected traffic spikes would rapidly push systems to their physical limits. Classical architectures also carried higher risks — risks that carry staggering costs within today's highly intricate financial systems.

ICBC (Asia): A Modernization Blueprint Under Hard Physical Constraints

ICBC (Asia)'s exploration of a partnership with OceanBase was aimed precisely at tackling these issues. At the time, while our system had already been validated across hundreds of clients, we were still newcomers to the Hong Kong and Macau markets. Among the more than 150 licensed banks in Hong Kong and Macau, not a single one had previously entrusted its mission-critical core banking operations to a self-developed Chinese distributed database.

In the end, the system upgrade enabled by OceanBase delivered a more than threefold performance boost for ICBC (Asia). Real estate in Hong Kong is notoriously scarce and costly, making it extremely difficult to establish ideal, physically and electrically isolated disaster-recovery zones. However, by deploying a cross-regulatory Shenzhen-Hong Kong dual-city architecture, OceanBase enabled ICBC (Asia) to leverage Shenzhen to implement a "two locations, three centers" topology. This dramatically bolstered their disaster-recovery resilience, slashed storage costs by more than 80%, and established a replicable modernization blueprint for other financial institutions.

Bank of Communications: Brain Surgery While Running a Marathon

Around that same period, Bank of Communications (BoCom) was also contemplating a comprehensive overhaul of its core systems. They placed high value on the scalability of distributed architectures, recognizing its vital importance in a fast-growing business environment. They decided to start by piloting OceanBase in their credit card business to evaluate the feasibility of replacing IBM mainframes.

I recall hearing for the first time back then that this type of project was colloquially known in the industry as "performing brain surgery while running a marathon." Intrigued, I looked into it and discovered that 70% of mainframe down-migration projects worldwide fail to meet expectations — the primary reason being sheer business complexity.

By any industry benchmark, BoCom's credit card system certainly fit the description of "extreme complexity." In terms of scale, it ranked near the top ten globally. Metrics like these might look like badges of honor from the outside, but for us at the time, they represented pure pressure. Across trillions in transaction volume, even the slightest error was something we could not afford. Crucially, cutover of the core system touched countless peripheral systems — a territory where no distributed database globally had substantial prior experience.

Ultimately, by completing the down-migration from IBM mainframes to a distributed architecture, we set a global precedent for a distributed system powering the core banking engine of a tier-one flagship bank. This endowed Bank of Communications with the foundational infrastructure capable of supporting its business growth for the next decade, while dramatically reducing the total cost of ownership (TCO) per account post-migration.

More importantly for us, we codified an engineering methodology that aligns technological exploration with mission-critical business stability. Subsequently, Bank of Communications migrated more of its core systems at its headquarters entirely to OceanBase — a monumental victory for distributed databases.

Four Requirements That Define a Modern Core

Whether modernizing classical architectures like ICBC (Asia) or handling extreme scenarios like Bank of Communications, drawing from Alipay's battle-tested experience and our work with hundreds of banks, we redefined the standards for core system modernization:

• The system must ensure continuous availability ("always-on") to meet the stringent business continuity demands of the mobile internet era.
• Across both transactional request processing and decision-making analytics, the system must operate in real time to respond instantaneously to user needs.
• Given modern application patterns and massive data volumes, the system must deliver flexible elasticity under extreme conditions while maintaining exceptional cost efficiency.
• The system must be natively distributed and service-oriented to empower agile innovation.

These technological advantages have earned us the trust of over 4,000 customers worldwide, including more than 150 banks, over 100 insurance companies, 150-plus asset management firms, and upwards of 100 fintech enterprises. The reason we have attracted so many financial clients is that they encountered virtually identical financial-scenario challenges to those faced early on by Alipay and Ant International — and our proven best practices aligned perfectly with their needs.

We often reflect with gratitude on the mobile internet era and the evolution of consumer habits, which laid the foundation for everything. It is with this same mindset that we now step into the era of AI.

AI for Innovation

Yet the landscape in the AI era is fundamentally different.

Statistics from Stanford University show that the core capabilities of large language models have long since caught up with human benchmarks, and OpenAI recently even welcomed everyone to the AGI era. However, McKinsey research reveals that outside of the IT sector, the adoption rate of AI in core business scenarios across most industries remains low — and in finance, it is even lower.

Is it because people are unwilling or do not know how to use AI? Clearly not. The real reason is that organizations dare not or cannot adopt AI, because their data simply isn't ready.

Data lies at the heart of at least three major bottlenecks. First, external models are growing more powerful, but they do not understand domain business logic — which resides exclusively within an enterprise's proprietary data. Second, while data assets are proliferating, data silos persist, leaving immense latent value untapped. Third, while autonomous AI agents have become the primary consumers of data, they remain far from genuinely unlocking real-world productivity.

Here is how we think about each of them, and what we have built in response.

Pillar One: Establish Business Semantics

Natural language data querying (ChatBI) is one of the most popular applications of AI in banking. But are these intelligent query systems truly intelligent? No. The greatest challenge is that financial systems demand absolute precision, whereas large language models are inherently probabilistic.

Take a simple prompt: "Check last week's loan disbursements for small and micro enterprises." The head office president, a branch manager, the head of risk management, and a frontline inclusive finance officer mean entirely different things when asking this question. Conventional technology fundamentally fails to grasp this nuance. Furthermore, the more departments involved, the longer the business chain, and the larger the data scale, the more pronounced the hallucinations and deviations become. Frontline staff asking about isolated details might get accurate answers, but when a bank president seeks to trace a lengthy business chain to pinpoint the root cause of an issue, the query output often goes completely astray. Therefore, simply translating natural language into SQL queries (Text-to-SQL) falls far short of solving the problem.

In collaboration with a major state-owned commercial bank, we explored an alternative approach. Rather than relying on simple similarity matching between questions and data, or leaning solely on manually constructed metric dictionaries, we introduced a semantic layer between the data and the user. This semantic layer establishes deep conceptual relevance to interpret intent, supply context, and answer inquiries. As a result, query response accuracy surged from 70% to 98%, enabling the system to support much longer execution chains, broader scopes, and deeper levels of inquiry and decision-making. Because this semantic layer is constructed semi-autonomously using LLMs, it captures more comprehensive information, is simpler to build, and continuously self-updates based on external data sources and user feedback — becoming smarter with use. Coupled with an enterprise-wide access control framework rooted in the database, data security concerns are comprehensively resolved.

In lending scenarios, this takes the form of a semantic ontology model that details the relationships between entities, their relevant attributes, and the actions that can be executed against them. This allows the ontology model not only to answer data queries, but also to serve as the decision-making engine and rationale for taking concrete actions. Only when semantics are accurately established can an enterprise's proprietary "dark knowledge" be made explicit, unlocking profound business insights. Superior technology leads to more precise decisions, faster operations, and lower risks. Establishing business semantics is therefore our first critical pillar in overcoming AI adoption hurdles.

Pillar Two: Process Multimodal Data as One

AI applications in finance are incredibly diverse. While the executive querying scenario illustrates how AI can better grasp the business context behind data to extract deeper insights, the financial sector also faces severe bottlenecks in processing complexity and efficiency stemming from multimodal data.

Take a mega property & casualty (P&C) insurer we work with: their existing multimodal repository exceeds tens of billions of records, growing by tens of millions — or even hundreds of millions — of new records daily. One facet of our collaboration involves identifying primary risk factors from mountains of claims data to optimize insurance products. Consider a real-world case: they needed to identify the top 10 intersections with the highest frequency of vehicle scrape accidents in a given region and year, determine the underlying causes, and partner with traffic authorities to optimize road traffic infrastructure — thereby curbing accident rates and reducing claims operational expenses. While highly valuable, this task was extraordinarily intricate. It required correlating and cross-validating an extensive array of information: accident liability reports, claims files, underwriting policies, road network maps, meteorological data, and traffic flow logs to successfully isolate these high-risk intersections.

Under legacy architectures, such data was fragmented across disparate silos. It required an army of data engineers utilizing different compute engines and writing custom operators — without native parallelism, resulting in low efficiency and exorbitant costs. The speed of data mining and insight generation lagged hopelessly behind daily data ingestion. This was not just a massive bottleneck, but a tremendous waste: highly informative data lay dormant as a mere cost center. Through our Lakebase architecture, data of any type can be stored in a unified manner without redundant copying or data movement, slashing storage footprints by a third on its own. By linking multimodal data natively, processing efficiency and accuracy are elevated significantly. Combined with a parallel computing engine where operators execute concurrently across massive datasets, tasks that previously took days are now completed in minutes.

The tangible business value delivered here was a simultaneous drop in accident frequency, claims payout costs, and the overall combined ratio for auto insurance. Blending unstructured data — such as contracts, certificates, audio, and video — with structured records like orders, customer profiles, and transaction logs is ubiquitous across financial services; any organization can unlock immense value here. Enabling integrated multimodal data processing is our second critical pillar in overcoming AI adoption challenges.

Pillar Three: Build the Data Foundation for Agents, Not Humans

While the first two challenges pertain to data consumption scenarios, the most profound shift is that the primary consumer of data has transitioned from humans to AI agents.

Agents broadly fall into two categories: consumer-facing (B2C) agents like ChatGPT or digital assistants (such as "Afu"), and enterprise-facing (B2B) agents acting as digital employees. As financial institutions deploy an increasing number of agents, three challenges inevitably arise. First, a massive population of agents inherently generates massive volumes of data. Second, multi-agent collaboration demands rigorous mechanisms for both data sharing and data isolation. Third, continuous agent self-evolution necessitates dynamic data mutability.

Behaviorally, users frequently spawn thousands of sub-agents for complex tasks without utilizing all of them. Most enter a dormant state, yet users expect them to be reactivated instantaneously on demand. This necessitates building a cost-effective data foundation at the logical level — a challenge we solved using logical databases.

Multi-agent collaboration has become the norm, where the linchpin of efficiency is sharing memory and context. In our experience, sharing distilled knowledge and memory across agents saves over 30% in token consumption on average. Improved comprehension also reduces conversational turns, lowering dialogue costs by 60% to 80%. However, this sharing must operate within a strict governance framework of access control. Because permission and identity management frameworks exist naturally as structured data within enterprise databases, context and memory retrieval can be governed seamlessly by existing access controls — making security both organic and highly efficient.

Furthermore, agents must continuously evolve. The cornerstone of evolution is logging execution trajectories to identify failure cases ("bad cases"), evaluating performance, and iterating. Contrary to common perception, this lies at the core of agent harness engineering. However, rigorous benchmarking requires real-world data, creating an "impossible trinity" among data authenticity, isolation, and testing efficiency. Testing directly against live production data risks contaminating clean records with transient artifacts; duplicating an entire production dataset for every test cycle incurs prohibitive data-prep lead times. To enable rapid, effective evaluation, we engineered a sandbox technology that instantaneously spins up countless isolated sandboxes sharing identical data baselines. By executing distinct feature tests across separate sandboxes concurrently, evaluation is transformed into a massively parallelized series of controlled experiments — vastly improving evaluation velocity. We believe these engineering foundations will underpin the new business models now taking shape across financial services, catalyzing a new wave of rapid evolution.

One Keyword: Unified

Whether as clients, industry partners, or technology enthusiasts, we all look forward to this ongoing evolution. I want to close by sharing our data engineering methodology, with the hope that it proves valuable to your own endeavors.

This methodology centers on a single keyword: "Unified" — bridging all existing silos, whether across disparate data modalities, heterogeneous compute paradigms, or platforms bound by different rules. As you will discover, real business value thrives precisely within the yet-unforged connections between these data domains.

Roughly twenty years ago, we integrated diverse capabilities into a single device, ushering in the smartphone era and completely transforming our daily lives. Now, by unifying all data, we will unlock the era of true business intelligence — fundamentally redefining our enterprises.

The journey over these years has taught us a fundamental truth: whenever business scaling encounters a bottleneck, the ultimate solution invariably hinges on data. The physical media of data storage, the paradigms of data analytics, and the principal actors utilizing that data will continue to reshape industries and propel successive waves of innovation.

As a data technology company, OceanBase stands ready to empower your innovation, always.