In the year when AI databases were released in droves, what did products such as the Electric Power Golden Warehouse KES AI upgrade?

AI-driven industrial upgrading has spread to the database industry this year. In January, Alibaba Cloud's PolarDB released capabilities including the AI data lakehouse (Lakebase), model operatorization, and Agent hosting, pushing databases from "AI-ready" to "AI-native"; in July it launched PolarDB Agentic Database for Agent scenarios;

In June, OceanBase packaged its AI direction into the OceanBase AI Database product portfolio; at the end of October, CETC Kingbase also officially released the KES AI database, emphasizing putting AI capabilities into the database kernel and building three core capabilities: in-database semantic construction, in-database computational inference, and in-database native data security control, so that data does not leave the database while still supporting the efficient and secure implementation of the full AI application chain...

Databases welcome new users

From Oracle to distributed clusters, from standalone machines to cloud-native, database technology has evolved for more than forty years, managing the data humans generate and use in good order. But after the emergence of generative AI, the primary users of databases are changing from humans to AI. They do not read reports or call the DBA; they only recognize context they can directly consume. Thus, a question that had never been seriously discussed in the past thirty years suddenly became a mandatory question for the entire industry: can machines "understand" the data stored in databases? As part of the AI foundation, databases have also entered an era of change.

To understand why database reform is happening now, one must first see clearly where large models get stuck when entering enterprises' core businesses. On this point, CETC Kingbase President Du Sheng said that databases require fast thinking and are responsible for providing a trustworthy, persistent, and computable data foundation. Large models, by contrast, are slow thinking, responsible for understanding, reasoning, and generation; the data of fast thinking is the fuel of slow thinking, and slow thinking needs fast thinking to continuously feed it data.

According to industry statistics, more than 83% of public data has been fairly fully developed, while more than 95% of enterprises' high-value business data still lies dormant in private-domain databases, and more than 80% of that is unstructured content. In other words, enterprises' most valuable data is "invisible, unreadable, and unusable" to large models.

The problem is not insufficient storage capacity, but that data lacks semantic description. Databases contain only table structures and field names, without business meaning. For example, a field in a database called customer may refer to a signed customer or a prospective customer; a field called amount means completely different things in the eyes of the finance, sales, and risk control departments. Models cannot determine this on their own, so they either guess blindly or answer incorrectly.

English text is being prepared.

The semantic models built at the upper layer often do not know when the table structures of the lower-layer transaction systems are changed or when field definitions are adjusted; synchronization links break at any time, and maintenance costs rise exponentially with the number of systems. The more hidden cost is organizational: lower-layer business systems change frequently without notifying the upper-layer governance department, while the governance department has no authority to control the various business systems, and the freshness of semantics is consumed among departmental walls.

In the AI era, the semantic problem has been amplified by AI. Industry research shows that among 236 analytical intents covering finance, human resources, and public security, successfully executed SQL queries contained 153 recurring "operationalization failures"—the query syntax was correct and results were returned, but the thresholds, time operators, or derived metrics it selected did not faithfully capture the analytical concepts in the user's intent. For example, "abnormally high-risk loans" are operationalized with a fixed threshold, whereas "abnormal" is essentially a comparative judgment relative to a reference group. Such failures do not report errors, but they cause enterprises to make wrong decisions based on a number that looks right.

Speaking about the importance of the semantic layer to databases today, Du Sheng offered his own judgment. He told us, "Semantics are the core foundation of AI architecture. Many questions that models answer incorrectly are not because the model cannot reason, but because the data it receives itself has no semantics."

The importance of the semantic layer is not just one company's claim. In multiple 2026 studies, Gartner repeatedly emphasized the same judgment—context is becoming new critical infrastructure—and predicted that by 2030, a universal semantic layer will rank alongside data platforms and cybersecurity as critical infrastructure. Put more plainly: in the past, enterprises bought databases to store and compute data; in the future, enterprises will also have to pay for the layer of capability that "lets machines understand data." AI projects that only connect to large model interfaces but lack a data semantic layer will most likely be unable to enter production environments.

On this basis, the path proposed by China Electronics Technology Kingbase is to move semantic addresses from outside the database to inside it. Rather than building another, smarter metadata platform, it lets semantic objects and physical tables coexist within the same kernel. When the physical structure changes, the associated semantic objects become invalid like stored procedures and trigger recompilation or manual confirmation; the change is pushed in front of business users instead of disappearing into cross-system synchronization pipelines. More critically, the direction of construction is reversed: semantics no longer pursue a one-time, top-down definition for the entire enterprise, but instead grow bottom-up with each business database as a unit. Those who truly understand what each field represents in the business system. Hundreds of databases each solidify their local semantics, and large models then extract them in bulk and aggregate them into enterprise-level semantics—a slower path, but one that is not easily broken.

Data not leaving the database remains the bottom line

The traditional approach is to export data from the database, send it to an external system for vectorization and retrieval, and then transmit it back to the application side. For one thing, data is repeatedly transmitted and moved across systems, so data movement overhead is high and computational efficiency cannot improve; moreover, once data leaves the database, the original permission controls, field-level masking, and audit capabilities are broken. For industries such as finance, energy, and government affairs, this is not just an efficiency issue but also a question of whether compliance justification and audit costs are acceptable—which is also why many AI projects remain stuck in pilots for a long time and cannot enter production environments.

"Data not leaving the database is not a technical choice but a compliance bottom line." This is Du Sheng's summary of data not leaving the database.—The in-database closed-loop model, while holding the compliance red line, achieves a semantic closed loop, reduces data movement loss, and improves the efficiency of the entire AI chain.

Against this background, the logic of "AI capabilities growing inside the database" becomes clear: vectors are generated inside the database, semantics are defined inside the database, and inference is completed inside the database under the ONNX framework. The database's original row-level permissions, field-level masking, and behavior auditing cover the AI chain; when a large-model Agent accesses the database, it reuses the permission system of the database account and cannot cross the account's fence.

English text is being prepared.

In the AI era, in addition to ensuring data does not leave the database, multi-model fusion is the most critical capability. The data forms faced by AI applications are no longer neat two-dimensional tables; documents, images, audio and video, graphs, and time-series streams appear simultaneously in a single agent workflow. In the past, enterprises mostly adopted dedicated databases for dedicated purposes: relational databases for transactions, vector databases for vector retrieval, and graph databases for association analysis. Behind one application were several engines strung together, with data repeatedly moved between databases.

Today, converged databases have moved from a bonus feature to a mainstream option for large-scale deployment, and the window for standalone vector engines is narrowing. Faced with the fork in the road between replacement and coexistence, Du Sheng's judgment is that replacement and coexistence proceed in parallel. "What the vast majority of ordinary customers want is one database that handles multiple types of capabilities, reducing DBA operations and learning costs. Only in extreme scenarios such as pure-graph ultra-large-scale social networks do standalone dedicated databases still have value for retention. The main battlefield remains relational databases; multi-model is a capability layered on top, achieving integrated fusion rather than starting from scratch." Du Sheng emphasized.

According to estimates in the CAICT "Database Development Research Report (2026)," China's database market size in 2025 was about US$9.49 billion (about RMB 67.8 billion), accounting for 7.2% of the global market; by 2028, the total size of China's database market is expected to approach RMB 98 billion.

Data released by IDC at the 2026 Bund Conference shows that in 2025, China's AI data infrastructure market size was about US$14.9 billion, with a compound annual growth rate of 37.7% from 2025 to 2030, reaching US$73.8 billion in 2030 and expected to reach about RMB 1 trillion by 2035, more than 10 times the 2025 level.

The traditional database market is growing steadily, but the real increment is on the AI side. IDC predicts that China's data generation volume will increase from about 76ZB in 2025 to about 147ZB in 2030, of which about 90% is unstructured data. Large models do not lack data; what they lack is the infrastructure that can "feed the data in."