The Future of Databases: From DB4AI to DB4DATA | GBASE Tech Webinar Featured Talks · Bai Shan

At the 2026 GBASE Tech Webinar, Bai Shan (Xu Ji), Co-founder of Baisheng Intelligent Computing and founder of the DBAIOPS community, was invited to deliver a keynote speech titled The Future of Databases: From DB4AI to DB4DATA. Starting from the historical patterns of database technology evolution, he deeply analyzed the new challenges AI brings to data processing and painted a forward-looking picture of databases evolving from a “data warehouse” to a “data intelligence engine.”

Historical Pattern of Database Evolution: Always Keeping Pace with Data Processing Needs

Bai Shan began by reviewing the evolution of databases. He pointed out that data processing has progressed from manual management to record management systems, then to databases, data warehouses, big data, and today’s AI data factories. Database technology similarly evolved from hierarchical and network models to relational databases, then to NoSQL, NewSQL, and today’s AI-native databases.

“Database technology evolution has always followed data processing needs” — this is the core principle throughout the speech. Relational databases have dominated for over forty years precisely because their ACID transactions, SQL standardization, and B+Tree indexes accurately met the core needs of enterprise data management. Today, the AI era is redefining the boundaries of data processing requirements, and databases must evolve accordingly.

New Data Processing Demands in the AI Era

Bai Shan identified three major new challenges for data processing in the AI era:

Exponential Data Growth

From TB to PB to EB and ZB, data scales continue to leap. For example, training LLaMA-3 requires over ten trillion tokens. China’s data production reached 8.1 ZB in 2022, with an annual growth of 22.7%, indicating an unprecedented scale.

Multimodal Data Becomes Mainstream

Data types have expanded from single text to images, videos, audio, graph structures, time-series data, and even embodied intelligence data. The traditional relational database design centered on structured data inherently has limitations when dealing with multimodal data.

Extremely High Data Quality Requirements

The AI era has shifted from 'just having data is enough' to 'must have high-quality data.' Data cleaning, professional annotation, and data governance have become key cost centers.

Meanwhile, real-time requirements are also dramatically increasing — the long-standing separation between online transaction processing and offline analysis is being broken, giving rise to HTAP (Hybrid Transactional/Analytical Processing). Vector capabilities have become a fundamental database requirement, and the combination of vector databases and large models has become a core capability for RAG.

Bai Shan analyzed: “Traditional RDBMS lacks vector indexing and multimodal data support, has limited HTAP capabilities, and its data governance systems are not intelligent enough. The fundamental problem is that the traditional relational database kernel cannot perform reasoning using multiple modalities of data simultaneously.” The conclusion is clear: Databases must evolve, and AI-native databases are on the way.

Eight Core Capabilities of Databases in the AI Era

Based on these challenges, Bai Shan systematically elaborated on the eight core capabilities that databases should possess in the AI era:

Multimodal Fusion and Vector-Native

Future databases will need a unified storage engine that supports structured data, documents, graphs, time-series, and vector data, enabling multiple access modes for the same data and eliminating data silos. Vector retrieval capabilities must be built into the database kernel, including quantization indexes like HNSW, IVF, and PQ, and support hybrid search combining vector similarity, scalar filtering, and full-text search to directly support AI applications such as RAG, recommendation systems, and similarity search — without requiring separate vector database deployments.

Lakehouse Integration and Real-time HTAP

Through a unified SQL interface supporting open formats like Apache Iceberg and Paimon, it achieves Zero-ETL, avoids data movement, and allows multiple engines to share a single data copy. The same data serves both OLTP and OLAP with millisecond query latency and real-time data updates, eliminating T+1 delays and enabling true real-time decision-making and real-time feature computation.

AI Autonomy and Natural Language Interaction

Databases should feature automatic parameter tuning, index recommendation, anomaly detection, and self-healing capabilities, with intelligent resource scheduling moving operations from 'manual management' to 'autonomy.' NL2SQL technology enables users to generate precise SQL queries using natural language, allowing business personnel to gain data insights without mastering complex syntax.

Compute-Storage Separation and Agent-Native

Compute and storage scale independently, with object storage providing low-cost massive capacity and serverless models enabling pay-as-you-go and elastic scaling. Databases directly serve AI agents via service interfaces, allowing agents to invoke query and management capabilities through natural language. Operations such as backup, recovery, and scaling are completed in one click — elevating the database from a backend system to a core component of AI workflows.

Bai Shan further proposed a bolder prediction: the future form of databases will integrate compute, storage, and inference engines into one, directly serving AI applications.

From DB4AI to DB4DATA: A Fundamental Leap in the Role of Databases

At the end of the speech, Bai Shan encapsulated the fundamental shift in the database’s role as 'From DB4AI to DB4DATA.'

DB4AI is about making databases better support AI — vector retrieval, multimodal storage, and HTAP real-time analysis — the evolution that is happening now. DB4DATA is about enabling AI to better manage and utilize data — intelligent operations, natural language interaction, and agent-driven — the ultimate form of future databases.

Bai Shan concluded: “A database is no longer just a ‘place to store data.’ A database plus business rules is a complete application system. In the AI era, the database is the ‘data intelligence engine,’ the core of a one-stop solution for data and business management.”

From DB4AI to DB4DATA, the role of databases is undergoing a fundamental leap. Multimodal fusion, vector-native, lakehouse integration, real-time HTAP, AI autonomy, NL2SQL, compute-storage separation, and agent-native — these eight core capabilities define the future landscape of AI-native databases. GBase database is continuously evolving along this direction, providing a solid data foundation for application innovation in the AI era.