Multi-Model and Multi-Form: Intelligent Connectivity for the Future | GBase 8c, the Next-Generation AI-Native Database

On June 25, the 2026 GBASE Technology Cloud Conference was grandly held at the auditorium of Tiankai Park in Tianjin. Zhang Yi, General Manager of the GBase 8c Product Division at GBASE, was invited to deliver a keynote titled "Multi-Model and Multi-Form: Intelligent Connectivity for the Future – GBase 8c, the Next-Generation AI-Native Database". Starting from the new challenges databases face in the AI era, he systematically introduced GBase 8c's technological innovations in multi-model storage, multi-form deployment, and AI-native capabilities, and demonstrated its practical implementation in critical sectors such as finance and telecommunications through real-world cases like the Bank of Luzhou and a major carrier's settlement system.

At the start of his speech, Zhang Yi pointed out that the arrival of the AI era imposes entirely new requirements on databases. Training data for large models has reached petabyte scale, making storage expansion difficult for traditional databases; AI training and inference workloads fluctuate dramatically, requiring on-demand scaling of computing resources; model inference demands millisecond-level responses, placing extremely high demands on query performance; concurrent access from AI agents to databases creates demands for millions of concurrent connections; and new scenarios such as RAG and semantic search require native vector support, with most AI applications needing hybrid retrieval of scalar and vector data. Moreover, AI scenarios demand much higher data freshness in warehouses, which traditional ETL data flow efficiency can no longer satisfy.

"The core design of traditional relational databases can no longer meet all these new requirements simultaneously," Zhang Yi said. "Databases must evolve, and AI-native databases are on the way."

GBase 8c: Multi-Model and Multi-Form – One System for Diverse Scenarios

GBase 8c is GBASE's self-developed next-generation AI-native database. Its core capabilities can be summarized as "multi-model and multi-form" – covering multi-model storage and multi-form deployment.

Multi-Model Storage: Row, Column, and Hybrid – One System for Both TP and AP

Zhang Yi detailed GBase 8c's three storage modes. Row-based storage physically stores rows contiguously, supporting high-concurrency transactions with MVCC and WAL logs, suitable for OLTP scenarios. Column-based storage physically stores columns contiguously, utilizing CU compression units and multiple compression algorithms for high compression ratios and low I/O overhead, suitable for OLAP analytics. Hybrid row-column storage allows a single table to have both row and column physical forms, using a delta table mechanism for row-to-column transformation, with the query optimizer automatically choosing the optimal path for OLTP and OLAP. The HTAP hybrid storage eliminates the data silos between OLTP and OLAP found in traditional architectures, enabling one system to simultaneously support transactions and analytics.

Multi-Form Deployment: Primary-Backup, Distributed, and Compute-Storage Separation – Flexibly Adapting to Different Scales

GBase 8c offers three deployment forms. The primary-backup deployment, with one primary and multiple backups, is simple to deliver and efficient, suitable for scenarios with lower data volumes that demand extreme single-machine performance. The distributed deployment features full-component redundancy for high availability and separation of compute and storage, allowing independent horizontal scaling of both capabilities, suitable for large data volumes and high concurrency. The compute-storage separation deployment uses shared object storage, with a stateless compute layer that supports second-level scaling and Scale-to-Zero, where compute costs approach zero when idle – especially well-suited for scenarios with highly fluctuating AI training and inference workloads.

AI-Native Capabilities: Vector Retrieval, Compute-Storage Separation, and Data Branching

Native Vector Storage and Hybrid Retrieval

GBase 8c features built-in vector storage and quantization indexes such as HNSW, IVF, and PQ, supporting hybrid retrieval of vector similarity, scalar filtering, and full-text search, which can directly power AI applications like RAG, recommendation systems, and similarity search. SQL joins can combine with vector search into complex hybrid SQL, generating bitmaps based on scalar filter conditions and then using ANN retrieval algorithms to obtain vector data, satisfying both bitmap filtering and vector search conditions.

Compute-Storage Separation and Scale-to-Zero

The compute layer is completely stateless – nodes hold no persistent data; state is continuously synchronized to the storage layer via WAL streaming. During high-load AI training, compute nodes are automatically added; during inference or idle periods, compute resources are scaled back, and at zero load, Scale-to-Zero reduces compute costs to near zero. When new loads arrive, the system wakes up within seconds to quickly restore services.

Copy-on-Write Data Branching Based on WAL

When creating a branch, there is no need to copy actual data; only the WAL log position is recorded, enabling branch creation in seconds. The new branch shares historical pages with the original data; only modified pages are copied upon writes. Multiple branches are completely isolated without impacting each other, supporting the parallel execution of multiple AI experiments, with fast rollback.

Unified Lakehouse Transactional and Analytical Processing

In terms of lakehouse integration, GBase 8c breaks the traditional model separating OLTP and data warehouses. Traditional approaches require ETL to move data from OLTP systems to warehouses, with delays of hours and data that is already stale. GBase 8c works in synergy with GBase 8a to achieve real-time data mirroring, allowing the same fresh data to simultaneously serve online transactions and data analytics, so AI applications and real-time decisions are based on the most current data.

Core Value: Reduce Costs, Increase Efficiency, and Simplify Architecture

Zhang Yi highlighted the three core values of GBase 8c.

Significantly Lower Hardware Investment Costs

GBase 8c supports a maximum single-node capacity of 10TB to 30TB, far exceeding the 2TB limitation of traditional solutions. With fewer physical nodes for the same data scale, hardware costs are reduced by 50% to 80%, while also saving on servers, data center space, power, and operations staff.

Substantially Reduce Application Development and Migration Costs

The natively distributed architecture is transparent to applications; the application layer does not need to be aware of underlying distributed details. It supports arbitrarily complex cross-node JOIN queries without middleware limitations. It allows smooth migration of existing centralized operations, avoiding extensive code refactoring and significantly reducing development and migration costs.

Greatly Simplify System Architecture Complexity

GBase 8c's built-in multi-storage engines for row and column data allow a single cluster to handle the mixed workload of high-concurrency TP transactions and complex AP analytics, truly realizing "one copy of data, two types of services" and dramatically reducing system coupling.

Practical Cases: Proven in Core Financial and Telecom Systems

Luzhou Bank's Credit System, Electronic Banking, and Distributed Core System

With business growth and increasing data volumes, traditional centralized databases could no longer meet high concurrency and large-scale data processing demands. Luzhou Bank chose GBase 8c as a unified foundation to reshape its data architecture from the ground up. After the project went live, results were impressive: overnight batch processing was reduced from 2.5–3 hours to about 30 minutes; data synchronization from 2–3 hours to just minutes; failover from manual operation to automatic within seconds. Core query response efficiency improved by 44%, data channels across electronic banking channels were connected, and core data such as customers, accounts, and transactions achieved centralized management and unified standards.

Settlement System of a Major Telecom Carrier

The traditional centralized architecture struggled with the flexible business needs of the settlement center, which handles tens of billions of call detail records per month. After deploying GBase 8c in a distributed manner, real-time settlement processing capacity reached 180 billion CDRs/month, and the first-tier system uploading capacity reached 15 billion CDRs/month. The monthly billing completion time for inter-network and inter-provincial services was controlled within 2 hours, and that for local SP services within 8 hours. The system now supports over 150 settlement services.

A Provincial-level Universal Health Information Platform

The platform is highly compatible with MySQL and Oracle syntax, minimizing migration and adaptation work. The central system is deployed in a distributed manner to support large data volumes and high concurrency, with a stored data volume reaching 5PB. The front-end devices at medical institutions use single-machine deployment, minimizing software and hardware costs, while unified access interfaces reduce business development costs.

From multi-model storage to multi-form deployment, and from native vector retrieval to compute-storage separation, GBase 8c has multi-model and multi-form at its core, providing a complete technological foundation for data management in the AI era. Going forward, GBase 8c will continue to deepen its AI-native capabilities, making databases a true "Data Intelligence Engine" for the AI era.