A Survey of Petabyte-Scale Data Architectures for Self-Serve Generative AI: From Foundational Systems to Intelligent Abstraction and In-Database Optimization

The integration of generative AI (GenAI) into self-service analytics platforms, such as Microsoft Power BI Copilot and Amazon Q, has necessitated a paradigm shift in data architecture. While these tools aim to democratize data access through natural language interfaces, their efficacy is contingent on an underlying data foundation that can manage the challenges of petabyte-scale data. This paper presents a survey and comparative analysis of three advanced architectural strategies designed to address the issues of query latency, cost, and semantic ambiguity inherent in such large datasets. The strategies examined are: (1) the Optimized Data Lakehouse, which focuses on foundational performance and cost-efficiency through open-source formats and high-performance query engines; (2) the Enterprise Semantic Layer and Knowledge Graph, an abstraction-first approach that ensures data consistency and mitigates AI hallucination by providing a structured context; and (3) the Instance-Optimized LLM (IOLM-DB), a specialized, in-database method that overcomes the high cost of per-row AI inference. A detailed analysis reveals that no single solution is universally optimal. Instead, the most effective approach for many organizations is a phased, hybrid architecture that combines the strengths of the first two strategies, reserving the third for specific, high-value applications. This framework provides a robust roadmap for building an AI-native data platform that is both performant at scale and semantically intelligent