# Postgres Inventor and Steward on AI Architecture and Data Sovereignty Michael Stonebraker and Bruce Momjian discuss why relational databases remain essential for AI. - **Source:** [www.theregister.com - Articles](https://brifing.si/source/2c34ea50-d341-4a10-925d-12c012a9ad10) - **Published:** 2026-10-29 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/80970f4a-5113-4c2d-86d5-9f68ba918c5f ## The Reality of Text-to-SQL Michael Stonebraker highlights that while public benchmarks claim high accuracy for text-to-SQL, real-world enterprise data yields zero-percent success. MIT’s internal data warehouse, containing 1,400 tables with cryptic naming conventions and idiosyncratic structures, proved too complex for current models. Even with prompt engineering and retrieval, accuracy remains insufficient for enterprise-grade production environments. ## Architectural Limits of LLMs Large language models struggle with structured data joins, which are fundamental to enterprise operations. Stonebraker’s team found that orchestrating complex tasks like transport scheduling via LLMs is prohibitively expensive. Their Rubicon system separates text processing from tabular operations, using a query optimizer approach to achieve ten times the efficiency and better accuracy. ## Database-Level Guardrails Enterprises require determinism to avoid legal risks, such as reversing approved loans. Stonebraker argues that databases already possess essential guardrails, including access control, comprehensive logs, and transaction triggers. He suggests that data-heavy AI agents should function as user-defined functions within the database layer, provided that debugging and versioning tools continue to improve. ## The Myth of AGI Stonebraker dismisses AGI as marketing hype designed to justify high valuations for foundation model vendors. He notes that no enterprise will allow proprietary data to be used for training public models. Consequently, the future of AI in the enterprise relies on systems that maintain strict control over data access and operational security. ## Independence as a Foundation Postgres remains a leading choice because no single company controls it. Unlike proprietary systems, Postgres is managed by a global community, ensuring vendor neutrality. This independence provides enterprises with data sovereignty, allowing them to retain control over their systems without the risk of vendor lock-in or sudden changes in corporate ownership. ## Community-Driven Development The PostgreSQL Global Development Group operates without a central office or legal entity, yet produces high-quality software. Bruce Momjian explains that the community acts as a global forecasting tool, identifying technical needs from diverse regions. This decentralized model allows for rapid, expert-led innovation that often outpaces the capabilities of billion-dollar commercial software companies. ## Design Philosophy of Flexibility Postgres was built to bend, allowing it to absorb features like JSON, GIS, and full-text search without requiring a complete rewrite. Momjian emphasizes that the project prioritizes long-term flexibility over short-term speed optimizations. By leaning on the operating system, Postgres ensures it remains a natural, reliable fit for cloud environments over decades. ## The Future of AI Memory Stonebraker and Momjian agree that AI requires a relational system to provide permanence and memory. While they differ on the speed of AI-generated SQL adoption, they view the database as the essential foundation for AI. Forty years after its invention, Postgres continues to evolve as the primary engine for structured data. ## Sources covering this story - https://www.theregister.com/databases/2026/10/29/partner-content-built-to-bend-postgress-inventor-and-a-steward-on-what-ai-really-needs/5305957 — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._