Cloud database systems are entering a period of fundamental change. Both software workloads and hardware architectures are being reshaped at once, breaking long-standing design assumptions. On one hand, AI-driven applications, including large language models, recommendation systems, and autonomous agents, introduce radically new workloads centered around embedding vectors, approximate nearest neighbor search, and hybrid queries that combine vector, text, and structured data. On the other hand, the end of Moore’s and Dennard’s scaling laws has slowed improvements in CPU speed, RAM capacity, and disk/flash capacity. The scale of investment in AI-driven hardware accelerators such as GPUs, TPUs, and custom engines along with CXL-based memory disaggregation has brought us into the era where the CPU is no longer the center of hardware innovation.

This Dagstuhl Seminar will address the disruptive innovation needed in the data management stack, from hardware to database systems to data-hungry applications, to enable software to leverage new (AI-driven) commodity hardware.

We aim to bring together leading researchers and practitioners from database systems, hardware architecture, and computer systems to rethink, from the ground up, how to co-design cloud data systems for the AI Era. Many directions will be discussed: What are the characteristics of cloud database workloads? What are the right storage formats, access methods, and data organization for hybrid queries and semantic operators? How to integrate exact and approximate neighbor search in database systems? Should we design hardware and software for tight coupling between CPUs and AI accelerators? Should relational and embedding operators both be executed on AI hardware? What metrics should we optimize for (queries per watt versus queries per second)? How should a system be designed for deployments with abundant memory via CXL technologies? How can we ensure reliability and defense against hardware silent data corruption for heterogeneous architectures? What are the most promising AI + DB architectures (Tensor processor, GPU, PIM)? And most importantly, how do we avoid the pitfalls of much past work, both academic and industrial, on accelerators or storage subsystems that wind up failing due to limited impact?

Creative Commons BY 4.0

David F. Bacon, Yannis Chronis, Jana Giceva, and David A. Patterson

This seminar qualifies for Dagstuhl's LZI Junior Researchers program. Schloss Dagstuhl wishes to enable the participation of junior scientists with a specialisation fitting for this Dagstuhl Seminar, even if they are not on the radar of the organizers. Applications by outstanding junior scientists are possible until January 16, 2026.

Creative Commons BY 4.0