Part I — The Layer
The AI stack has one unresolved layer
Models, compute, cloud, analytics: each adjacent layer has consolidated into an oligopoly of well-capitalised incumbents. The operational data layer has not.
This layer sits between analytics — where warehouses and lakehouses operate on historical data — and applications, where agents operate on live state. It is architecturally distinct from both. Analytical systems optimise for large-scale reads over stable schemas; the operational data layer supports low-latency writes over evolving schemas, with the characteristics agentic applications require: vector retrieval, semantic search, persistent memory, and transactional consistency over heterogeneous data.
No single incumbent dominates it. Multiple architectures remain viable; hyperscalers, established database vendors, open-source communities, and adjacent platforms are all contesting the position. Which approaches prove durable — and on what horizon — are open questions.
Why the layer matters
Every AI-native application that graduates from prototype to production needs infrastructure here: durable state, retrieval over accumulated context, memory across sessions, transactional guarantees for tool use. It is the substrate on which the next generation of enterprise software runs — and at scale, a control point whose significance extends to the enterprises that depend on it, the ecosystems built on it, and questions of national technological competitiveness.