From systems of record to AI-native hubs
An AI-native database is a data platform designed to serve autonomous agents and AI workloads as first-class users, blending structured and unstructured information, minimizing data movement, and supporting inexact, context-aware queries rather than only human-crafted SQL. Instead of treating databases as quiet systems of record, we are turning them into active hubs for data movement, decision-making, and AI integration. This is not a cosmetic upgrade; it is a clear architectural pivot. Historically, operational stores like Postgres fed warehouses and search clusters through endless pipelines and copies. Now the pressure runs the other way: analytical systems and AI applications need direct, timely access to operational data without spawning yet another synchronization job. The next frontier for data platforms is not where data lives, but how gracefully it flows between agentic database systems, AI data platforms, and downstream services without breaking consistency or budgets.
Google’s agentic vision: humans step back, agents step up
If you want to see where AI-native databases are headed, listen to the people building the largest ones. At a recent cloud summit, database leaders Sailesh Krishnamurthy and Yasmeen Ahmad described a future where “humans are not going to be using data platforms in the next three to five years. It’s going to be humans orchestrating agents, and agents actually doing the work.” In that world, business users stop staring at dashboards that only answer predictable questions and instead use conversational analytics to probe data more freely. Agentic workloads demand flexible, context-aware queries, mixing natural language, previous interactions, and structured schemas. Krishnamurthy calls this “AI native infrastructure,” where vector and text indexing sit alongside graph technology to combine structured and unstructured data and tolerate inexact results and uneven data quality. The scale is staggering: Spanner now runs 7½ billion queries per second at peak and holds about 23 exabytes of data, while BigTable handles roughly 7 billion queries per second and double-digit exabytes.
Postgres shows how interoperability beats raw storage
Traditional databases are not quietly fading away; they are reshaping themselves to survive in an AI-first world. Postgres, long trusted as a transactional system of record for customer interactions and financial operations, is now being reimagined as a central hub that reduces the need to move data at all. The harder problem is no longer storage or raw performance, but database interoperability: sharing operational data across analytical systems, AI applications, and services without creating yet another brittle pipeline or duplicate copy. Technologies such as logical replication, change data capture, and foreign data wrappers help Postgres participate directly in these larger ecosystems, letting organizations extend AI and machine learning over the same source of truth they already rely on. As AI highlights how much effort teams spend shuffling data, not just saving it, it exposes a limitation that has been growing quietly for years and now demands architectural change.
Ghost and the rise of agent-centric database services
The clearest sign that we are in an agentic era is the arrival of databases built for agents first, humans second. Tiger Data, the team behind TimescaleDB and years of Postgres engineering, has released Ghost, a database service designed specifically for AI agents. Ghost targets one of the most pressing infrastructure problems in AI: developers need databases built for collaborating with agents, not retrofitted from tools designed only for humans. Agents constantly experiment and must be isolated so their failures do not pollute shared environments; Ghost responds with unlimited Postgres databases, fast forking, and instances ranging from ephemeral to dedicated always-on. Crucially, Ghost’s per-query pricing makes these isolated databases cheap enough to treat as disposable, enabling trial-and-error workflows at a scale that used to be impractical. A free tier with 100 compute hours per month, 1TB of storage, and support for hundreds of databases and forks further lowers the barrier to agent-heavy experimentation.
AI-first design, evals, and the practical impact on users
AI data platforms promise convenience, but they also introduce new operational disciplines. Injecting AI into every interaction means paying for tokens on top of traditional compute and storage resources, so careless agent design becomes a cost problem as much as a technical one. Krishnamurthy points to proxy models—tiny models inside the database—that can cut token use by about 400x and reduce latency by 30x–100x. Yet efficiency is only half the story; reliability now depends on eval sets, representative question suites used to verify that AI-generated SQL and inexact queries produce acceptable answers. Meanwhile, AI-first services like knowledge catalogs aggregate structured and unstructured data into unified context for large language models, turning enterprise search into a core ingredient of agent reasoning. Ordinary users may stop worrying about joins and indexes, but they will feel the shift as dashboards give way to conversational interfaces and as agentic database systems quietly orchestrate their data behind the scenes.
