Databricks Vice President: Claude Code Is Just the Beginning; AI Factories Need New Infrastructure
Databricks Vice President Nikita Shamgunov said in a recent interview that Claude Code still attracts attention, but the next focus is the “AI factory,” meaning a platform that shifts from assisting code generation to delivering software like a factory.
Shamgunov said that software ultimately still needs to be deployed on some kind of infrastructure, and that future AI factories will run better on agent infrastructure. He specifically pointed out that state is crucial, such as database state, and that this may be one of the hardest problems to solve.
The related discussion appeared on the MTS program. Shamgunov also introduced Databricks’ Lakebase Postgres architecture, emphasizing that its low cost and branching capabilities make it suitable for scenarios in which agents create databases at scale. The Neon platform had previously seen a large number of database instances created by AI agents.
Original sources:
Title: Databricks VP @nikitabase argues Claude Code is just the beginning... Publisher: MTS (@MTSlive) Publication date: 2026-10-29 URL: https://x.com/MTSlive/status/2115951648090841446 Nature of source: media original post (paraphrased interview) Verification status: full text of the original post actually read for this report (2026-10-30) Supported content: Shamgunov’s remarks on Claude Code, AI factories, agent infrastructure, and database state
Title: Hyperscalers Have the Wrong Primitives for AI | Databricks Publisher: MTS (YouTube) Publication date: 2026-10-30 URL: https://www.youtube.com/watch?v=ip3v3WfaDlw Nature of source: interview video Verification status: description and partial transcript actually read for this report (2026-10-30) Supported content: background for Shamgunov’s discussion of agent infrastructure and Lakebase; does not prove the full context of all quotations
ABAB AI Insight
Nikita Shamgunov was previously the founder of Neon, which provides serverless Postgres and has been acquired by Databricks; he is now a Databricks vice president responsible for related database products. Neon data shows that the share of databases created by AI agents rose from about 30% to 80%, and he once predicted it could reach 99% in the future. After Replit Agent launched, it once created about 13,000 databases in a single day.
The Lakebase architecture separates compute and storage and supports branching, similar to a test environment, allowing agents to create and modify database state at low cost and decide whether to merge it. Shamgunov said that the primitives offered by existing cloud hyperscalers are better suited to traditional workloads than to the large-scale, short-lived, parallel database operations of agents.
Compared with coding tools such as Cursor and Windsurf, Claude Code and similar products still focus on developer assistance, while “AI factories” point to platforms that can directly generate and deploy complete applications from prompts, such as Replit and Lovable. Database state becomes a bottleneck because each generated application usually needs persistent data, and agents create databases far more frequently than human developers do.
This view distinguishes code generation tools from the infrastructure layer and emphasizes that under an agent consumption model, low-cost startup, branching, and state management for databases become key constraints. Shamgunov also noted that Lakebase serves both agent scenarios and large-scale enterprise applications, but database state management remains an unresolved problem.
ABAB News · Cognitive Law
After code generation tools become widespread, software delivery is still constrained by the capability of the underlying state infrastructure.
When agents create databases far more frequently than humans, low cost and branching become key design constraints.
From assisted coding to factory-style delivery, database state management is said to be one of the hardest problems to solve.