Redesigning the Data Stack for AI Agents, Not Humans
LTAP architecture (Lakehouse Transactional Analytical Processing) is Databricks’ new design that unifies transactional and analytical data in one lakehouse so AI agents can read, reason over history, and act on live information without shuttling data across separate systems. Instead of optimizing databases for human analysts running occasional dashboards, Databricks assumes autonomous AI agents will become the primary database users, issuing rapid tool calls, running loops, and triggering actions at machine speed. Co-founder and CEO Ali Ghodsi framed the shift bluntly, noting that organizations have “effectively doubled their workforce, just not with humans,” and that legacy infrastructure has turned into a bottleneck. LTAP is Databricks’ answer: a lakehouse technology stack tuned for AI agent databases, with governance, storage, and compute laid out so agents can move from query to action on a single, consistent source of truth.
From Two Databases to One LTAP Lakehouse
For decades, enterprises ran one database for online transactional processing and another for online analytical processing, then linked them through ETL jobs and replicas. Databricks argues that agentic applications break this pattern because agents need current transactions and long-range analytics in the same place, at the same time. With LTAP, transactional and analytical data share a single storage layer in open formats on cloud object storage, while separate compute engines handle each workload. This design grows out of Lakebase, Databricks’ Postgres-based operational database, which separates compute from storage and places operational data directly in the lake. On top of that, Databricks adds native vector and full-text search, real-time ingestion through Zerobus in Lakeflow Connect, and Git-style branching so agents can clone a database state, experiment, and discard it without waiting on slow infrastructure.
Targeting Agentic Applications and New Competitors
By turning the lakehouse into an LTAP architecture, Databricks is moving into territory long held by ClickHouse, Splunk, and Oracle, but with AI agents as the first-class users. The company is pitching a single, governed data platform for agentic applications: one environment where operational events stream in, analytical views stay current, and models and agents query everything through the same lakehouse technology. Earlier hybrid transactional and analytical approaches (HTAP) usually meant high costs and proprietary lock-in, while so-called zero-ETL patterns still kept multiple copies of data and left it prone to becoming stale. Databricks is betting that database consolidation around open formats and shared storage will be more attractive in an era when AI agents perform continuous queries, execute tool calls, and drive automated workflows rather than waiting for humans to ask occasional business questions.
NVIDIA Acceleration for the Agentic Era
Databricks’ expanded partnership with NVIDIA anchors LTAP in an accelerated hardware stack aimed at AI agent databases. NVIDIA GPUs already power Databricks AI Runtime for training and fine-tuning models on governed enterprise data, with Hopper GPUs and NVIDIA Quantum InfiniBand networking set up for multi-node distributed training and future support for the Blackwell architecture. On the inference side, Databricks Model Serving uses NVIDIA hardware and Triton Inference Server to deliver low-latency, high-throughput model calls that agentic applications depend on. The companies are now focusing on the next bottleneck: CPU-bound work like tool calls, orchestration, and analytics inside agent loops. NVIDIA Vera, described as a next-generation CPU for agentic workloads, reinforcement learning, and CPU-based data analytics, is positioned to keep LTAP-driven agents responsive even as their reasoning chains and database interactions grow more complex.
