Data + AI Summit 2026: the announcements that matter
A recap of Databricks Data + AI Summit 2026 in San Francisco, focused on what is relevant if you run data engineering and AI/BI on Azure Databricks.
Anton Corredoira
The most honest line of the whole summit came from Ali Ghodsi himself, as recorded by SiliconANGLE: “This is going to get extremely expensive. We are just scratching the surface.” Thirty thousand people watched agent demos at the Moscone Center from 15 to 18 June, but the question that lingered was more businesslike: which architecture keeps AI alive in production, and who keeps the costs under control?
Below I have filtered the announcements for what is relevant if, like most of my clients, you run data engineering and AI/BI on Azure Databricks. I have skipped the pure data-science and security tooling.
LTAP wants to make your CDC pipelines redundant
The biggest architectural promise is called LTAP (Lake Transactional/Analytical Processing): transactional (OLTP) and analytical (OLAP) workloads on a single copy of the data, in open formats. If it delivers, an entire category of CDC pipelines disappears, pipelines that exist only to copy data from operational systems into the Lakehouse.
On timing, Databricks itself is careful: its own blog talks about rolling it out “in the coming months”, and that phrase means something different on a keynote stage than it does in a sprint planning. Worth following, but plan nothing around it until it reaches preview on Azure.
In the same corner sits Lakehouse//RT, now in beta: a real-time path that, according to the press release, delivers sub-100ms latency at 12,000 queries per second on Delta and Iceberg tables. Interesting for operational use cases, though a vendor’s benchmark holds until your own workload shows otherwise.
Unity Catalog: governance catches up with reality
This was the most important track for most teams, because this is where the answer to the keynote’s cost question lives. The full list is in Databricks’ Unity Catalog roundup; this is what stands out:
- The Unity AI Gateway was positioned as one runtime governance layer for models, tools and agents, with cost controls, unified tracing and hard spend caps. Those caps are the reason to take this seriously: agentic workloads have no natural cost ceiling. Do read the fine print: pieces such as contextual service policies and the guardrails are still in beta.
- Domains moved to Public Preview and Business Glossary reaches preview soon: a governed layer of business meaning, for people and for agents.
- Metrics expanded with multi-fact and parameterized metrics, materialization for faster queries, and imports from Power BI and Tableau. Most of these pieces are still in preview or beta.
- External Lineage is GA, extending lineage beyond Databricks to source systems and BI reports.
- Cross-region governance reaches preview soon, with cross-cloud to follow. Managed disaster recovery is already GA, but as part of a new paid Mission Critical add-on; factor that into your cost picture.
Genie becomes a platform, not a feature
Genie One is GA: an agentic coworker for business teams, alongside Genie Agents and Genie Code. For architects, the layer underneath is more interesting: Genie Ontology, a self-improving knowledge graph that extracts business meaning from your tables, queries, dashboards and more than fifty connected apps, and sits as the context layer under the whole family. One caveat from practice: a graph like that is only as good as the metadata it learns from. If your workspace is mostly tables without descriptions, that is what the ontology gets to see.
The pricing model stands out: no seat-based pricing, but up to 10 dollars of free usage per user per month, after which you pay for the AI that is actually consumed. That fits the keynote’s cost message and keeps experimentation cheap without letting the bill drift out of sight.
Microsoft shops got their own round of announcements: Genie in Microsoft Teams and M365 Copilot is in beta, letting colleagues tag Genie in a Teams thread and get governed answers from the Lakehouse. On top of that, OneLake interoperability is GA, so you can query data in OneLake through Unity Catalog without copying it, and an Excel add-in arrived in public preview.
Lakebase and apps
Lakebase, the managed Postgres layer for operational data, gained cross-cloud and cross-region disaster recovery, git-style branching, autonomous operations and Lakebase Search (beta) for hybrid vector and full-text retrieval in the same Postgres instance. That last one saves you a separate vector database next to your operational data. Building apps got easier with App Spaces, Serverless Micro Apps that scale to zero, a Genie App Builder and an Apps Marketplace.
This is exactly the foundation under the data-driven apps I build on Databricks, so I welcome the direction.
Agents and platform, briefly
Outside my core focus, but worth knowing: Agent Bricks expanded with broader model choice and managed agent memory, and Omnigent was introduced as an open-source meta-harness for combining agents, with a managed version in beta on Databricks.
In closing
Three things deserve a spot on your roadmap. The Unity AI Gateway, because without spend caps and policies you will not keep the costs of agentic AI under control; do track which pieces have left beta. Genie Ontology and the metrics layer, because reliable self-service analytics stands or falls with that context layer, and that layer in turn stands or falls with your metadata. And LTAP, once it reaches preview on Azure, because it could cut a lot of plumbing. The rest was a good show; these three change architecture decisions.
Want to talk through what this means for your setup? Get in touch.