Databricks Data + AI Summit 2026 brought 32,000+ to San Francisco, where CEO Ali Ghodsi's "context problem" thesis framed the week. Genie, Agent Bricks, and the new Unity AI Gateway reorganized enterprise AI around four priorities: Context, Cost, Control, and Choice.
June 15-18, 2026 | Moscone Center, San Francisco | 32,000+ in-person attendees | 75,000 virtual | 800+ breakout sessions
Databricks Data + AI Summit returned to Moscone Center in San Francisco from June 15-18. The Summit brought together more than 32,000 in-person attendees and another 75,000 joining virtually. Ahead of the event, Databricks said participants would represent more than 150 countries. Across four days the event packed in 800-plus breakout sessions, 25-plus hands-on training and certification courses, and a multi-day hackathon run in partnership with OpenAI to build agentic data apps for social impact. Databricks bills it as the world's largest data, analytics, and AI conference, and this edition was built around a single, sharpened argument.
That argument came from CEO Ali Ghodsi's keynote. His central point was that enterprise AI no longer has primarily an intelligence problem - it has a context problem. Models are already smart enough, the pitch went; what holds enterprise AI back is the surrounding context - the governed data, the business definitions, the access controls - that lets a model act reliably inside a specific organization. If a CFO cannot get an assistant to explain why margins moved, Ghodsi argued, that is a context gap, not a model gap. He structured the priorities around four words: Context, Cost, Control, and Choice, and every headline announcement mapped back to one of them.
The most concrete expression of the context thesis was Genie. What began as a natural-language query tool was repositioned as an agentic coworker that produces reports and artifacts across data types and is available through the web and native mobile apps.
Genie One reached general availability on the web, moving it beyond the experimental stage, while its native iOS and Android apps launched in public preview. It connects to more than 50 workplace apps and data systems. Around it, the company filled out an entire Genie family. Genie Agents (GA) turn a conversation into a reusable, governed workflow that can be deployed into Teams and Slack. (GA) is a coding assistant trained on Databricks platform context. and were both announced for private preview shortly after the Summit - the former generating governed applications from a plain-language description, the latter monitoring pipelines, detecting failures, and proposing fixes on its own.
Underneath all of it sits Genie Ontology, a live context layer that automatically extracts business knowledge not just from Databricks but from the workplace apps where definitions actually live - Google Drive, Jira, Slack, Confluence, and SharePoint. The through-line is that a model is only as useful as the semantic scaffolding around it. Genie Ontology is Databricks' bet that the differentiator in enterprise AI will not be the model alone, but the governed context layer feeding it - the context problem from Ghodsi's opening argument translated into a product strategy.
If context was the philosophy, agents were the strategy - and the strategy was openness. Agent Bricks was expanded into a full enterprise agent platform and, crucially, opened to external agent SDKs including the Claude Code SDK, LangGraph, CrewAI, and OpenAI's Agent SDKs. Databricks reported that customers had already built more than 100,000 agents on it. Rather than force teams onto a single proprietary framework, the company positioned its platform as the governed data foundation underneath whatever agent tooling a team already uses.
The platform also grew the plumbing that production agents need. Managed Agent Memory services, backed by Lakebase, persist conversation history and preferences. A secure Databricks Sandbox isolates agent code execution. Model support expanded beyond the usual names to include Kimi from Moonshot AI and Grok models through a partnership with SpaceX. Separately, Databricks open-sourced Omnigent, an alpha-stage meta-harness for combining agents across frameworks under shared cost budgets; a managed Databricks version is available in Beta.
Read together, these releases map cleanly onto the Choice pillar. Databricks is wagering that the durable position in the AI stack is the layer that stays framework-agnostic - the place data lives and gets governed - not any one agent runtime that could be displaced next quarter. The featured-session lineup reinforced the posture, with Anthropic, CrewAI, LangChain, and LlamaIndex sharing the agenda rather than competing off it.
Two pillars that rarely headline a keynote - Control and Cost - got unusual airtime, a sign of how central both have become as agents move from pilots to production.
On control, Unity Catalog drew heavy attention. New Unity Catalog Metrics let organizations define a KPI once as a governed object that is queryable from SQL, BI tools, APIs, and agents alike. Databricks also introduced a new four-level Unity Catalog namespace and laid out a roadmap for governance across regions, clouds, and accounts, with cross-region support coming first in preview. Domains (public preview) organize data and AI assets along business lines. On security, Databricks launched Lakewatch (private preview), an agentic SIEM that keeps security data in open lakehouse formats, and agreed to acquire detection-as-code vendor Panther.
On cost, Ghodsi was blunt that agentic consumption is going to get extremely expensive, and the answer was Unity AI Gateway (Beta), a runtime control plane for models, agents, tools, and MCP servers. It adds spend caps, smart routing to cost-appropriate models, PII guardrails, tracing, and an agent registry. The pitch to a data leader anywhere from São Paulo to Singapore is that agents are only deployable at scale if every query, metric, action, and dollar runs through a governance layer that can be audited. That is precisely the friction that has kept many enterprise AI pilots from reaching production.
For all the agent talk, Databricks used the Summit to make clear it still intends to own the storage and query layer underneath. Lakebase, its serverless Postgres on open object storage, reached general availability with Git-style branching, point-in-time recovery, and a multicloud architecture. Two forward-looking bets pushed further. LTAP (Lake Transactional/Analytical Processing) aims to unify transactional and analytical workloads on a single governed copy of data stored natively in Delta/Iceberg. Lakehouse//RT, introduced in Beta, is a real-time tier powered by the new Reyden engine, with benchmark results in the sub-100-millisecond range.
Rounding out the data story were an expanded Lakeflow ingestion and pipeline suite, the push-based Zerobus Ingest API, managed Iceberg and Iceberg v3 support for interoperable workloads on a single copy of data, and OpenSharing, a Linux Foundation protocol for sharing data, models, and agent skills. A meaningfully expanded Free Edition folded in Genie Code, serverless GPUs, Lakebase, Agent Bricks, and Lakeflow Designer, lowering the on-ramp for the next wave of builders.
The keynote stage doubled as a map of Databricks' partnerships. Co-founders Ali Ghodsi, Matei Zaharia, Reynold Xin, and Arsalan Tavakoli-Shiraji were joined by Microsoft Chairman and CEO Satya Nadella in a pre-recorded fireside, OpenAI President Greg Brockman, and PepsiCo Global Chief Data and AI Officer Magesh Bagavathi. The pairing of Microsoft and OpenAI on the same program, alongside the joint hackathon with OpenAI, underlined how much of the current AI buildout runs through a small set of overlapping alliances.
Customer stories from PepsiCo, Mastercard, Reliance Industries, and AstraZeneca helped ground the announcements in large-enterprise use cases across consumer goods, financial services, energy, and pharmaceuticals.
For teams that sell into or buy from the enterprise data-and-AI market, the Summit sent three practical signals:
The 2026 edition marked a clear shift in how the enterprise data world talks about AI. A year ago the conversation was about model capability; this year it was about everything around the model - the governed data, the semantic layer, the access controls, the cost meter, and the agent frameworks that turn a capable model into a deployable system.
Databricks' answer, across Genie, Agent Bricks, Unity Catalog, and the Lakebase/LTAP data layer, is that the winning position is the one that owns context and governance while staying open to whatever models and agents come next. Whether that thesis holds is the story the industry will test between now and the next Summit. For teams evaluating where the enterprise AI market is heading, the signal from Moscone was consistent: the hard problem is no longer intelligence, it is context.
