Why Databricks’ Data + AI Summit matters for global data engineering teams in 2026

Ali Ghodsi opened the keynote with a claim that is simple until you try to act on it. AI does not have an intelligence problem. It has a context problem.

If a CFO cannot get an AI to explain why margins moved last quarter, or a sales leader cannot get it to surface the next obvious upsell, that is not the model failing. Often the model is no longer the bottleneck. What is missing is context, the data, the definitions and the business logic the AI needs to give an answer anyone would act on. Get that wrong and the machine will still answer. It just will not be right.

For anyone who runs data engineering, that reframing matters more than any single feature. It moves the hard problem off the model and onto your estate. And it was the thread running through everything Databricks announced at Data + AI Summit 2026.

More than 30,000 people came to the Moscone Center in San Francisco from the 15th to the 18th of June, with tens of thousands more watching from more than 150 countries, at what Databricks calls the largest data and AI conference in the world. The four co-founders headlined, Ali Ghodsi, Matei Zaharia, Arsalan Tavakoli-Shiraji and Reynold Xin, the people behind Apache Spark, Delta Lake and MLflow. Satya Nadella joined by video, Greg Brockman from OpenAI in person, and PepsiCo’s data chief brought it back to a live deployment rather than theory. More than 20,000 organisations now run on the platform, including 70 percent of the Fortune 500.

Ghodsi gave the room a map worth keeping. Four priorities: Context, Cost, Control and Choice. Almost every announcement fits one of them. Once you see the map, the product news stops looking like a list and starts looking like a strategy.

Start with context, because that was the point of the whole event. Genie One went generally available, pitched as an AI coworker that understands your data rather than a chatbot bolted onto it. Underneath it sits Genie Ontology, a self-improving layer that keeps learning what your business means by its own terms, so that “revenue” or “active customer” resolves to the same thing every time the AI is asked. It builds that picture by reading across your tables, dashboards and queries, and the Jira, Slack and SharePoint where a business actually explains itself. That is the unglamorous work that decides whether an agent can be trusted, and it lives in your data model, not in the model weights.

Then the plumbing, and this is the one data engineers will feel first. Databricks set out an architecture it calls LTAP, Lake Transactional and Analytical Processing, where transactional and analytical workloads read and write the same copy of data in open formats, Delta Lake and Apache Iceberg, instead of keeping two copies in step with fragile ETL. The serverless Postgres layer, Lakebase, handles the transactions. Alongside it came Lakehouse//RT, a real-time engine now in beta, powered by Reyden, an in-house joke that stands for “Reynold’s Dream Engine” after co-founder Reynold Xin. Databricks benchmarks it at under 100 milliseconds at 12,000 queries a second, running straight against Delta and Iceberg tables under Unity Catalog governance, and says early customers have seen up to sixteen times the performance of their existing real-time serving stacks. Claims like that get tested hard over the next year, and should be. For teams who would rather not hand-build pipelines at all, Lakeflow Designer, a drag-and-drop natural-language pipeline builder, went generally available.

Control is where the agentic story either becomes real or falls over. Unity Catalog picked up Metrics, turning business definitions and glossaries into governed objects rather than tribal knowledge scattered across dashboards. Unity AI Gateway extends that governance to the agent runtime itself, so what an agent is allowed to see, do and spend becomes a policy rather than a hope. Catalog Federation and OpenSharing push the reach across systems you do not own. The message was blunt. If you want agents in production, governance is not the paperwork you do afterwards. It is what lets them run at all.

Security was the other headline. Lakewatch, the lakehouse-native SIEM Databricks launched earlier this year on the Open Cybersecurity Schema Framework, keeps security data in the same governed lake as everything else instead of a separate and expensive silo. At the Summit, Databricks agreed to acquire Panther, its third security acquisition after Antimatter and SiftD.ai. That is a company deciding trust and governance belong at the core of the platform.

Set the announcements side by side and one intent runs through them. The lakehouse itself is settled. The new question is what you build on top of it, agents that actually work in production, grounded in governed context and cheap enough to run at scale. Databricks reported more than 100,000 agents already built on the platform, and was named a Leader in Gartner’s 2026 Magic Quadrant for AI Platforms. It is behaving like a company that means to own the whole stack, from the transactional store to the agent runtime to the governance layer, as a single platform rather than a bag of parts.

Which changes the job. For a lot of teams the last decade was about moving data from A to B reliably. The next few years are about something harder and more interesting, engineering the context that makes AI trustworthy, governing what agents are allowed to do, and running those agents like the production systems they now are. The pipeline skills do not disappear. They get pointed at a new target. Databricks put that shift on the record at the Summit by launching what it calls the industry’s first certification in context engineering, a formal credential for designing, assembling and governing the information an agent sees. The syllabus reads like the new job description: retrieval, memory that persists across sessions, tool access, and control over exactly what an agent is allowed to know. When a platform vendor builds an exam for a discipline, the discipline has arrived.

Nowhere is the shift better funded than in Asia. Databricks named Simon Davies, a former SAP and Splunk leader, as its head of Asia Pacific and Japan in April, after the region grew more than 85 percent year on year in its most recent quarter. It is putting money behind the words, with a US$300 million investment in Australia and New Zealand and a new regional headquarters in Singapore. That sits on top of real Japanese momentum. Toyota built its unified data and AI platform, vista, on Databricks. NTT DATA committed to training 500 Databricks engineers by fiscal 2027 as part of a roughly $200 million investment. MUFG Bank chose the platform. And in January, Databricks achieved ISMAP certification in Japan, which opens the door to public-sector procurement, the market where reliability and governance are the price of entry rather than an afterthought. Databricks reported over 100 percent year-on-year growth in Japan for the fiscal year ending January 2024, with ANA, Bridgestone, AEON, Konica Minolta, SoftBank and Nikkei among the names on the platform.

Here is why the Summit’s emphasis lands so cleanly in Japan. In Japanese enterprise, a data platform rarely survives contact with procurement, operations and audit unless it is governed, reliable and accountable by default. The context problem Ghodsi described becomes a governance problem in short order, and governance is where Japanese buying logic and this new Databricks direction meet.

If you lead a data engineering team building across Europe, Asia or the US, the takeaway from this Summit is not any single launch. It is the direction. Treat context as a first-class engineering deliverable. Bring your metrics and definitions under governed control before you point an agent at them. Plan for agents as production workloads, with cost and access built in from the start. The teams that do this methodically over the next year will be the ones whose AI actually gets trusted with a decision.

Ultra Tendency was in the room in San Francisco, and turning platforms like this into production systems people rely on, across Europe, Asia and the US, is the work we do. If you are wrestling with lakehouse, governance or agentic AI across regions, this is a good moment to talk.