Databricks Data + AI Summit 2025 Keynote Recap: The 5 Biggest Announcements

There we were again—in the sonically aggressive techno-scape of Moscone’s ballroom, waiting for the next spate of industry-defining announcements to echo through its halls. It was a full-on visual and auditory assault.

However, as soon as Ali Ghodsi’s tailored blazer hit the stage, the announcements came fast and furious. 

Missed Wednesday’s keynote? Check out our recap to find out what was shared, what’s coming next, and what it all means for the “data intelligence platform.”

Revelations this way. 

The Walk Up

In contrast to Snowflake Summit’s bizarre live brass and orchestral sections, Data + AI Summit attendees were met with an onslaught of electrified bass accompanied by the absolute maximum number of speakers and stagelights allowed in this country.

The entire audience lost the ability to hear at least four frequencies before the keynote started.

While Databricks wasn’t able to book the same live musicians as their contemporary, they did continue the tradition of galvanizing welcome videos—complete with a stream-of-consciousness blend of words and phrases about challenging the status quo and curing cancer or some such.

It was loud. It was bewildering. It was a lot. And that was entirely by design (I think). Databricks was gearing up to make some big announcements, and they wanted everyone in the audience to know it—so galvanize they must. 

And as soon as the onslaught abated, Ali Ghodsi marched onto the stage to deliver the goods.

Ali Ghodsi’s Opening Remarks

Ali opened the keynote in typical conference fashion by running off a list of impressive statistics about the event—including 22,000 in-person attendees, which after reviewing my notes is up an additional 6,000 attendees from 2024 (if those figures are accurate). 

The growth in attendance—and the sheer energy from the crowd—reflected a momentum that’s been amassing around Databricks’ democratized vision for data + AI for some time. You would think the Beatles were hosting a reunion tour the way this crowd was cheering. 

Ali began his remarks in a familiar place, by discussing the fragmentation of the data + AI estate. 

“…the reality on the ground is, it’s still actually really hard to succeed with data and AI—especially if you’re in a company that’s been around more than 10 or 15 years.”

The reason—at least according to Ali’s officially approved point of view—is that the data + AI estate is still too fragmented. It’s long been Databricks’ position that in order for enterprise data + AI teams to succeed, they need to verticalize—and that position is on full display in this year’s announcements.

“Each of these systems has data sitting in it, and that’s what’s locking people in…each of them has their own security model, their own access control, their own governance.”

If there’s one thing Databricks excels at, it’s telling their story. Nothing on this stage happens by accident (unless you were at the cookie demo fiasco last year). Databricks Summit keynote is a carefully curated collection of announcements and interviews designed to articulate one very particular vision for the future of data and AI.

In this case, that vision is a fully-integrated data + AI ecosystem inside the Databricks platform. 

So, with that in mind, let’s dive into the top 5 announcements from Databricks Summit to see how that vision is playing itself out. 

Unity Catalog Advancements

If you thought “democratization” was Ali Ghodsi’s favorite word, you would be right. 

Following its acquisition of Tabular back in 2024, you knew open table formats were going to make an appearance at DAIS in 2025—and appear they did, alongside several enhancements to Databricks Unity Catalog. 

As part of a bucket of Unity Catalog enhancements, Ali kicked off his keynote by announcing complete support for both Apache Iceberg and Delta Lake within its proprietary catalog solution, including reading and writing from any system “with fine-grained governance, to performance-optimized, Iceberg managed tables.”

According to Ali, this enhancement is designed to eliminate lock-in by enabling seamless interoperability on the back of open format—furthering Databricks’ stated intention of reducing fragmentation and giving data back to the owners. 

Ali went on to announce a couple new enhancements designed to bring business users closer to the data (and platform teams closer to the business users), including:

• Unity Catalog Metrics which brings metric definitions into the data platform traditionally siloed within BI tools 
• And an internal marketplace feature that’s designed to surface high-value data, AI, and BI assets for business users

While seemingly disparate in scope, the Unity Catalog umbrella (and its features) acts as a tent pole of Databricks’ vision for democratization—so you can expect a majority of Databricks’ forthcoming announcements to ratchet into this solution. 

Databricks Goes Free—And The Crowd Goes Wild(ish)

While not necessarily a platform enhancement, it’s certainly an announcement, and it bears mentioning here—if for no other reason than how jazzed the audience got about it.

After some more narrative meandering about democratization, Ali dropped a bomb by announcing the launch of a new Free Edition of Databricks alongside a $100M investment in training designed to teach the fundamentals of data + AI engineering and data science through the lens of the Databricks ecosystem. 

Make no mistake though, this is marketing at its best. Along with complete access to a production version of Databricks, learners will be able to:

• Build AI agents and applications
• Collaborate on ML projects
• Creative interactive dashboards
• Query and analyze with SQL
• Learn how to build pipelines
• Get coding help through the Databricks assistant (robot teachers)
• Invite collaborators
• And access free training content on the Databricks academy.

While free training is an admirable pursuit, I can’t help but wonder if focusing learning within a single, largely abstracted and fully-verticalized platform will actually create more of a skills gap than it solves. 

Food for thought. 

Lakebase Joins The Race

At about the midway point of his announcements, AlI revealed Databricks’ questionably named answer to the problem of fragmentation between analytics and operations for AI applications.

Touted as ​​a first-of-its-kind fully-managed Postgres database, Lakebase is a database architecture that adds an operational database layer to the company’s Data Intelligence Platform for production AI.

Setting up the reveal, Ali stated that, “database technology was built for a different era.” And he’s not wrong. Database technology was developed at a time when ML modeling was still the most innovative thing happening in the data space.

As the name implies (mostly), Lakebase splits the database into a base and lake layer,  where the data sitting in traditional transactional databases is instead stored within an open format data lake (Databricks Lakehouse for Ali’s purposes) and the processing is handled within an open source base layer (in this instance powered by Postgres). 

Again, this is all about unifying systems, architecture, and teams around one verticalized data + AI platform—Databricks.

Powered by Neon technology which Databricks acquired earlier this month, Lakebase reportedly offers: 

• fully independent storage and compute
• continuous autoscaling for compute
• low latency (<10 ms), high concurrency (>10K QPS) and high availability transactional needs
• And “branching” to quickly spin up new copy-on-write database clones for low-risk AI development and experimentation—and easy deprecation

I’m generally a stoic observer at these events, but I’ve gotta admit, this branching feature sounds pretty cool. 

Agent Bricks Could Verticalize Agent Creation

Ali got one thing absolutely right; “Agent hype is everywhere.”  Unfortunately hype doesn’t equate to real-world success.

One of the single greatest factors impacting agent success is trust. Gartner recently released a report that identified the same from its own research. 

So, it was no surprise to see Databricks announce a new agent builder that included an evaluator solution to optimize performance as part of its data + AI suite.

Pegged as a “new, automated way to create high-performing AI agents tailored to your business”, the referentially named “Agent Bricks” builds agents based on a single description of the task to be performed (potentially…if you get the description right…), and then creates task-specific evaluators and LLM judges to assess performance.

The new tool features:

• An information extraction agent to structure data
• knowledge assistant agent for evaluations
• multi-agent supervisor to string together multiple agents
• And a custom LLM agent to support domains

Arguably the coolest feature of Agent Bricks is an ability to programmatically assess the relationship between output quality and cost to recommend the right mix of power for a given goal.  

Databricks Wants You to Build More with Databricks Apps

Last but not least is the new Databricks Apps release.

If you haven’t guessed, Databricks wants you to do more inside of Databricks. The GA release of Databricks apps is the next phase of that dream.

Designed as a verticalized (there’s that word again) app builder, Databricks Apps ostensibly abstracts the stack decision/development process away from builders by managing everything in Unity Catalog.

The Databricks website describes this as “simple, production-ready apps with built-in governance.” Ali described it as “everything you need to build the best data intelligence app already deployed.”

The audible “wows” from the crowd demonstrates that this has been a felt need for some time among developers. 

Where You Build Matters a Little—Trust Matters a Lot

You can learn a lot from what’s said on a keynote stage—but you can also learn just as much from what isn’t. 

While the Databricks team spent a lot of time discussing the unification and democratization of the Data Intelligence Platform, they spent very little time discussing the interdependence of the data and AI systems that it aims to support.

Don’t get me wrong, it’s not that it wasn’t there—it was all over these announcements. But unlike previous years, the interdependence of these resources didn’t need to be stated—it was all but assumed. 

Gone are the days of discussing the data and AI workflows in a silo. The message from the Databricks stage was clear—data and AI are two halves of a whole; and they need to be managed and governed that way. 

Whether you build inside your lakehouse with a pre-baked suite of tools or get your kicks from hacking a stack together yourself, your agents and applications will only be as valuable as they are reliable—and that starts with managing each layer of that system end-to-end.

Deploying scalable reliability tooling that extends beyond just the data or the outputs to view the entire structure (data, system, code and model) as a single observable system—and enabling those tools with thoughtful incident management and governance processes—is the preeminent step to trustworthy data + AI products. 

Contact our team to find out how data + AI observability can help you deliver scalable, interoperable, and comprehensive reliability to your data + AI estate.

Our promise: we will show you the product.