Data + AI Summit 2025 | Keynote Analysis

Clips
News
More from Data + AI Summit 2025

George Gilbert

Principal Analyst, Data & AI

SiliconANGLE & theCUBE

Inside Databricks’ AI revolution: 10 key takeaways from the Data + AI Summit

Artificial intelligence is becoming more intuitive, interactive and embedded, not just in systems, but in the way companies build, think and create. From live coding and autonomous agents to unified governance and automation, Databricks Inc. is moving AI closer to the application layer and aligning it with developer intent.TheCUBE Research’s John Furrier and George Gilbert break down the keynote analysis during Databricks’ Data + AI Summit.“You sort of tell the AI engineer, the software developer, an agent, what you want to do,” said George Gilbert, principal analyst for theCUBE Research. “In the past, all it did was generate code

From lakehouse to live code: Databricks bets on integrated AI experiences — theCUBE analysis

Databricks Inc.’s vision for an enterprise world powered by artificial intelligence is based on vibes — or rather, vibe coding.At Databricks’ annual conference this week, the company announced a series of data and AI products on its open analytics platform, including the serverless Lakebase database, Agent Bricks for automating agent building, and tools that enable coding with AI through natural language prompts, also called vibe coding.The show floor at Databricks’ Data + AI Summit.“The apps themselves are getting much better,” said George Gilbert (pictured, right), principal analyst for theCUBE Research

Building for a new era: Databricks takes on pain points of complexity, lock-in and cost for enterprise AI

Ali Ghodsi, co-founder and chief executive of Databricks Inc., believes that data and artificial intelligence estates are fragmented, costly and much too complex. Today, he explained how his company’s flurry of announcements will address these issues and more.“It’s actually really hard still to succeed with data and AI,” Ghodsi (pictured) said during his keynote remarks today at the company’s Data+AI Summit in San Francisco. “It’s a complexity nightmare of high costs and proprietary lock-in. It’s slowing down the organizations. We think this is the biggest problem.”Part of the solution will involve AI agents and an ability to simplify their development for enterprise use

play_circle_outline Keynote featured Ali Ghodsi discussing Databricks's evolution toward an AI operating system.

play_circle_outline Overall keynote assessment given an A for relevance and content variety, but integration issues exist.

play_circle_outline Integration concerns remain, particularly with legacy applications and existing database estates.

play_circle_outline Databricks is positioning itself to compete in high-transaction environments like financial apps.

play_circle_outline Unity Catalog praised for multicloud capabilities and enhancing data governance across platforms.

Info
Transcript

Keynote Analysis

George Gilbert

Principal Analyst, Data & AI SiliconANGLE & theCUBE

TheCUBE’s John Furrier and George Gilbert unpack the biggest announcements from the Databricks Data + AI Summit 2025 keynote. They explore Databricks’ transformation into an AI operating system, spotlighting the launch of Lakebase and a deeper push into operational transactional databases.

The analysts examine Databricks’ growing AI stack, including the new Agent Bricks framework and support for the Model Context Protocol. Gilbert underscores how these innovations pave the way for agent-native development with built-in monitoring and enterprise-grade p... Read more

explore Keep Exploring

What are the recent developments and contributions made by Ali Ghodsi and the Databricks team in the AI market? add

What is the assessment of the platform evolution and what was observed during the presentation? add

What aspect of the keynote suggests that Databricks is focusing on integration? add

What are the implications of Databricks potentially challenging OLTP databases in the context of new applications and existing systems? add

What is your perspective on the use of Delta tables and Iceberg in multicloud and hybrid cloud environments, particularly regarding their integration and computational capabilities? add

bolt Powered by CUBE AI

Keynote Analysis

search

>> Hello, welcome to theCUBE here in San Francisco for live coverage all day today with Databricks's summit. It's the Data + AI Summit. I'm John Furrier, host of theCUBE. We are here also with our NYSE Wired team. Brian Baumann is also scouring the floor, getting some stories. And we're kicking off the keynote analysis with George Gilbert, our leading analyst, who's been tracking, now I'd say hardcore, for a long, long time, but over the past year and a half, the role of agents, digital twins, the role of this data layer we've been focused on at theCUBE Research and SiliconANGLE and theCUBE. We've been calling this, been watching it very closely because we knew this was going to be the enablement point for all the AI infrastructure that's emerging. And as the AI stack starts to emerge, platform engineering analytics merge, and in the middle of that is generative AI. George, great to have you on. Keynote does not disappoint. Ali Ghodsi and the Databricks team, every year they bring the A game, they bring new developments, some surprises. Ali Ghodsi being very humble and admitting he was wrong twice. I think that might be a record for Ali, but I mean, really good mojo. And again, Databricks continues to bring strong directional, I think, relevance to the AI market. I really love what they had. There's a lot of questions I have around how that's all going to work. We talked to all their customers around why they love Databricks, but now we're starting to see the picture, the models. It's not so much about what model here and there. We heard Ali Ghodsi talk about that, the intelligent platform, apps, Agent Bricks, which I loved. And then you had Dario come on from Anthropic and really kind of lay down some nice commentary around how he sees the AI going. And of course, they close it out with Jamie Dimon from JPMorgan Chase. Probably there's a deal in there around IPO or financing or a discount because he's a hard get. He delivered what I thought was a very pragmatic, intellectually honest speech, talking about how they're using AI, little political stories there, how he sees China. But overall, just a nice package for Databricks. Overall as a keynote, I give it a strong A in terms of content, the variety, but also relevance. Really interesting. They're evolving from a data warehouse, data lakehouse, I should say, leader. They disrupted the data warehouse. And where they've come from, they're now moving into an AI operating system kind of vibe. You're starting to see that those elements we've talked about, you've been covering and reporting with Dave Vellante on Breaking Analysis. This is now prime time for all the work you've been doing. So I have to ask you, what is your assessment of their platform evolution? What did you see on stage? What was your big takeaway from the Databricks's keynote?

George Gilbert

>> Three big topics. There was the Lakebase, which is their move into operational transactional databases, and we'll unpack that. Applications, which at first looked kind of like a janky bit of a shell, but there are a few things coming together that make it very impressive. And then finally, agents, which I took a look at what was existing pre-announcement and that was promising, but what they're rolling out, depending on the timeframe, was very impressive. So, let's unpack one by one. Lakebase is basically they took a PostgreS-compatible database that they were working on, that they also recently bought. It sounds like they bought something to accelerate the efforts that were in-house. And it's basically like Amazon Aurora, which was the big innovation where they made a cloud-native PostgreS, and in their case, also MySQL database. So they separated compute from storage so you can independently scale them, only I think the new wrinkle here is it can talk directly to openobjectstore. So you can get at the underlying tables and replicate them by directionally into and out of the lakehouse, and also-

>> It's a transactional database. This is-

George Gilbert

>> Yes, transactional database because that supports then operationalizing the analytic apps that they formerly built. And one other thing that was interesting is because it's this database is native to your development life cycle, so you have versioning and branching, an agent can create a database. You can see if the agent did the right thing. If not, you can kill it and go back to main branch. So that's useful now when we have agent-native development.

>> So is that LakeFlow, are they democratizing essentially ETL?

George Gilbert

>> No, no, no. They're probably going to talk about it tomorrow. This is the transactional underpinnings of extending what were formerly analytic apps. So now you can say the analytic informs, and then this is what do I do? This is operationalizing.

>> Ah, all right. Got it. That's lakehouse integration basically?

George Gilbert

>> Yes. With lakehouse, because that's the bidirectional synchronization. Then the apps themselves are getting much better. This was something Snowflake pioneered with container services and native apps so that your application inherits the governance of the underlying data and platform. But here, they go a step further. With the underlying Lakebase, they can support transactions. They have your choice of the popular, both Python-based and JavaScript-based UI frameworks. But what was really amazing was they co-opted the whole vibe coding, things like Windsurf and Replit and Cursor. They showed Supabase generating an app, sort of live-coding with a UI, with an underlying transactional database. It had a supply chain dashboard. So underneath it, it has all the components. It generates the DB SQL queries for the warehouse. It had underlying Genie for the metrics and dimensions that it's accessing, and then it deploys the whole package as a container. Now, I don't know how much of that is all shipping, but the point is this is coming together in a very powerful way.

>> So just to unpack that, first of all, they did comment on stage about the complexity of vibe coding into production, which is a whole nother discussion. But the fact that they're putting it into container, this essentially validates the analysis that you have had with Dave and I on front-end development becoming really an easy thing to do. Now the integration becomes a concern. What did you see in the keynote that gives you a promise that Databricks is all over this integration piece?

George Gilbert

>> Well, the integration, you're pointing at the one piece that still, it's a big piece that still needs to happen, which is your existing application estate, your operational ops where Genie starts to give you a harmonized view of the metrics and dimensions, but right now, the metrics are like individual metrics. It's not a whole data-estate view of all your metrics. So, that still has to come. And then talking to your existing application estate, we get into that with agents. It's not quite there yet.

>> All right, got it. All right. So what else do you got?

George Gilbert

>> So on agents, this is under the whole Mosaic AI umbrella.

>> Agent builder was the big news.

George Gilbert

>> Yeah, with-

>> I'm sorry, Agent Bricks.

George Gilbert

>> Agent Bricks.

>> Agent Bricks.

George Gilbert

>> So they put a whole bunch of pieces together that they had been working on last year where with agents, a hard part is you have to build an agent that learns because agents get better based on data and use. And so first, there's how do you perceive the state of the business? That's a work in progress with Genie. But then Agent Bricks was they have these building blocks where you say, "I want an information extraction agent," or, "knowledge assistant agent," or, "multi-agent supervisor," and then something like a quality that monitors those individual agents. The point is it's not 100% there, but they track all the traces of the agents in MLflow, and then they generate judges. The reason they need to generate judges, which is for evaluating the performance of the agent, it's very advanced because that's among the hardest things right now, which is the agents are probabilistic so you don't know, especially on an informational agent, how did it do? So they generate these LLM judges that say, "Okay. The answer was good on these aspects," and then they collect all the traces, lets them be like the data dog for agents, so they can observe every step of every agent operation and improve it.

>> Yeah, keep track of the agents. Are they hallucinating? Are they drifting? Are they accurate?

George Gilbert

>> That's exactly... They can track all that.

>> Pruning the agent dataset, I love that piece.

George Gilbert

>> Yes, yes.

>> That was a big one.

George Gilbert

>> But also, you can see the pieces of it coming together of an extremely advanced agent development kit. The one thing they added MCP support, Model Context Protocol, which is the wildly popular, it's like HTTP, HTTP equivalent, but for agents. And that starts to get you to your operational application estate. So like when SAP exposes and Salesforce expose parts of their apps as MCP servers, the agents can talk to them.

>> It wasn't clear to me MCP's commitment. They announced support of it. They weren't specific. Did you hear anything in the analyst side?

George Gilbert

>> They did say the agent development tool can talk to MCP servers. So like any tool, it's rather than hardwiring each tool into your agent, you talk through MCP and it understands how to talk to the tools.

>> I think MCP was the biggest surprise in the past between last year and this year, emerging as that kind of rallying point where people can have an MCP node and have interactions with other agents.

George Gilbert

>> Right, agents and existing applications and existing tools.

>> Yep, got it. All right. So I have to ask you on the... Do you really think that Databricks is going to challenge the OLTP databases in this AI market? Because OLTP was usually restricted to financial transactions, but it's a high-transaction environment. Think micropayments back in the old days. There will be a lot of transactions going on between agents. It needs an OLTP layer. What is your thoughts on this? Do they get it right? Will they be the true system of record for OLTP or do they have to be?

George Gilbert

>> Well, for new applications, they have something that's very promising, like low latency, high concurrency because they separated compute from storage. That's why they paid $2 billion for Neon. So, for new applications. But remember we've got 50 years of applications that it's our sediment. Those all have databases that you're going to need to talk to, not just the databases, but the applications.

>> What is your view on the migration? Because one of the things we're seeing a lot of action on is, one, AI net right now is with all the coding capabilities they mentioned, Anthropic's got a great one. The tooling is now highly agile, customizable. So we're seeing a lot of people talking about migration from databases and then moving into Unity as like this open system of record, system of AI record, because you can run your data anywhere and still use your own AI and ML. So what does that piece mean? What does this-

George Gilbert

>> So we've been talking for a long time about the need for analytics and for AI to talk to a harmonized data state because the data and the artifacts, like your ML models, even your agents, they all need to understand in a common language what your data and application estate looks like. Unity starts to up-level that and to give you common governance.

>> How do you see the traction for Unity Catalog? I talked to a lot of their customers. They love Unity. They can store their data anywhere, still take advantage of the ML. Do you see it in a multicloud environment and hybrid cloud environment? What's your analysis on that?

George Gilbert

>> I think yes, multicloud, and certainly for all your Delta tables, and they're working on converging with Iceberg, but we need to understand better. There's a lot of work to be done to converge those. And so yes, you can probably track Iceberg tables in Unity, but can you run all your compute jobs, your Databricks native compute jobs on Iceberg? Not 100% sure yet.

>> So I have to ask you, you've been tracking the momentum of both Snowflake and Databricks as well as a variety of other factors in the market. How would you describe the platform evolution of Databricks from a progress standpoint? Obviously, I gave it a high mark on the A for the keynote, not an A+. No one really gets an A+ unless they really smash it. What is your view on their platform evolution? They want to build a trusted AI-native platform for both developers and enterprises. Clearly an enterprise play. How do you see their platform evolution? Is it enterprise-ready? What's your view on their evolution? Give them a mark where the progress is.

George Gilbert

>> It's really impressive how they can buy companies for technology and rapidly roll it in to extend their footprint. Like when they bought Mosaic, I don't know, a year and a half ago, or was it some-

>> Yeah, yeah. About a year ago.

George Gilbert

>> A year ago?

>> A year and a half ago, yeah.

George Gilbert

>> So now the whole agent development platform is growing out of what they did with Mosaic. But by talking to Genie, you can understand your data really well, and then by talking to Neon, you can operationalize the transactions. But again, this is for new apps. They still have to do work to get to your existing apps.

>> I like this agent-first strategy. I mean, that's the word that's been kicked around, but Mosaic brought a lot of that mojo. How do you view the open-governed interoperable side? Because Iceberg, I wrote a couple posts up there, got some comments, but people are like, "It's not just about Iceberg." Obviously, they have Unity. It's pretty clear Iceberg is becoming the de facto format. Pretty clear there. No debate. But how does Iceberg fit in? Because you got Lakebase, you got Iceberg, you got Lakebridge coming on migrations, so you got all this now, plumbing, connective tissue between performance from the heart infrastructure, and now this layer. It's perfect for what you've been covering.

George Gilbert

>> So take Lakebase, the operational database, the transactional database they just bought. That will replicate into Lakehouse tables. It's not clear. They didn't say whether it's Delta tables or Iceberg. You certainly could read Iceberg if it's replicating in on Delta. But the point is they're building this platform that takes your operational database needs, it's getting very strong on the analytic format where Snowflake was very strong for your dashboards. The Genie talks to your data. The presentation part is still immature, but the structuring of the data to present it underneath, it's getting stronger.

>> Yeah. I love the platform evolution of Databricks. Go back 15 years when we started working together on big data with Hadoop and that's how they all started, from Spark, and we've been covering their journey. We used to riff on theCUBE and off privately, like, "There's going to be a data SRE-like role, large-scale data, and then developing on top of data." The apps thing got my attention because to me what that signifies is that the vibe coding kind of direction will be a rapid runtime development of apps where the data will be generated, the apps will be generated, but it'll have to be powered under the covers by a highly robust, highly scalable, low-latency data infrastructure layer, which we all see that. So, connect the dots.

George Gilbert

>> So the vibe coding, so something like a cursor or they showed Supabase where you sort of tell the AI engineer, the software development agent, what you want to do. In the past, all it did was generate code. But if it's now aware of what's inside Unity, it can talk to the Databricks tools, it can talk to Databricks the way you've modeled your AI, your internal data, and it can use that, and it can talk to then also as an MCP server, I believe Unity might be an MCP server, but certainly anything that was within the Mosaic Agent Bricks. The point is the vibe coding is starting to be an umbrella over the Databricks data and tools and that's when all the pieces come together and they become very, very powerful.

>> Yeah. I think the big bet obviously with Databricks is AI-first, a new class of a developer/data engineer, developer, enabler, whatever you want to call it, but the world's changing. And then the legacy abstraction, so legacy doesn't go away, whether it's legacy or modern legacy, as we call it. That'll be abstracted away with connectors, right?

George Gilbert

>> Yeah.

>> So you're going to have a lot of connective tissue between the legacy and then this new era of development.

George Gilbert

>> Yeah. So you put your finger on the biggest open piece, which is the connecting to legacy. So right now, lots of vendors have connectors and so you could put a MCP server in front of the connector to make it sort of easy to expose to an agent or an agent development tool, but still, that means each connector is its own language. So you have a Tower of Babel of hundreds or thousands, maybe, connectors. At some point, we need to abstract that. That's the digital twin we've been talking about. That makes it... Then you really are leveraging your entire estate.

>> If you're watching this video, you got to go to thecuberesearch.com. Check out the work George and the team are doing. Amazing work. Dave Vellante, George, myself, Rob Strechay, the whole team, a lot of digging in this area and we've been working on it for a while. It's now coming to fruit. Final question as we wrap up, obviously I love the demos. Clearly, you have two major areas we've been seeing a lot of advancements on with this first gen of AI value, coding and sales and marketing. Why? Because the data's all out there and no one's going to get hurt from sales and marketing. But the analytics business, so you now have the previous analytics teams, analytics databases. So you have coding, and sales and marketing, but then this analytics, these KPI, metrics are talked about. So you got the convergence of analytics, which is a data practice, and now these new areas of value creation. Where's the value extraction for the enterprise? Because in the demo, they were very clear, like, "Hey, vibe coding is great. We're doing it. Let's ," but now getting into production is where the enterprises are now. So the enterprises will be the value extractors, as well as some creation, but for the most part, the goodness that comes out of Databricks and others is the value that they're creating or developers. How will the enterprise extract the value in your opinion?

George Gilbert

>> Okay. So Databricks, the pieces are coming together, we got to see when they're being delivered. But with the vibe coding, it makes it vastly easier to start coding. But the key with the vibe coding is can it use the Databricks tools? So therefore, can it read data that you've modeled in Genie? And then can it operationalize? It informs the decision. Then can the agent take action? You've got the transactional database to take action, but right now, that's for new apps. That missing piece was how do you talk to existing apps? And I... Thanks, John.

>> All right. CUBE coverage after this short break.