Fireside Chat: A Practical Conversation on AI, Data Stacks and What Matters for the Next 24 Months

Summary

This video is a no-slides fireside chat between Databricks SVP of Product David Meyer and zeb.co VP of Data and AI Architecture Josue Bogran, covering honest, unfiltered perspectives on what is genuinely changing in enterprise data stacks and AI adoption over the next 24 months.
The conversation addresses why traditional BI tools like Tableau and PowerBI create semantic latency, copy problems, and governance debt compared to shifting business semantics directly into the Lakehouse using AI/BI and Genie.
Lakebase—Databricks' instant database branching capability for OLTP workloads—is discussed as a new category where David Meyer argues that every new database created going forward will inevitably become a Lakebase.

Fireside Chat: A Practical Conversation on AI, Data Stacks and What Matters for the Next 24 Months

There is no shortage of AI predictions. There is a critical shortage of honest conversations about what enterprise leaders should actually do next. In this no-slides, no-hype fireside chat, Databricks SVP of Product David Meyer and zeb.co VP of Data + AI Architecture Josue Bogran — one of the most followed voices in the Databricks community — dissect what is genuinely changing in data stacks and enterprise AI adoption over the next 24 months. They'll surface the most persistent misconceptions leaders have about the modern data stack, reveal what separates organizations generating real returns from those still in experimentation mode, and share candid insight from Databricks' product roadmap. Questions are shaped by the community in advance. This is the session for leaders who want an unfiltered signal, not vendor positioning.
Talk By: David Meyer, SVP, Product Management, Databricks (E-Staff Care Internal Speaker); Josue Bogran, VP of Data + AI Architecture, zeb.co ;

Chapters

FAQs

What separates organizations generating real AI returns from those still in experimentation mode?

According to this fireside chat, the key differentiator is having a unified data foundation that eliminates the latency, copy proliferation, and governance debt created by traditional BI tool architectures. Organizations generating real returns have shifted their semantic layer and business context directly into the Lakehouse rather than maintaining it in separate BI tools downstream.

Why does Databricks argue that traditional BI tools like Tableau and PowerBI create problems?

Traditional BI tools require semantic definitions—metrics, hierarchies, business logic—to be maintained in a layer separate from the data itself, which creates latency between data updates and dashboard freshness, multiplies data copies, and introduces governance complexity. The discussion in this video frames AI/BI and Genie as a way to keep semantic context close to the governed data in the Lakehouse.

What is Lakebase and why does David Meyer believe every new database will become one?

Lakebase is Databricks' managed PostgreSQL offering with instant database branching capabilities, designed for OLTP workloads that need transactional consistency alongside Lakehouse analytics. David Meyer argues that the combination of Lakehouse governance, AI integration, and instant branching makes Lakebase the inevitable architecture for any new transactional database being built today.

How does Databricks think about the relationship between the platform and AI model providers?

The fireside chat describes a durable separation of concerns between Databricks as a data and AI platform and the model providers that supply foundation models. Rather than competing with model providers, Databricks focuses on providing the data layer, governance, and routing intelligence—including tokenomics for directing AI queries to the right model for value maximization.

Full transcript

[00:09] Hi everyone. Thank you so much for joining. I hope you've all had a wonderful morning with the keynote and you're having a great start to the summit this week. Uh so my name is Brier Cleavvens. I'm an account executive here at Data Bricks from Australia if you
[00:21] haven't picked up on the accent already. Um but super excited to be hosting the data strategy track in this room today and tomorrow. So, first up this morning, we've got a fantastic fireside chat with David and Josh. So, um, David is our
[00:35] senior vice pres president of product management at data bicks where he leads product strategy and execution and Josh Bogrin is the VP of data and AI architecture at Zeb and one of the most followed voices in the data bicks
[00:48] community. So Josh is also a data bicks MVP and product advisory member as well as an adviser to product companies. So in this no slide fireside chat today, um
[01:00] we've got David and Josh who are going to be talking about what's changing in data stacks and enterprise AI adoption over the next two years. And they're going to be digging into the misconceptions that leaders uh hold
[01:12] about the modern data stack, what separates organizations seeing real returns from those still experimenting, and sharing some candidates from the product roadmap. And a lot of the questions that Josh will be stepping through today are actually shaped by the
[01:24] data bricks community that he's part of. So we've got 25 minutes of the fireside chat and then we'll be opening up to the audience um for Q&A. Uh when we get to the Q&A, what I will do is I'll ask um everyone who has a question just to line
[01:37] up at the two mics um that we've got um either side of the room just so I'm not running around everywhere. Um but with that, Josh and David, I'll hand over to you for the fireside chat. >> Thank you very much for that. Well,
[01:49] thank you so much everyone for being here with us. Uh, it is an honor that this is probably the first session after the >> Should we say which one of us is which? >> I'm David Josh. >> Yeah, that sounds
[02:01] >> exactly. Don't ignore the shirt. So, no, I'm Jose and uh and this is David. Um, here's the I have some bad news and some good news for you. The bad news is that
[02:13] our slides are not up. The good news is we don't have any slides. And you're probably going to go through a lot of those today. So hopefully you enjoy this. This session is only going to be a success if you participate. So I'll ask
[02:27] some questions here and there. Make sure that again at the end we're going to have space for questions. Feel free to drill him. Don't drill me. Drill him. Uh but uh yeah. So thank you so much for being here. Um so two quick takeaways
[02:41] too. I mean two key things. I want you to go go out from here having a practical understanding of data bricks's road map. Second, I want you to be able to go back
[02:54] home or go back to work with perhaps the answer or an idea on how you're going to be able to solve some problem or you know whether strategic or technical
[03:06] problem in your organization. So with that said um I'm gonna start off with a question. How many of you here use PowerBI?
[03:22] >> Just a few of you. Uh, how many of you use Tableau? >> So, I think those are great tools. >> What What about both? >> Sure. Both. You use both. Um, so but then the next question is for
[03:35] you is what if data bricks had something that was good enough that you didn't have to pay any licensing fees? Would you like that? >> Can I get that? >> Only one person.
[03:47] >> Okay, there we go. >> I mean, look, you got it's it's joyful to spend money. >> You know, we're consumers. >> Maybe if you're B Lei or something. But yeah. All right, so with that said, here's our first question. I'm sure a
[04:00] lot of you have seen AIBI dashboards, Genie or whatever data bricks renames it to next. Um, so I really want to understand uh David, you know, historically uh data bricks was all
[04:12] about data engineering and AI or ML. Now you've moved a lot more towards the the BI layer and obviously you know the chat conversational experience and all that. Those are very two different things.
[04:25] >> Yeah. How should customers think about what is moving into data bricks versus what's should stay or is going to stay longterm within PowerBI, Tableau,
[04:38] Excel or Google Sheets. Google Sheets. So yeah, I mean I think there's that's we could talk an hour on that topic. Um the uh the the the goal of everything
[04:51] data bicks has done since founding is to try to simplify the data and AI stack. So the advent of the lakehouse um was to prevent copying the data
[05:05] cleansing copying the data to a warehouse in order to do analytics. That added a second copy a second governance scheme latency cost. So being able to do that directly on the data fundamentally
[05:18] simpler the data warehouse the dedicated data warehouse could go poof and you could have strictly better performance, cost, latency, simplicity, governance. Um all these other things we talk about
[05:32] and like Bill was on stage and tried to make a bunch of things disappear was to simplify so it's less brittle, it's easier to manage and you can get to your data and AI goals easier. So traditional
[05:46] BI now all these BI vendors PowerBI Tableau have evolved into a lot of modern things but traditional BI had the same downsides as traditional data warehousing. It was a copy of the data.
[06:00] You had to do import to make it go fast, right? There was latency because it wasn't operating on the live data. There was separated governance. You had to replicate the governance in the BI tool.
[06:12] Now, we love PowerBI, we love Tableau. These are incredible ways for a lot of people that are familiar with that form factor to get access to the data, but the entire industry and AI is
[06:24] accelerating it has been shifting left. Your semantics, if they're trapped in BI can't be used for AI. You need to shift these left into the data platform
[06:36] oftentimes in justest decorating things with AI. um capturing the ontology through all the data processing to be used by AI. So these same things have our customers saying you know here's the
[06:48] funny thing so data bricks in 2013 to 2015 1617 what was the the primary interaction it was a notebook it was the first collaborative notebook in the
[07:00] cloud to do big data processing that was what data bricks was people created BI dashboards in notebooks because the data was there and these
[07:12] dashboards sucked and they were hard to write and they were hard to share and they couldn't be run concurrently for lots of users and our customers were like please please you have this SQL
[07:24] engine this warehouse now give us an ability to manifest these dashboards there shift left and we follow where our customers ask us to go so I think that you will always have these curated
[07:38] dashboards rigid or not in these tools because they're useful they're useful for parts of the population, but there's new use cases that they're not as useful for because they're a copy, there's
[07:52] latency, they're somewhat rigid with all the DAX queries behind them. And that's where we're seeing AIBI dashboards emerge. So, are we going to replace BI? I think you will replace BI because
[08:06] you'll expect a different modality. you saw in um it was Ken's demo um I think or Alisa's demo where Jeanie was just alerting her to something that was
[08:19] happening. She didn't create a dashboard in AIBI. Something happened in the operations of the company and it says hey Elise you
[08:31] need to know about this. So, I think that's where we're getting to. Instead of, you know, having to ask the question, that's already a failure. I refresh my dashboard every morning. That's already a failure. You need to be
[08:44] alerted to what the system knows you need to know. So, I mean, I think all of these things will exist and there'll just be lots more users and lots more agents engaging with data on their own
[08:57] terms. And that is what's going to drive people to these new modalities. >> But >> I didn't answer your question. >> Well, I was going to say, but if you could, you know, if is your goal that
[09:09] every single person here becomes a AIBI dashboards user and the other >> No, I don't want anyone to be dashboard users. I want every person here to be connected to the umbilicus of Genie and
[09:23] having data fed to them like nutrients when they need it most. I did not expect that analogy, but uh you know, moving on. But no, that's a great
[09:35] answer. Um I'm gonna ask a question on lakebase now. First off, how many of you know what lakebase is? >> Okay. So, how many people have used
[09:49] lakebase and branched a database with lakebase? The rest of you are lying. You don't know what it is. >> It's not. Um so just just to set the stage here for this question, linkbase is data bricks is transactional database
[10:04] uhp based on postgress from the neon acquisition. So for applications basically or agents. So with that said, as Lakebase expands data bricks into
[10:18] operational Postgress workloads, what workloads is linkbase ideal for today? >> Yeah. >> And what do you expect it to become ideal for in the next two years?
[10:32] >> Okay. So all my questions are going to have the same answer. So the goal is to simplify data and AI, right? Um Ali talked about and Reynold
[10:47] talked about the unification of OLAP and OLTP in a singular columner data structure which is pretty radical. I don't really believe it yet. I haven't used it much.
[11:00] It's brand new so I don't expect you to believe it yet. Um, when we talked about Lakehouse in 2019, everybody thought it was a beautiful idea, but it was technically unachievable
[11:12] and we were like the mother of all anti-atterns and stuff. And now everybody has a lakehouse. So, I think that the the concept of a lake base, let me just break it down really quickly. If you let think of your imagine in your
[11:25] mind your favorite database, the database you love that you ask all your questions of how long will it take to clone that database? If you need to make a clone of
[11:38] it to do dev on or to run a test on, you you're lying to yourself. If it's if it's less than tens of minutes, it's probably hours. Okay, that means when you're going to write test code, you're
[11:50] going to write a off old data, old test data, synthetic data, etc. because you just can't have these clones and these clones are expensive and they take, you know, they to run 247 and and all of
[12:02] that stuff. So, because we've separated compute and storage in Lakebase, you can literally branch in less than a second and the branch is free.
[12:15] Okay, only, you know, it has all the data in production, but it doesn't move any of the data. there's just a bucket over here and it does, you know, new stuff. It writes to the new bucket and then emerges it on read. Okay, so long story
[12:28] short, it's fundamentally new. Every branch of your code can have a database branch of production data totally safe without impacting the production system. It's a fundamentally
[12:40] whether or not you know it, you all have hated databases with a passion inside you. You have this this disgust of them because you can branch your code. You love git but databases don't abide by
[12:53] the same inner dev workflow. Now they do and that's why I said if you haven't branched a database you don't understand them because you haven't felt it. Okay I'm going to get to the answer but but I want to I want to quickly just to challenge I just want to make sure that your assumption is correct. Who here
[13:07] does software development? Who which of you has been happy with uh your your how your database has worked >> like keeping your production your deaf
[13:21] environments. >> Okay. >> Yeah. It hasn't just >> to the new the new world of work especially the work with agents and of course with lake base and neon we see 90% 90% of the databases created are not
[13:33] created by humans they're created by agents at the behest of humans. Okay. So um anyway so this new category just like in 2019 lake houses were ludicrous now
[13:48] every analytics platform is a lakehouse today lake bases are ludicrous you see you know Oracle posting hate articles about them >> which I think is pretty awesome
[14:00] >> don't get to my next question which is are you going to replace Oracle >> well so so in in four or five years every database will be a lake base. It just it's it's strictly better. It's
[14:12] paro optimal. Okay. Now, it will be us. Every vendor will be doing it. It won't just be us with our crazy neon lake base. Um I don't know if people will adopt the term, but every database will operate this way. Oracle is an amazing
[14:26] database company. There's a lot of amazing database companies. And here's the trick. Nobody moves databases. If your database is working, nobody's going to move it. But tomorrow is much bigger than yesterday. There's going to be more
[14:39] code written this year than in history, right? And there's going to be more databases created in the next couple years than in history, probably in the next couple months. Agents are spinning out these things like crazy.
[14:53] So, you're not going to move databases, but new databases, you'll learn as soon as you try it. They need to have the separated compute and storage. It's strictly better. And they will be lake bases. Some fragment of them will be on
[15:06] the data bricks infrastructure in neon but all other databases will become lake lake lake lake lake lake lake lake lake lake lake lake bases just like all other analytics platforms became lakeous I'm going to move a little bit from data
[15:21] bricks specific products but just to say you know a lot of enterprises are investing in data bricks while also seeing rapid progress from model
[15:33] providers like OpenAI and and others. >> Where do you see that durable separation between data bricks and the model companies and where should customers
[15:46] expect overlap? Okay, so AI is amazing. You know, you can vibe code pixel perfect things and they work. Um, I guarantee you if you call up one of
[16:00] these companies and say, "Can your forward deployed engineers replace my data bricks?" What do you think their answer will be?
[16:12] Come on. Yes. Four deployed engineers can do anything. I >> was going to say I I'm I'm in a consulting company. Just a couple hundred bucks, hundred grand and you know, sure. >> But like here's the here's the trick. We open source it all anyway. That's why
[16:25] the AI agents are so good at this stuff. You know what's hard? You know what's hard? Running it at scale in a durable way. They could create a point in time simulcum of our data platform that would
[16:38] work when you tried it, but no one wants to be at the pager duty back end of that thing. So you know the open core of data bricks is available for you to run on your laptop but you don't want to run
[16:51] unity catalog for you know thousands and thousands of machines and and millions of users and all that. You don't want to do it. It's a it's hard just like when Spark was created data bicks was the
[17:04] company that ran Spark. Spark was open source but no one wanted this thing's hard to run at scale. So I mean um the AI uh all these companies are great.
[17:16] Open AI is great. Anthropic is great. You know SpaceX and all of these um I'm not I just need this is just like you know a security blanket. I'm not going to use my pen. Um but the uh the key is
[17:30] we're a great partnership, right? Because these incredible models need access to data. You need it to be governed and bulletproof. You know, we spin up 50 million machines a day to crunch data. We process uh hundreds of
[17:44] trillions of tokens a month with all of the leading AI vendors because you need that orchestration machine to be run by people that aren't you with pager duty people that aren't you if you're going
[17:56] to trust your whole business on it. I have a probably a probably the toughest question that I have here today. All right.
[18:08] So obviously data bricks benefits from the whole AI everything but you know we've seen that every new AI model that comes out
[18:21] it's more expensive more powerful but more expensive more expensive bills keep going up obviously that's why you even came out with AI gateway >> but the question is though and I think a lot of people probably ask themselves this question is
[18:33] >> what happens to organizations that lean heavily on AI when AI becomes less and less subsidized by venture capital.
[18:45] >> Yeah. So, this is where I think the story is actually really exciting. There's lots of fear in the news, people running through their AI budgets, but here's the thing. Um, so, you know, I
[18:58] use Claude as a consumer a lot. I love Claude. Um, and I can choose my models and I always have it set on the like most elite modern model, the most token spendy model. And I ask the dumbest
[19:12] questions, you know, I ask questions that, you know, I could answer. I put it in the in the, you know, like math questions, not really math questions, but things I could easily look up, you know. Um, it's human nature to think you
[19:26] need the the latest model. It's engineers nature to think they need the latest model. Um something I think we're not going to demo it till tomorrow, but in this AI gateway we have it evaluates
[19:38] the complexity of the query and can route it to the appropriate complexity model. Okay? So I guarantee you don't need opus for half the stuff you're
[19:50] asking opus. You just don't. Right? and you can route it to newer models and we're going to announce some tomorrow that are very fast, very efficient with tokens. Okay. So, what's going to happen
[20:02] is the customers that we we have prototypes of this stuff that we've had for a while and now we're going to be announcing them and releasing them tomorrow. But we have a lot of our customers and this general field I like to call tokconomics. It's economics of
[20:15] tokens and making sure the marginal utility of each token you spend is optimized based on the need of the query going to it. um they have their token usage going up exponentially crazy and
[20:29] their cost graph totally flat totally flat because just as these expensive models are coming out a lot of these you know distilled models other models are incredibly inexpensive and
[20:42] they're only about four months behind right so the best thing you can use today will be available in four months for cheap to route the lower complexity questions to so I think um you're
[20:55] Everybody's going to have to put together the infrastructure to manage their tokconomics, to manage the costs in their company to make sure they can unleash the AI and not fall behind. >> Value maxing.
[21:07] >> Value maxing. Yeah, that's what Patrick called it yesterday. The co-founder uh Patrick Wendell in PAB instead of token maxing, value maxing. And I think that now with the what everything that Unity
[21:19] catalog and the AI gateway sees captures all this telemetry that allows it to be the most powerful value maxing engine for you whether your AI is inside of data bricks or just outside. I can use
[21:32] you know cursor and route everything through AI gateway and all the models I'm using and value max. Now I'll ask one more question before we have the mics open. So by the way if
[21:45] anyone wants to ask a question cur you to be lined up on the mic. Um the last question is you know there is a lot of noise around
[21:59] AI replacing everything right replacing humans uh SAS applications and entire workflows. Um
[22:11] from what you see with real enterprises just so you guys know David is very customerf facing so he spends a lot of time interacting with very large enterprises. So from what you see with
[22:23] real enterprises what is actually being replaced >> what is being augmented and what is mostly fantasy that sometimes gets you know just uh puffed up in the news.
[22:37] Yeah. I mean, it's interesting the uh it used to be like I have I have two kids. One's 21, one's 18, and I used to give them advice. It was silly. Um but I
[22:49] used to be like, "Learn the hard skills, you know, learn the hard skills and then you'll be able to apply them to all these different walks of life. Like be an engineer like your dad." Neither of them chose that path. Um but, uh you
[23:03] thought engineering was hard. I mean, that's kind of been the cultural understanding of it. Um, but it's but AI is ripped through coding, ripped through engineering. And why is that?
[23:15] Because it's provably correct. You have test suites, test cases, etc. And the rest of the the company has a lot more ambiguity in it, you know, and that's a lot of what Ali was talking about today.
[23:27] getting all that more ambiguity context understood so these agents can do the work you need to do as co-workers. Um so I really have no idea what the answer to
[23:39] your question is. Um but I think that every organization is going to figure out how to reorganize themselves. The first thing every HR department is doing now,
[23:53] their interview process in every company of every HR department is different today than it was three months ago. I guarantee it. They're looking for different kinds of people. They can understand how to leverage these things
[24:06] and instead of being everything being blocked by humans and wanting human approval of everything, figuring out how to replace that as skills so they can, you know, leverage AI to make their
[24:19] whole business move faster. So, it was uh that was a total non-answer to your question. I just realized, >> you know, uh want to try it again? You got 30 seconds to try it again. Give give a >> Well, say the question again. What
[24:32] what's fantasy and what's real? >> Like they they go they go back to to their work today. >> What's real is is
[24:44] >> mundane things. What makes something mundane? You get up in the morning, you have a plan for your day and a lot of your day sucks and some of your day is awesome. That's work, right? Um the
[24:56] stuff that sucks is mundane. Meaning you don't need to be there. I don't need to be clicking on all this crap and all these that goes away, right? Um but what will never go away uh and I think the
[25:10] Pope recently said it well um is uh the human ingenuity thinking, strategy, the true driving of things will require
[25:23] humans to be orchestrating it because that's not generative in a predictive sense. that's generative in a creative sense and I think that's uh something that I think will always have humans driving.
[25:36] >> I would say my my simple answer to this would be that what is not going to be uh to be replaced easily is going to be trust. >> Whether that is trust in platform I just made it up.
[25:49] Whether it's trust in a platform or trust in a human >> uh you know again that's seems very altruistic but I think ultimately building making yourself or your company
[26:02] have a reputation that you will get it right. I think that's the safest thing you can do and the most ambitious thing you can do uh you know to survive all the AI hype you know. So that that's my
[26:17] thought. You can use that. >> I like that trust. >> Thank you. Thank you so much everyone. I hope you found good value in this session. Uh and
[26:29] If if u I've done a lot of content too. I've interviewed David a few times. Mate, Rainald, and Ali. Uh feel free to connect. U you know happy to talk to you too. So thank you so much. Thank you.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.