Confluent, Inc. (CFLT) Earnings Call Transcript & Summary

Research there real quick. But now the main event, we'll have our guest of honor, Jay Kreps, Co-Founder and CEO of Confluent, joined by -- joined with him is Matt Hedberg, our Global Head of Technology Research, and Matt played a significant role in the Imagin report. So thank you.

Thanks, Mark. All right. Here we are. It's a great group here. Thanks for doing this, Jay.

Yes, happy to be here.

And then Shane, Shane is around here somewhere. I don't know where Shane is. Thank you, Shane, wherever you are.

He is getting lunch. Yes. He's got lunch. Yes. I don't know we got to somehow get lunch on demand after that. That's good.

Well, thanks, everybody, for doing this. We have 41 minutes. I've got probably 2 hours of questions, but we will -- there's going to be some mics around here. So I do want to leave an opportunity for questions. Jay is -- he's an incredible founder, visionary tech enthusiast. He obviously is the CEO and Founder of Confluent. He also sits on a number of boards. And so just as sort of a visionary as we think about the future and it is dovetailing off of our Imagine report, I think it's a great opportunity to kind of talk about the future, talk about AI, talk about Confluent and really talk about the super cycle that we're seemingly is upon us. So Jay, thanks for joining us.

Yes. Yes, excited to be here.

So let's start with the obvious elephant in the room here, this AI super cycle that we're in here. You've seen a lot of technology cycles in your career. Both of us. We don't have a lot of hair. So I think we probably both have seen similar cycles over the years. But AI is...

Each one needs you know...

Yes, we need the hair upgrade cycle. AI is causing a compute upgrade cycle. And a lot of people in the world think that we're -- the world is short compute. So I guess just when you sit here today, how do you compare this AI cycle to prior cycles that you've seen? Let's start there.

Yes. I mean -- so first of all, I think it's significantly bigger. Most of what we would talk about as cycles or cycles really within a single paradigm of computing. And I think this is kind of opening up the other side. If you think about all the problems we could solve with computers, we had a way of solving problems with very precise business logic and rules and carrying that out. We've got better and better at stacking those on top of each other. But this is really opening up a much broader set of problems that we couldn't address with technology or software before. And now we're increasingly able to do that. There's always questions about the pace of that progress. And I think that is an open question. What we've seen so far is actually waves of excitement and concern and excitement and concern. But if you look at the actual progress on AI benchmarks, it's been extremely consistent -- there's a line going up. A line going up doesn't mean it goes up forever. But nonetheless, your best prognostication of where the next point will fall is probably on that line. There's an open question about the rate of economic value unlock relative to just creating intelligence. I think that's a very open question, right? But nonetheless, I think the hype is justified, right? This is a very big deal.

So doing -- playing the compare and contrast game, if you think about the Internet, which was arguably -- and mobile, we've seen that significant trends over the last several decades. Where are we at in this AI cycle? I mean I've heard people say maybe we're like 1997, 1998. Like where are we in this cycle from your perspective?

Yes. I mean we're still very early. A lot of it -- this one is weird because it really hinges on this kind of effectively research progress on the problem of intelligence. As I said, the line is going up. And the assumption is it will continue. That has to be the most likely outcome you would predict, but it's not the only outcome, right? If you look at value realization, it's still very early, right? We're creating an artifact, which is interesting. There's clearly a consumer use case, which has taken off and it's answering interesting questions for people. the enterprise use cases have started to move. And in some areas, there's very big movement. I would look at the progress in coding as an area where there's clearly -- this is going to be something that's universally adopted and a very deep value, right? So like doing a lot of the work, and you can see the progress in that domain very rapidly. And so I think that's -- in many ways, if you're looking for whether it's the kind of leading indicator or the exemplar or the canary in the coal mine, depending on how you think about it, I do think coding is like one type of knowledge work that's being by AI very fast. And so I think that's an interesting phenomenon. It's certainly an interesting one for me since I kind of came up as a software engineer and worked in that industry and kind of watch it happen. So for me, it's certainly possible to appreciate the impact maybe more than if I was looking at progress in drug discovery or something I know nothing about where maybe from the outside, you might not know exactly what's happening.

Sure. So we're not here to answer the question about if we're in an AI bubble and where we are in the cycle. But presumably, in your answer, if you're starting to see some early production workloads from an AI perspective and you talked about code suggestion and there's some other evidence of AI benefiting the economy. Presumably, we're still pretty early in the cycle. And I would imagine that as the years progress, we'll see more evidence of this. But any kind of additional thoughts on that?

Yes, I think that's right. I mean I think my observation with kind of big general purpose technologies of any kind is that we kind of underestimate, if anything, the impact they can have and then we underestimate how long it will take to fully realize the impact. So that was certainly the case with the Internet. There's a certain amount of just having it play out. This may be a bit faster, but it's still -- it's not months. So -- and then yes, where are we at kind of day-to-day and different types of work? We're very early, very early. And enterprises are very early in the cycle of figuring out how to use this stuff. And so I think that's probably -- depending on how you think about it, I think it's probably good news, like we'll see a lot of change that comes out of that. But there's clearly like a very major new capability that's being introduced that's going to impact how companies operate, I think, across the board.

Yes. This is a TIMT conference. So I'm a software analyst, but AI has impacted other aspects of TIMT differently than software. We've been under this death of software narrative from an AI perspective now for seemingly a couple of years now. You guys have a different model. It's not a seat-based model, so you don't -- you're not under some of that similar pressure. But just from your seat, how do you see AI changing software in the future? What are some of the bigger change factors that we need to be aware of?

Yes. Yes. I think I've heard different aspects of this. First of all, I would say the most obvious near-term implication that you would see today out of what's happening is there's going to be a lot more software, right? So like all this AI coding is generating applications. There's going to be a lot of them. And the impact of that may be different in different areas. But certainly, we kind of sit in the data layer and enabling that, we feel like, that's certainly for the foreseeable future, a very powerful trend for us. I do think that there are some nuances that companies will have to get right to navigate this. The challenge to seat-based licensing, I think that's real, but it's kind of not fundamental. I think you definitely need some kind of consumption pricing for AI because it's very expensive, right? So you can't just tag it in like the seat-based model kind of assumes effectively a fixed cost of your software and then you're effectively pricing to the number of people that you're providing value to, but the actual cost of serving a seat is negligible. For many of these AI workloads, especially as AI is taking on some background tasks where it's doing work like all the time, it's a very expensive proposition, which is not going to be just baked in for free. It's going to have to show up in the pricing model. But if you think about that, I don't think that that's a fundamental challenge for those kind of...

Or seat-based models.

Yes. I mean, like, look, there's always nuances in pricing to capture value when companies evolve. It just -- as long as there is value to capture, companies will figure out how to price against it. So it's a change, but I just -- people have brought that up to me as a kind of fundamental threat. It just doesn't see.

We don't see it as -- it's not as high.

No. I think a more fundamental challenge is I do think we've conceived software applications as being primarily these little islands of UI. And if you think about how these systems are going to work together, that's become less true over the years and AI is probably making it even less true, right, where I do think the access to the data and APIs that drive the functionality is going to be as important as the thing you see on your phone or web browser, right? And so to the extent that companies primary moat and stickiness is humans being very familiar with clicking certain things in the UI, I would say that, that moat is less effective, right, in a world where there's a fair amount of kind of gentic AI happening. We're not in that world yet, but that's kind of the direction. So I do think you could see that as a change. Again, kind of in the layer we're in this kind of data and infrastructure world, I think that's good. That means, hey, we're going to be accessing more of these things. There's more kind of data to harness. But you could see that as a change. And then look, I think we're in a dynamic environment. Whenever there's a dynamic environment and many things are changing, then that's something every company has to master in how they operate.

So if you were to put on like 5 years from now, because it feels like there's just definitional changes going on all around us right now. And technology has evolved over the years. And certainly, some thrive, some demise. What does -- and this is -- I'll get into some Confluent questions here in a second. But like what does software look 5 years from now? Is it just a series of APIs and connectivity, which certainly would play into Confluent's strength. But what does software look like in the future from your perspective?

Yes. I mean I don't think any change is ever that black and white. The old thing always remains. There's -- people still want to see a UI. It's not just going to be that something comes in and by codes, some back into the whole company. But I think what we're going to see is more interest in automating the things that happen between the applications, which may be done by kind of humans and clicking through it, and that's going to be a big direction of change. I think overall, you're going to have a lot more software in companies, the ability to create software faster effectively means that. I think the ability for companies to generate custom software that helps them do what they do is going to be higher. So I think you're going to see more of that. And I think that you're going to see an environment in which companies that harness that well are more successful and companies that don't figure that out are going to be less successful. Like whenever there's something like this that's changing in the economy, that becomes a vector of competition that becomes very important to faster. And so I think it will become a focus for a lot of companies to make sure they're on the right side of that.

So when you think about moats, moats change as technology evolves. One could argue that open source is becoming much more prevalent and maybe proprietary code is less relevant in the future. How do you think about building moats that are sustainable for some of these change factors when we think about -- is it access to data? Is it partnerships? Is it management quality? When you're advising companies and you're about Confluent itself, how do you think about these moats changing in the future?

Yes. I think all the classic things matter, right? Getting scale your relationship with customers, anything that has any kind of network effect across like for us, we kind of sit at this layer that connects a lot of the different data systems. So something like that where you have to get everybody to agree that kind of stickiness is very important. So I think all of those classic things matter. I think what you're likely to see is just a step change on the productivity in software. And you're going to see a widening of the scope of the types of problems that can be solved by software systems, right? And so more software doing more stuff. That's the change. We've seen little versions of that in the past. We went from a world where people were coding software applications in assembly language, which if you've ever tried to do it, is incredibly slow and unproductive. We went to something much more productive with higher-level languages and libraries. Now that happened drawn out over a period of time. But nonetheless, what was the result of that? Well there was more software, it was ultimately net more valuable even though it was easier to create. The value of software engineers actually went up in that time period because they were more economically productive, even though per unit of software, it required less software engineers to make it because it was easier to program. I think you'll see kind of some exaggerated version of that where you're -- I think you're going to see a step change in the ability to create software. And I think that, that's going to create a lot of demands on these different data systems and layers and infrastructure and cloud. And I think you're going to see a fair amount of economic value come out of that. And you could see it as being analogous to some of these other big productivity jumps in technology, but I think this may be sped up quite a bit relative to some of those more slow progress in programming languages, operating systems, networking, et cetera, which may be played out over some decades.

You've talked to the importance of real-time data. I grew up in a world. I was a COBOL programmer back today and...

It's the battle days.

Yes, the battle days.

You know all about...

I know all about that. I know about batch processing, and it's a painful world. But like when you're talking to customers today, how -- like there's this debate, how important is real time? And like where are we going in the relevancy of real time and just sort of thoughts around that.

Yes. Yes. I'll describe it just from first principles, like that as software is doing more to run companies, being in sync with what's happening in the company becomes more important. And that's extra true with AI, which can kind of suddenly take on these bigger lumps. And as an example of that, a lot of the data processing that was more sophisticated that we had was often kind of in service of business intelligence. It's kind of in the back end of a business, you run a bunch of data pipelines, so you can see some report. And the way that you would make decisions or take action is some smart executive looks at the report and call it so and so and says, "Oh, look at this, right? And I think what we're seeing over the last 5 years, but even more over the next 5 years is kind of a move to having that loop be closed in software systems, where you're kind of looking at the operation of the business and you're taking action on the operation of the business in software, right? And to do that, you go from something that's kind of a snapshot at a point in time to something that has to be in sync with what's going on. And you would see this very clearly in some of the use cases, right? So a customer of ours has AI-based support interaction. This is kind of very classic like the first AI use case was like help automate some customer support. The product just fundamentally doesn't work if the data that AI has is out of gate, right? So if it doesn't know what you've done, what has happened in the product, what you bought, all the aspects about you and what you're doing, it doesn't really matter how smart the model is. The model could get 10x smarter, it still can't do anything useful for you. And so being able to harness that and access it becomes very important. And I think that's fundamentally a driver of this kind of real-time use of data, and I think a very strong tailwind for us when we think about what it is companies need to do to harness AI.

We're seeing a couple of companies accelerate growth. And what's interesting is that a lot of it's coming from nondigital native companies, the broader economy. What is your sort of lens on who's adopting AI right now? Clearly, the digital natives are and you have a number of digital native customers. But talk about sort of the broad economy. Where are we in that level of adoption?

Yes. Yes. I mean, well, first of all, there's a funnel that you would see from the training of big models to the kind of adoption and usage of those models, right? And there's a time lag that would occur there, where we're kind of training models today -- and those are going to be better and more useful and they're going to generate usage and economic value in the future. And so there's some ramp-up there. So the first run-up you see is anything that is kind of in that supply chain for training. If you look at the kind of effective usage of AI, I think what has moved fastest, as you said, was effectively digital native companies packaging up some use case around AI and taking it out to customers as a way for them to adopt that use case faster. So that's the first category. So they're a customer of ours like Cursor is around coding, right? This is a coding tool that has an agent that tries to make changes for you and an IDE that helps you get coding suggestions. They're not actually building the model. The model would come from Anthropic or OpenAI or whomever, but they're kind of wrapping that up in a way that's easy for software engineers to consume. And so I think you've seen that probably move the fastest. I don't know that, that's the biggest portion of value for companies. I mean that use case is extremely valuable, but I think a lot of what's valuable to companies is solving their core business problems with AI. That takes a little bit longer because they have to figure out how to do what Cursor did, but they're not starting with a blank sheet of paper. They have to do it with all the gnarly data and systems and processes and regulations that they exist with. And I think they're still moving very quickly on those problems. But Quick here has a different benchmark in that type of organization. And I think we are seeing progress in those use cases. I think there is a learning loop for organizations of what types of problems can be solved now, what's the type of team that they need to be effective? How can they structure the stuff so it is successful. But I think we're seeing those start to move as well. I think those are arguably the biggest unlock because it's now taking this new capability and it's applying it to the unique thing you do as a company. And those range across industries in insurance, there's amazing use cases in claims processing in health care, there's a million things from billing to patient interactions to whatever. All of these are domains where there's incredibly messy data and putting that together and being able to bring some actual intelligence to bear is a huge unlock, but it takes time to move in any of those environments. And so that's a little bit the lagging indicator where I don't think we've started to see the bulk of that impact in organizations. There, you do see more a mix of success and failure, right? These organizations often do take a swing at something and don't quite get it right, come back at it. That's true, of course, in the start-up realm as well. You just never hear about the ones that don't make it. For us, we start to -- we're kind of right at the heart of the data flow for a lot of these use cases. We start to benefit as they hit scale and kind of have something production that's real, which is probably the right place toy. It means that as you're getting that revenue, it's sticky and durable, something that's going to continue. So certainly, if we look at the kind of realized revenue, we would see more out of these kind of AI start-ups than we would out of big enterprise. But I think if you fast forward a couple of years, that's probably reversed.

And you've got the largest frontier model out there as a customer of Confluent, and they've moved from a cloud version to an on-prem version. How do you -- like what is that right? Like if you look forward and some of these big companies are -- could be unique in how they consume software and technology. But do you see this moving trend towards bringing some of these elements on-prem or DIY? How do you think about that balance in the future?

Yes, not broadly. I mean most of tech effectively kind of moves as a unit and is more in the cloud than it was in the past and probably using more SaaS services than in the past. I do think these LLM companies are fundamentally different entities. They just do not look like an enterprise SaaS software start-up. They don't look like a consumer -- a traditional consumer Internet company in many ways. The internal structure, the capital intensity, their approach to solving problems, just like on almost every dimension, they're very different. I'm on the Board of Anthropic, and I've seen some of this kind of firsthand. I think similar things are true of some of these other efforts. And so yes, you just don't see traditional tech start-ups with an interest in building chips, supply of power, data centers. I mean, there's much more kind of first principles construction element to it. And so yes, I do think they're in that sense, probably more akin to these kind of large hyperscalers than they are to the average tech company that runs on top of the hyperscaler.

In terms of that, the ability for the average organization to use native open source, unsupported open source. Do you -- I mean, you're probably talking to customers on a daily basis of the value of Confluent versus pure Kafka. How does that evolve in the future when we think broadly of open source? Because I would imagine the proliferation of open source is only going to accelerate, just broadly speaking.

Yes. Yes. Our value proposition has been around a few things, right? First, it's creating a version of the open source that's kind of fully managed that can make it very cost effective relative to doing it yourself, even just in your spend on the kind of infrastructure to run it. And that's possible because we run things multi-tenant where we're kind of pooling the usage of many customers. It's possible because we've done a good job of just kind of closely engineering to what's needed in the cloud. And then it's possible because we also have huge advantages, certainly in excess of 1,000:1 in the operational capabilities for running these kind of big distributed data systems. So it's kind of getting that thing but better and more elastic and more as a true cloud service. That's the first value proposition. The second is really bringing to bear all the parts of the problem in real-time data, like really solving that, not just being an ingredient but being kind of a complete platform. I think Databricks did a good job of this in the analytics realm around Spark. We've done it around Kafka and streaming. I think that's increasingly what customers want is they want kind of a broad platform where all the parts work together. And in our business, the part of the business that's around the connectivity, the governance of data and the real-time processing with Flink has been very fast growing and a big part of the story to customers when they think about, okay, not just how can I get the real-time data, but how can I build around that? That's an essential part to them. And I think that, that kind of complete package makes it quite differentiated. And then there's a third thing in addition to a real cloud-native offering being a complete platform. The third thing is really making something that works across all the environments that they operate in. It's probably unique in this streaming problem. The role of the technology is to kind of act as like a central nervous system that plugs together all the applications and parts of the company. And so it ends up having to span all the environments that a company runs in and just getting that to work across edge, on-premise, different cloud providers, having that all connected, it's actually quite complex. And so to just make that work for customers is actually a big unlock as well.

The managed element of that is important, especially as this world becomes even more connected and open for that matter. How do you see, just broadly speaking, open source software proliferating across the entire infrastructure stack? Obviously, there's areas that are fully penetrated, whether it's the OS layer or.

Yes. Yes. I mean it has continued to succeed. You continue to see all the models working, right? There's proliferation of open source. There's a ton of proprietary software that's out there that's successful. What I would say in the area that we're in, what it acts as is a kind of standardization mechanism. And so the technology industry, it really loves a stable standard everyone can build against. And open source is one way to get that. It's not the only way, right? But like even for -- if you're a company like Cisco, you benefit from the fact that everybody has chosen GCP IP, and that's the way data is going to flow on networks. For us, that kind of fundamental protocol of how data is going to flow across an organization is Kafka. That's kind of one. So that's the default. We have the leading offering around that. That's a very powerful thing for us. Ultimately, once everybody has agreed like that, then the whole ecosystem comes in and starts to build around that as well. So you get all the integrations into all the systems that would be adjacent. That ends up -- that kind of ecosystem generates a lot of the value that customers realize in a way that we could never do entirely on our own. And I think that, that's a very valuable and sticky story. Like when we were starting the company, there was active competition from these non-Kafka layers that were trying to do a similar thing. So Amazon had a system called Kinesis, there was alternative open source things. But ultimately, that kind of standardization or network effect of something that just kind of plugs into everything, works with everything, fits all the architectures, that kind of meant that the thing that got ahead stayed ahead and the thing that got ahead was Kafka. And obviously, we worked hard to make that happen. But I think that's a powerful force going forward and that it just becomes a very difficult thing to displace in companies once it's kind of installed as that layer across.

One of the other things that I think a lot of people think about with AI is consolidation, big, bigger data wins. I guess a 2-part question. How do you think about the competition with that hyperscaler level? And then maybe secondarily, how do you think about this -- like is there this broader consolidation play? Like do you think we continue to see they get bigger, continue to see M&A in the space? Just kind of thoughts around that.

Yes. Yes. The competition with the cloud providers has been relatively stable for some period of time. So the early dimension was they each tried to make -- the way the cloud providers work is they have 2 categories of systems. They have systems which are proprietary to them, which they invest in heavily on the R&D side. So that if you're in Amazon, that would be things like Redshift or Aurora. In our space, it was a system called Kinesis, and they like those because if you adopt them, it's only available in Amazon, it doesn't work anywhere else. Each of the clouds have those. But ultimately, the customers vote with their feet as it were. And for some of these layers, the proprietary cloud system is one, for some of these layers, the open system is one. And Kafka clearly won in the streaming space. I think Postgre is winning in the kind of database serving space. And so those early proprietary systems kind of all died out. And most of the cloud providers just fell back on like, well, okay, we won't do any R&D in the streaming area. We'll just put the open source on some surface and see that will get something. That's not the hardest competitor to compete with. It's ultimately not different from the kind of DIY open source thing. So I think we've been quite effective at building differentiation against that. I think each year, that differentiation grows rather than shrinks. And as our platform becomes more complete and solves kind of a broader set of problems across that even adds to that. So yes, I think that's been very consistent and not a very kind of stable competitive dynamic with them. And I would say that's just about the streaming bit. By and large, because we're a connectivity layer between all these different data systems and layers, if there's 300 products in the cloud, we cooperate with 295 of them and drive consumption to them with all the data that flows. We compete with a handful of streaming things. But by and large, the clouds end up being pretty good partners that actually help us along.

I'll ask you one more and then see if there's a question out here. In terms of like when we think about the model builders in the future, obviously, a lot of them are customers of Confluent today. Do you see -- how do you see that dynamic evolving as they continue to go after more TAM? I mean, could that be the new competitive frontier in the future for the broader technology landscape, maybe Confluent specific?

Yes. I mean this -- I would say a few things. So it's clearly possible for the LLM providers to do more around their offering. And this is what you're asking about. When you say model build, so you would see that in some of the expansion of functionality that OpenAI has had. There are certainly kind of vertical AI start-ups that we're adding stuff that has just been developed out there. Anthropic has added coding tools, which have done extremely well. There's a tendency to extrapolate from that to like, well, everything will just be developed and I've been -- we kind of started at the beginning of the conversation with past -- and I think in each of these cycles, there's some companies that are doing extraordinarily well and people imagine that they will just do everything, right? And I think there are certain constraints that prevent that. So when I was at LinkedIn before founding Confluent, I was there early and we were sort of social network and there was Facebook social network. And whenever we would interview somebody, they would be like, well, this is fine, this social networking thing. It seems like it's turning into anything. But if Google does it, you guys will just die. like they'll just wipe you out because Google was that company at that time where just they could do no wrong, everything they did was genius. They were very technically innovative, et cetera. So the assumption was if they just -- if they ever got up out of bed in the morning and had any inclination to do that thing, it would just completely kill Facebook, LinkedIn and all these other companies. The reality was they tried really hard and it totally didn't work. And what's the lesson in that? The lesson is doing things is hard. Big companies have a lot of force that they can apply, but they can only do a pretty small number of things in parallel, maybe 5 things, maybe a little bit more if you're very innovative, but you can't do 100 things. And so I think what we've seen is there's going to be a ton of opportunities around this that will get explored. I think we'll see some things get drawn into these bigger platforms, but I think that's natural. I mean any of these expanding areas, you're going to add capabilities, but I think there's going to be a broad set of applications around that are successful. I think the model companies are going to be very successful. I think many of the infrastructure layers, I think, will benefit as we will through this overall rollout of AI.

So TAM expansion.

Yes. I mean, look, I mean, I think very fundamentally, there is something very valuable happening with AI. And so I think we're going to see that get captured through a lot of different...

Are there -- I'll be able to pause here. Is there any questions for Jay?

I guess we answered them all.

When it comes to the large language models, the simpler question is, how many models does the world need? I guess the essence of that question is if you look at the history of tech over the years, there's always 1 or 2 winners in every new emerging tech. How do you see that sort of consolidation when it comes to the large language models? Do you think AI is so different that we can support multiple models or the specialized models? Or do you think it's going to be a consolidation like everything else we have seen?

Yes. I mean, first of all, I would say many of the people in the room are probably smarter students of the business landscape than me and probably have as much to say about it. what I've seen at least is these enterprise markets have a certain rationality. So the buyers are thinking broadly. And so they actually like a certain amount of competition. So coming into the -- coming in as a new creator of LLMs today, if you were to go start a company, I think you're probably jumped unless you have some very unique insight that obsoletes everything that's come before, which is the level of capital investment and process knowledge, like the number of small improvements that you have to accumulate to be competitive is very high. So I think it's very hard for a new player to come in. At the same time, enterprises actually like to have a couple of different people who will sell to them and they like to bake them off and get them and that's actually how they control their cost structure. They don't want it to be the case that somebody captures too much value in that chain. I think that's true up and down the stack that each player kind of looks at the margins of the layers underneath and if it gets a little too big, then they want to have an alternative source. And enterprises just look forward with a little bit more. And I think you would have seen that play out in the cloud. when we were starting Confluent very early on, we released the cloud services. And at that time, many people thought it was just all going to be Amazon. It was just like it's just hopeless. Nobody else is going to do anything. It's just all AWS. They had multiple years of lead. They were just ahead in functionality. They had the talent. They had the momentum. They had the best conception of it. They were more focused on it. So everybody at that time was like, okay, the hope for Confluent to add a cloud service in this world that's just going to be Amazon doing everything in the cloud is not so good. The reality was, first of all, Amazon couldn't do everything, right? Even they have bandwidth constraints. Secondly, customers actually will competitors into existence. The drug Microsoft along until it was a very credible cloud provider and they could rely on that. Similar thing with Google. Ultimately, the success of Amazon both motivated those competitors, but also motivated buyers to kind of get them there. So I think you get a certain kind of industrial logic that plays out where you get competition, but not unbounded competition. It's rational enough that nobody -- everybody makes some money, but nobody makes too much money. I think you kind of see that play out a little bit. And the forces are a little bit intrinsic that produce it. So I wouldn't be shocked to see that happen as this advances, but we will see. And to some extent, your guess is as good as mine. I don't think that there's no force in AI models that prevents that. Ultimately, there's barriers in capital, there's barriers in research, et cetera. But there's no effect by which a small lead magnifies 1,000x into something that's inescapable. And so other than just the accumulation of intelligence itself. So we'll see how it plays out.

Thanks. What about the DeepSeek? What role -- I mean, how do you see that? Do you see them? Do you customers?

Yes. I thought it was interesting. I mean the most interesting part is probably the U.S.-China aspect of it. I think that they're impressive and that it was a relatively small team that was able to get something that was kind of at least close to the frontier with obviously sizable investment, et cetera. many of the things people thought about that story were not true, right? So there was some data point where it was like, it cost $10 million to train the model. Of course, yes, each training run might have cost $10 million, but it's on a cluster that has a cost, right? So like yes, so that onetime slice was $10 million, but it was a lot of $10 million lined up. So many of the things people thought, I think we were just not the way it works. But yes, I thought it was an impressive effort to get something that was kind of up there in competition. It doesn't -- in this area, being good at inference and the cost of inference is so high, it's not like pure software in that if you take an unoptimized model and run it on chips, it may be more expensive. for you than just using one of these APIs where they've just really optimized for that model all the way down to very low level code. And so I think if you see how that's affected the market, my perception is still the bulk of spend is going to these top models, right, Anthropic, Google, OpenAI, et cetera. There's clearly a tier of things that are kind of going to open source. I do think that this is an area where, again, the rationality of the buyers kind of wheels it into existence. They want it to be competitive. And open source is always good at finding niches. That said, the nature of these models, they're called open source was not really the same as open source in that like if you have the weights for version X, you cannot make version X plus 1, right? You don't have the source code. So you can't actually keep producing it. It's more like having the compiled binary, right? So if I gave you the compiled binary for Linux, you can't improve Linux, but you can run it for free. That's more or less what it is. And yes, you can kind of fine-tune the weights, et cetera, but it's a little bit of just pushing it back and forth. You can't actually make the next iteration. So some of the analogies with open source kind of break down. But yes, fundamentally, I would say that there's always in software, there's a commoditization game because it's easy to give away stuff because the fixed costs of just the software part are low. That's true in our business. That's true in many of these other businesses where you're coming up with a way of giving away some value that costs you nothing and you're creating commercial offerings that are going along with that. And that's -- I think everybody does some aspect of that.

Maybe just to wrap up from a -- I always like to ask the moonshot question. So we think about maybe a 2-parter for both Confluent when you think about some of the moonshot opportunities for Confluent, but then the broader AI, do you have any sort of like bold predictions about AI?

Nothing super interesting. AI is so discussed at this point that are talked about. The most interesting questions are around kind of really the rate of progress in different use cases. When can we do x? The reality is nobody knows. Even the people building the models, you don't know exactly what's going to unlock when. I think the biggest thing to that's outside of the normal LLMs is obviously some of this physical world AI and robotics. If that moves, I think that's a whole other dimension that is equally mind-blowing and weird in terms of the implications. Early indications are there is some progress in that area. There's obviously a lot more investment going into it than there was 5 years ago. So I think that's an interesting one to watch. But predicting the rate of progress on these what are effectively scientific problems is just not easy. Most of people doing it don't really have that much more information than anybody else. So I think we'll see what happens. But if that moves. That's a big one Yes. As for Confluent, our role, we feel, hey, we're very well positioned. One of the fundamental ingredients for these AI use cases is data. The need for that data to be kind of real time and in sync with the business is just very fundamental. If you want to act as part of the business, you got to have data that's in sync with what's happening in the world. Otherwise, you're -- it's like trying to walk around in your daily life using a snapshot of what was there yesterday, right? You're going to bundle these things. So we think we have a big opportunity there. We had some new functionality at our conference a few weeks ago around real-time context data, the processing and generation and serving of this. I think that's a big area. I think it's going to turn into a big use case for us over time. So I don't -- I wouldn't call it a moonshot because maybe we already put in the we're halfway to the moon or whatever. But that's certainly, I think, a big opportunity in this space for us.

Right. Well, that was a lot to think about. Great insight as always, Jay. Really appreciate your time from all of us at RBC and everybody attend, thank you. That's my pleasure.

Thanks, everyone.