Datadog, Inc. ($DDOG)

I can't even see any of you guys -- view in the middle -- Olivier is obviously a much more interesting and important person than me. But for those of you who don't know me, Peter Weed from Bernstein, I cover software. There's actually 2 of us from Bernstein who cover software that you might know, Mark Moerdler and myself. I do kind of IT infrastructure, dev type software, which is kind of my long-term background have been kind of around it since the late '90s, and you may know Mark, my colleague, he actually was a database entrepreneur back from the same time. So he covers databases and mostly like functional line of business applications. And we are really fortunate today to have Oli from Datadog join us. Oli is 1 of the founders and the CEO of Datadog, a company that has been really instrumental in category that has really kind of emerged over the last decade that we now call observability, which is all about helping like when you ship a cloud application, you expect it to run. And this is a company that helps ensure that it's both up and it's running at performance before you even put it into production, you can predict how it might behave. Now I recognize the audience there's probably going to be a variety of background in data and kind of where they are. I think it may be relatively obvious from some of the stock price action recently that things are going pretty well.

Oli, I thought it might be just interesting to kind of reflect on the last kind of 3 to 4 quarters and the acceleration that you've seen and kind of where has that really been coming from?

Yes. So I mean, look, First -- thanks for having me here. We have seen acceleration for the past few quarters. I mean, I won't bore you with the exact numbers. We deal in the filings. But we've seen acceleration. What's the most exciting to us is that it is not a specific customer, specific set of customers, specific side of the business. We've really seen acceleration across the board. And the way we look at the customer base today is we look at both the -- what we call the AI natives, which are largely companies that were started in the last few years built on AI or companies that were selling a little bit before that, that are the -- essentially the large builders and the most important building blocks of they are today. So we see acceleration in that category, which sort of makes sense. I mean, everybody knows the AI is happening right now. So -- but what's even more interesting is we see an acceleration of the business outside of that. So in the non AI companies. And for us, that's a mix of the older cloud natives, which existed before AI and also all of the larger enterprises, whether they are large enterprises or mid-market, and we've seen acceleration across all of that. So on the AI side, obviously the AI companies are consuming a lot of infrastructure, building a lot of applications, running a lot of applications, and that's what we see on day-to-day basis. On the rest of the market, we see also more applications, but we also see some growing adoption of AI itself. So we have -- we've disclosed in the last quarter that we've seen very large increases in volumes of our AI-specific products, whether that's the traffic we see to our MCP services or whether that's the traces we get into LLM observability product that really gets traces produced by AI. So we see inflection across all those points.

Oli, I think you've done a really nice job kind of helping people kind of understand like how you get paid and how you kind of ride the cloud wave. Maybe maybe describe like how you think about like where your value comes and like where your growth equation kind of comes from over long periods of time?

So the value comes from helping customers be on top of the complexity they create their applications. So meaning they ship applications, they have to understand what infrastructure is part of that application, how it's running, whether the application itself is running, if it's fast enough, what the users of that application are doing whether the application is creating the right outcomes for the business, whether the application is safe and secure. So all of that, that's what we do for our customers. One way to think about the growing importance of what we do is that, over time, developers have become way more productive. So when you go back 50 years more than that, developers were hand coding every bit on punch cards and writing every single piece of code that was running-- then after that, there's been higher level languages, better interfaces, compilers, then after that, you've seen open source software. You've seen cloud, you've seen SaaS and internal -- at every single step of the way, you see an increase of productivity. What it means when you get higher productivity is that you produce less or you produce more in less time. which means that you don't really get to [indiscernible] as much to it, you don't understand what's actually going on. And you see an even bigger acceleration of that today with the coding agents, where engineers can in a few minutes, with open application. They just have no idea what's going on. They probably don't even read the code, but they combine all sorts of different components into an application. So that increase of productivity creates a dramatic increase in complexity, and the problem we saw for our customers is we actually understand that complexity. We manage it for them. And in the future, we want to be even more an automate basically the way the applications run after the [indiscernible].

It's interesting. I think like 1 of the bare cases that has been pervasive around software is like now that we actually have these agents this complexity can now be handled and can't we just go ask whether or not it's Anthropic, Claude or any of these LLMs to just do it for us, how is my system doing? Tell me when it's down. Like, why does Datadog still have a role in a world where we have these AI agents? And maybe it's not today, but maybe 3 years from now, they get smart enough to do this.

So I mean, the 2 things I'll say. First of all, the companies that are building these kinds of models are also using our product to manage their infrastructure and their applications and everything else. So it is a hard job. It is consistently a harder job than writing the application itself. And that is why there's been this race has complexity over time. And everybody is going after it, including the leading companies AI [indiscernible] AI builders all the hyperscalers. So that's 1 thing. The other thing I will say is that the kind of problems we saw are somewhat different in nature from what these fundamental models or financial models are built for. and operate at very different levels of data volume and real-time requirements. So we have 5 or 6 orders of managing more data than which is typically put in an we need to operate in the real time. So the analogy I would give to understand where we stand with the prime we saw versus what these models do is compared to self-driving cars. And so mean today, it's working pretty well. You have self-driving cars everywhere. You could, if you wanted take a photo from your dashboard on your car and upload it to Claude and ask what you should do. And it will give you an answer, and it will be a pretty good answer. And we'll tell you actually you should make -- if you turn on the right, and that's why, and this is what's going to happen -- the problem though is that Claude will do that, but it will not drive your car in real time. If it did, it would do it poly, were worse than the algorithm that's or the mall that's running in your car. And also, it would cost you $10,000 a day to drive your car. So that's a way to understand what these models do relate to what we do in our case, for observability, automating the life cycle of an application, automating the availability, the validation, the security of an application.

I think the other bear narrative like when you get through that, oh my gosh, it's actually a difference between like deterministic programmatic software and kind of probabilistic software, it's like, okay, well, I still have these code suggestion tools. And we've got some open sores like why can't I just like to ask it to build me this bet a programmatic software late, you're obviously not seeing that happen, like what's the disconnect between that narrative in the reality of like why the most innovative companies are turning to use Datadog as opposed to do it themselves.

Yes. I mean it doesn't make economic sense for you to do it yourself typically. And the reason for that is this is -- what we provide gives enormous leverage. So typically, for any dollars customers spend on us, they're going to spend $10 or $20 on their cloud provider. They're going to spend $20 to $100 on their engineering team. with the rise of cutting agents and [indiscernible] in general, maybe some of that $20 to $100 depending on their engineering team is actually going to go to these AI models instead -- so we're talking about a huge amount of leverage. So when the -- and by the way, we help our customers make the most to optimize their spend on their cloud provider, on their engineering team, on their AI levels on all of that. So we're basically the small part of spend that helps our customers make the most out of the rest and accelerate and make more money, so at the end of the day, it makes no sense when you have them to do that. I'll give you a few more examples that we've had, so more recently. So we mentioned over our last earnings call that we also had a number of hyperscalers static adopting our products in their case for helping with their the development and training of their AI models. And these are companies that culturally basically don't use any commercial software. They write everything themselves. They use open source, they have to completely homegrown by nature. Even those companies had to come to the realization of that, actually, it didn't make sense for them to do that. They would get better outcomes. They would get faster delivery of what matter the most to them. They would get better economics of it, and that would be better for the business in the long run. We didn't build Datadog to serve a handful of hyperscalers. That's not our business model. A business model is going is to go after every single company out there, have products with very broad appeal. We think though that this is a great example of these hyperscalers just going for the first time, maybe through what every single other company has to think through, which is, okay, so what do I need to own? What should how do I get the best return on investment? What am I trying to achieve? I need to get growth, I need to get better economics at scale? What's the best way of doing that. I have finite resources, which should I focus on. And I think that's what we're seeing at play here.

So if -- can't be replaced by AI, just asking questions can't be coded, this place still has a tremendous number of start-ups, like I spend a lot of time on Hacker News, for better or worse. I would say, every month, there's somebody who's like celebrating the fact that they're trying to build some new observability platform. And you and I were talking about this a little bit, but beforehand, my history, I'm an old product manager -- years ago at Microsoft in the middleware category. And the 1 thing that has stuck with me this whole time and when I worked at McKinsey and other places, is that I always watch the buyer. And for me, what I noticed is that the greatest chance for a new entrant coming into a market was what the buyer changed, right? So like you could say like, why does Salesforce exist and replace prior CRMs. Well, because the buyer moved from the CIO to the head of the function. Why is Datadog, I think, an amazing company that replaced prior generations of what we might have called monitoring software. Well, because we got the site reliability engineer and a new set of expectations and requirements when we move cloud delivery and management. If we look at AI, it appears, and I think you've made a good case for when we shed in the past, like the roles might compress again. There may be several roles that need to get redefined if we're really going to have high velocity shipping security might need to be grouped in as well. And that looks like 1 of those scenarios kind of like the site reliability engineer that collapsed some old roles together and gave you an opening into the market, but I think you guys have been thinking about this almost from day 1 about like why that type of change actually is kind of like what you're almost hoping for because of the upside as opposed to being a threat.

So a couple of things on that. So the first thing I'll say is I mean there's a -- I think you understand when you say there's another new engineer building and observative product every month. I think it's every 2 weeks -- and it's been like that since the very beginning of the -- so I still remember the first fundraising meeting I took in the -- with VCs way back in 2010. And the first word that came out of the mouth of the investor where all monitoring because that's what it was called at the time, monitoring credit market. . And then it went down here from there. The -- and it's always been like that. And I can't blame people for doing that, and that's what I did. I mean engineer, I started an absorbing company, and that's naturally I think we did approach it a little bit differently though. So our starting point was not, hey, let's build a better product [indiscernible] we're using that, we had it. We're going to build a better one. Or we're using that is too expensive, let's build something else. The starting point for us was, hey, I used to come from development. My cofounder used to come from operations -- technical operations. The 2 teams fought all the time -- and we thought, okay, so , maybe we can bring them together and have 1 platform that bridges the gap between the rules and things like that. And that vision actually lended itself very, very well towards adding more roles, bringing more people into the mix, expanding the footprint over the lifetime of the company. And that's what we've done. We were actually quite lucky because -- what we didn't understand when we started that was that the collapsing of the role at the time between DevOps, we're at the center of cloud adoption. The whole DevOps movement was born at that time. And that's when you saw those rules become closer. And we saw that pull in the market. but you see even more of that happening today. I mean you see securities being brought in, you see with the advance of cutting agents, now instead of large deals with specialists, you're going to have smaller teams with people wearing many hats across product, design, engineering, security and many other functions. So I think we're going to see even more on that. The last thing I would mention because you talk about resisting disruption. And we built also the company very deliberately in a way that keeps us in touch and honest when it comes to the low and bleeding edge of the market. So we have more than 30,000 customers. But the bottom half of those 30,000 customers represented about 1% of our revenue. So we do not serve that bottom half for money, it actually cost us money to serve those customers. We do it because those customers are -- they are small companies, they're individual students. We also have people on the free tier that we only have included in that. And what those people do is they keep us honest with respect to the simplicity of the product. They also help us see what's coming in the market because we see they adopt all of the new stuff way before the large companies do that. So we see that coming. We get many of what we need to build for. And whenever we break things because we try hard to accommodate our largest customers that have very sophisticated needs and that make our products more complex. We see -- actually, we hear from those smaller customers, and we also see in the numbers that really something is not working as it should be working there. The alternative is what most software companies do, which is they quickly understand that most of the money comes from the higher end of the market. As a result, they focus to care mostly about their large depressed customers. and then they end up being on the path towards future escalation. And before you know it, you end up with products that are what I call enterprise abomination, and we all live with them every day in the office. We know exactly where they are. And I think that that's -- you don't come back from that basically. That's where we become irrelevant over time.

I think we've focused a lot of this conversation is like some of these like bear cases around the people have. And -- maybe we should turn it a little bit to some of the opportunities, right? So you all don't rest. Obviously, I think you spend more in R&D than pretty much the entire rest of the sector put together. So you're shipping a lot of things. And if we look historically at the S curves that the business has gone through, maybe you started with metrics, actually really kind of collaboration and the metrics and then APM and then logs, and these have each kind of grown to being kind of $1 billion type opportunities. There are several things that you've been putting out that you, I think, have been very hopeful that can become some of those next desk curves. -- whether or not it's security, whether or not it's AI monitoring, the GPUs, modeling that you guys have been putting out. when you look at the acceleration that's going on right now, how much can you kind of thank those new things for this versus kind of your existing footprint?

What's interesting is most of the acceleration, like mathematically, it comes from the existing stuff. And when you think of -- like we have the I think we said like 14 of the top 20 AI companies using us, including the very largest of those. Like they use the same traditional products as the rest of our customer base. So by and large. So that gives you a sense of the shape of the market in general. When you look at the observed category, so there is gotten a number that we put every year, but we're #1 there was ability according to Gartner in -- or at least in the closest -- the category had as closer to observability, I forgot the exact name of the category. . And -- but we only have about 13.6% of that market right now in a market that is growing pretty far -- pretty fast. So it doesn't take a ton of imagination to see where there's a 5x from there just from the core capability, looking at infrastructure applications, end users around those applications, logs and all of that stuff that's that. The new part of the market is everything that relates specifically to AI. Part of that is what we call AI for Datadog, which is who we can do more for our customers and really expand beyond the observatory category. So basically automating the whole for them, not just being in the business of telling humans where things are going, but being in the business of fully running and automating the application, it's promotion to production, it's repairing it when there's an issue, making sure it's secure at any point in time, there's a lot of automation we can do with that. Another area is what we call [indiscernible] which is observing and understanding all of the AI components our customers are using. And mostly it's fall into 2 categories. One is the AI models our customers are using within their applications or within their agents, which are nondeterministic models and which can't just be tested in development and then validate a ship to production, you have to keep observing them all the time. Like they're like humans, you have to get to know them and you spend time all the time and you keep watching, and you see what's going on and you keep adjusting. So that's 1 big area. Another area is adapting to the whole new world of agentic coding when developers can produce a lot more applications, a lot faster. But would they still come do faster is getting those applications to production. Now they still need to make sure that those work that they do what they're supposed to be doing, that they provide the right outcomes for their business that they are secure, that they keep working once they've been put into production and when other people are also making changes to the world around it, so that's a huge -- right now, we see a huge increase in the volume of co-change that are being made and applications that are being shipped, and there's a big opportunity for us there, too.

I think there's probably like 2 kind of natural places to take this. Maybe we start with the Datadog for AI. So if we think about some of the things that are going on, they're obviously like there's some very basic uses of AI, but as we start to get to agents look, maybe beyond coding copilots, there isn't like a lot of depth of use of agents. But as we get more and more of that, there are some kind of unique observability problems that come up with essentially monitoring like the behavior of the agents over time. And there are some companies that have come out, things like LangChain and BrainTrust and these types of things that are designed to kind of tackle some of that job, like how do you see Datadog working alongside of or instead of some of these kind of born and AI components that are there to observe kind of the behavior of agents?

Yes. So I think there's 2 extremes to that market. There's 1 side which is everything that's in production. And when you talk about things that are in production, it doesn't make any sense to separate the AI from the non-AI components. For 1 thing, what the agents do is -- so they have a lot of reasoning, but also they use a lot of tools. And the tools are actually apps. So when you monitor the whole thing, you basically end up monitoring like largely apps with a little bit of AI in sprinkle. So when you're on production, it doesn't make any sense to have different tooling, and that's something that we will -- like we develop, we have an element [indiscernible] product for that reason and something that we will want to own in the end. On the other side of that spectrum, you have what happens in development. So when developers are tinkering, they're playing on their laptop, their testing various things. Historically, that side hasn't been very platform in that in a company, you don't have 1 of them and everybody plugs into it, but whether you've had many of them. And these were more a matter of choice and taste for the individual developers. So the equivalent there could be the IDEs, the developers we're using. I'm old enough to have leave some of the AMAX versus VIAs. So I remember people being very opinionated about the editors they use and things like that. And whereas absorbability, you have only 1 production application. So like everything has to plug into 1 environment to 1 thing in the end and that's platform in nature. So what's not completely clear to us today is how much of that is they've clearly developed a purely local non-platform lots of it of usage in any company, some homegrown, some and source, et cetera. And what part is in of basically back there, we have to go to own fully the production side of things.

So when you kind of look at that opportunity going forward, I think perhaps 1 of the challenges is it's very dynamic, like how we want to do all of this stuff is somewhat of a work in progress. How do you think about your own investment in kind of being part of the frontier there, so you're in the right place at the right time when things start to solidify and become a little bit more common across organizations.

Yes. So I mean we spent a ton of time like watching our customers and what is they do and how they work. And so we have the luxury of catering to every single layer of the stack in the customer. So most of our business is with large enterprises. But we also serve a lot of start-ups in AI, whether they are super large-scale model makers or the companies 1 level below that are largely AI native and building on AI. So when you look at the largest, it's interesting to see how they work but it might not be completely transferable to the rest of the market because those companies have unlimited access to inference for 1 thing. And also for a number of them like almost vertical wall of demand in front of them. So they are not in the situation that both of the companies are in. So they might make different choices and that might the way they operate might not carry across. When you look at the tier below that, though, like the companies that are, I mean, still very successful going fast in AI, but are still buying most of their influence from somebody else and have to make some hard prioritization goal in terms of what would they work on and not and have more relatable growth for the rest of the market. I think from there, there's a lot we see and we can learn in terms of where the world is going. We see the trends pick up like a few quarters before the rest of the world might and we get some interesting insights in terms of how our products are being used and we need to be able to support those companies as they do that.

And I feel dumb. I feel like I should have almost even asked the question in a slightly different way because I realize for the audience, they might not fully appreciate your product-led approach to the organization, which you're kind of describing some of the outputs of product-led is something that I have been kind of in love with ever since like I got enamored with Atlassian flywheel approach background in 2010, and I tried to apply it to several businesses like building it out of Dropbox and things like this. like maybe help people understand what is special about that product-led approach to the organization with the growth function and with product management being so tight with discovering these opportunities and delivering and why that allows you to stay so innovative?

Yes. So the whole business is built around bottom-up product adoption. So we -- our product is typically adopted by the record files and then the people who actually do the work. We build around a unified platform, which is very important, meaning every single new product will build as part of that platform. Whenever we do M&A and we acquire company we do platform. So everything is part of that tightly integrated in 1 single platform. We charge per usage. So the whole business model is usage. I mean there's a few tiny parts of the revenue that are based. That's because we compete in categories that are purely based. But the vast majority of our revenue is completely usage-based, which also is interesting because right now, it allows us to I mean basically, we don't have any transformation to undergo to get into an AI edge where the notion of seed gets disrupted and you have to charge for outcomes for value per something. We're already usage base, so we can very easily adapt our model. Everything around the adoption of our product is -- everything around the business is geared towards the adoption of the product. So when we land the customer, we're already deployed typically and we just see the usage start growing from that time. What this gives us as a business is a great way to understand which products to build and not build for our customers and what is valuable or not. Again, we use age based. We [indiscernible] about having that usage or having new products being attached to new SKU, meaning that we get very clear signal about what's valuable and what's not in our products. which allows us to keep developing a lot of the right things. So that's -- and if you follow the business, were knowing the industry for very quickly expanding the product footprint very quickly, improving the existing products and be very good at growing new lines of business. And we have a lot of tasks in our earnings around the adoption of those businesses. The other things we get from having this product-wide approach to the adoption and bottom-up adoption, is that we have a very efficient go-to-market. So we spend significantly less as a percentage of the top line on go-to-market, which then the typical enterprise business is which allows us to invest about 30% of the top line on the R&D, which then feeds the flywheel of building more product and building more differentiation and being relevant for the long term.

Maybe the asterisks on that now has been that you are actually starting to build some enterprise sales capability. like what drove that evolution where you keep kind of the product-led core, but you realized that some enterprise overlay was important. And how has that kind of worked in...

Enterprise has been around for a while, like we built that team many years ago. So the -- all that's baked in. What's important to understand is that it's enterprise but bottom up. So it's never the case or maybe I can happen once. But it's never the case that you call up the CIO and you're super impressive in your golf game and you end up with a deal that gets done, that's not will happen. . In our case, we start with the teams, the teams adopters. And then we make sure that we connect between the teams that are adopting us and the leadership in those enterprises, so that they understand what is their buying and what the ROI is and things like that. And so that we can enter growing and longer-term agreements with those companies. But even with the large enterprises, we mostly land very small. -- the ASPs when we land are in the, call it, mid-5-figure to low 6 figures every year. which is very small for this kind of enterprise deals. And then we grow those customers for the very long term. And we announced on the earnings call, we typically have very interesting land deals or consolidation deals we make in the 6, 7, 8 figures annualized. But most of our deals start more and they grow and they -- we tend to have very, very efficient sales, thanks to that. The 1 thing that we're doing today that is a little bit different is the new sales motion we're adding is that we're adding some overlays for selling security very specifically to the [indiscernible] because that's not the most we had before. And we find that it's a little bit different -- different enough to warrant a slightly different team to handle that. But that's specific to security to a CISO in the largest enterprises.

And by the way, maybe talk about that a little bit like why is security such a natural part of what you're doing? And obviously, you have security vendors that are also making some acquisitions into what looks like competitors in your space. Like how do you see that playing out between like your role in cybersecurity and what some cybersecurity vendors are doing?

Yes. I mean, look, the end of the day, the thesis we had when we started the company that you should bring Devon ops together also extends to security. It doesn't make sense for security to be separate. -- the teams would work well together, they should speak to same language, achieve the same role. And we see the same -- we see this accelerated by the adoption of AI, where you'll have the same person basically having to carry to wear many hats in care about security as well. So we think that, that bottom of adoption from the developers and the ops people is going to reach into security. And is, in the end, going to be -- it's going to become much more of a bottom-up motion than a upon motion in terms of how that adoption. From a technical perspective, most of the signals you need to secure an application, you get observability. You have everything that happens in the production environment, you have everything that happens on the way to production in the core, in the testing environment. You have the behavior of the end users, you have the behavior you have the developers. You know exactly who's changing what and when, so all of that is the information that you need to fill into the security models, and we're uniquely positioned to get that as well.

So we went down the path right now of like some of the expansion opportunities organization and the commercial motion and why that's really helped you be successful. But I think you pointed out there's like the other side, which is AI for Datadog. And while it's still relatively early, there's kind of several investments that you're making there, everything from the its AI agent to some of the new modeling stuff with Toto that you guys are doing. Maybe help the audience kind of understand some of those things that actually can add incremental value, reduce the human time, add more predictability that you guys can do?

Yes. I mean, the high-level goal there is to automate right, and to do more than observed and to really automate -- so 1 thing a lot of our customers go through is so they use us to observe and monitor the applications and something [indiscernible] 30 a.m. and we work them up and then they start Zoom call with 20 people on it, that last 3 hours from 3 a.m. to 6:00 a.m. The site is down, business is upset. I mean you can imagine -- Yes, very costly, very disruptive, not anyone's favorite part of the job. So there's a huge opportunity to automate there. And it used to be fairly hard do that. But with the new technologies in AI are really opening doors that were just not quite available before. So what we do today is we have products under operations side and on the security side, they do that automation. We have a bit, that's a name. So sorry, agent BI security agent. And what they do is they run investigations for you. So for example, in the case of the outage I mentioned earlier, instead of having like a 3-hour Zoom call with 20 people on it, 4 minutes in at an analysis that says, "Oh, by the way, I find out what was broken. This is what this is what it is. These are the 3 people who know how to fix that. This is the fix I propose, not sure we do it. And I mean that's already life changing for those customers. And typically, this is the way they like they see they turn this on, maybe they forgot the tuning on and then the gain this incident. And then they get on to that long zoom call, and then after the Zoom call, they look at their notification and they were asked, "Oh my god, 4 minutes into it, you had it already. So -- and that's the light of moment from them and they start that. We have a similar agent on the security side that does security investigations. And there, the focus is mostly to cut the noise. So the teams that do security response have tons and tons and tons of investigations to run. 95% of them are more even are just benign. And we -- a big part of the job there is to make sure to get to the super quickly, so they can focus on other things. And so we do that as well. So that's what we do today. If you squint, you could say, hey, some of that you could also do from you use cloud or and you give them MCP tools to acquire your data and maybe they can come to the to the similar reasoning and come to a similar conclusion. The next level of what we're building though is about having that intelligence directly inside the data plane at with all of our real-time data and to drive the automation itself, so models that are specialized that can run on very large volumes of data at a very attractive cost and with very low latency, so we can fully automate the operations. There are some first outputs of that. You can see, we published 2 versions of a time series model like a financial time series model, which is called TOTO and you can look for it online. This model is very interesting because it is a model that is -- that was trained on observatory data almost exclusively, but that generalizes very well to other types of time series. So it can actually -- like it's state-of-the-art across domains, not just for viability. And so it's actually really good at what they're forecasting even though it was trained on mostly application of the ability data. What's also exciting about it is that it's a model that scales, meaning that you can train a larger model from the same data set with the same architecture and even the same hyper parameters and you end up with high performance. And that's not something that has ever been achieved in time series models before. We have been achieving very famously in language models, starting like the most visible case of GP2, and we know what happened next. I think there's -- we start to see the same opportunity with these models. So now we're busy basically extending those models beyond time series, when I say time series. -- are mostly timestamps and values. We're extending that towards including the rest of observed data, whether that's logs and traces and application topology and network data and user data and all of that stuff, so we can really fully model and be predictive on our customers' infrastructure. And the idea there is if we build that properly, we can automate a lot of our customers' operations.

And I think the point being that like an LLM is designed to kind of represent human language. This is a completely it's a semi structured set of data that you would love to have a model like an LLM, but now this is a machine learning model optimized for that. As a result, its performance is much better than you can ever get out of LLM.

Yes. And also, these are -- like these are very different types of LLM are very bad time series for example. And so these are -- the architectures are even though the models are transformer based, so a lot of deep learning, transforms a lot of it is similar, but the architecture of the mall is actually fairly different.

Yes. Well, LLMs don't understand ordinality. So like you need a model that has that as a structure of it for this to work in different ways. We can obviously go much deeper on some of these things. But before we run out of time, I think it's also like important to perhaps reflect on some of the deployment model flexibility that you've been providing customer I mean it's almost like exhausting when you think of like the level of innovation you guys have been able to put out. I think that you could probably help us reflect on like from a buyer standpoint, 1 of their frustrations when they're forced to have a single deployment model that may not be ideal, whether or not it's for data residency or credits they already have. So maybe help us understand some of what you're doing with bring your own cloud and some of these models.

Yes. I mean -- and just to zoom out, customers in general, they face just an explosion of the amount of data and applications and everything, right? So -- and so there's -- in absorbability, there's a perpetual quest for efficiency, and making sure that would you spend on measuring auditing, rating your application remains predictable and within range of what you were expecting. For that, there are 3 core mechanisms. One is you need to have a feedback loop that lets you understand what you're consuming and what's valuable and what's not and what you need to do more or less of. And that's something we built big part of the product is making sure that the right people have the feedback loop at the people who care and who capture their information, see exactly how much value they're getting out of it and how much is costing them. So that's 1 part. Another part is to be to get ever smarter on how much of a sample you need to understand the data. And that goes back to all the smart building the models, et cetera. etcetera. And the third part is to keep innovating on how you can, how efficient you can be at storing and acquiring and managing the data and all the various topologies you can employ to make that work for various customers. So on that last one, One thing we started doing last year, and we see quite a bit of demand for and we're accelerating on is, we are doing some we call bring you on cloud options for our customers to deploy our product. meaning that they can store the data on infrastructure that they manage themselves even though we still run and update the application for them. And the reasons for that are -- so we see cost at large scale is 1 of them actually have petabytes of data, that's something that you probably won't. Other reasons would be data resiliency lows, and we see the world in general is fracturing into smaller zones as opposed to one big go. So we expect a lot more of that in the future. Even though today, it's still a native for most customers, we expect that to be more front of mind, maybe 1, 2, 3, 5 years. It's hard to tell when exactly. And the last reason might be that customers have -- maybe they have a large data center, they already built and they want to use. Maybe they have a large commitment with a specific cloud provider, and they want to use the what they've committed already for that. Like there's all sorts of reasons why customers might want to control more of the deployment there. Historically, we had focused on being purely SaaS and there were a few reasons for that. One is, I mentioned earlier that we thrive on being product-led and product adoption, meaning it's very important for us to be able to iterate very quickly. and also to understand what customers are using and not using and what's valuable and valuable. . So we never offered an on-prem version of our software early on for that reason. Like we thought it would be a huge tax on innovation. And also it would make us a lot more a lot number in terms of who we build product. And we've seen that play out. Like most companies that start having a product that has some success in the enterprise, get offers, they can't refuse early on of having buildings on-prem versions of their product. And that's a great idea in the short term and a horrible idea in the long term because those companies end up slowing down a -- the way we built this technology now is bring you on cloud, allows us to have the best of both worlds because we still manage and run and update those applications and meter those applications. But at the same time, we can run a data plan fully within control of our customers and on their infrastructure.

And between this and some of the stuff you guys have done with FedRAMP perhaps this even leads to some of the opportunity you could see the scale around because like if I look at many of the other software companies I cover, somewhere between 7% and 15% of their revenue come from federal. And I think you guys are -- if David's comments earlier, we're accurate closer to like 1%. So you could see even just on that 1 customer category, the type of revenue opportunity that it opens. We only have a couple of minutes left. You talk to a lot of investors, be talking a lot of these different events. What do you think the most underappreciated aspect of Datadog's forward-looking opportunity is?

I mean, there will be 2 halves to it. 1 is the world is exploding where they are right now. And this means so much more stuff so much more complexity, so much more infrastructure. And we see that play out with the large dialytic today, but that's going to come to the rest of the market. And so that's a huge opportunity for us. If there's only 1 category that remains, and I know everybody here has been debating the software ever going to be worth anything ever again. Well, I mean the answer Genies yes. But the -- if there's only 1 category that is left in the world, it's observability because the 1 job that remains for us humans is to control the machine and understand what it's doing, understand whether it's delivering the right outcomes, how much it's costing and whether the intent of the -- of what we wanted to do is actually what would get performing in the end. So I think that's a huge opportunity for us. There's a lot to be there. It's a challenging market because it's changing really fast. So there's a -- that muscle we built on product innovation, I think, is going to be brought to bear, but that's -- we see that as a huge opportunity. The other half, the more boring half is that all of that is driving the need for standard observability. Those large AI companies are consuming large amounts of infrastructure and applications in the same way the other companies are -- and as I said earlier, like we have less than 14% of that market, and that market is growing and is going to grow faster, I think, now as the -- as you see this huge CapEx investments in infrastructure and general and it's not use GPUs anymore. It used CPUs and everything else as these agents need to use tools. And so I think we -- that part of the business, the boring the existing part of the pan, that's not new form factors, your plication or anything. That part has another 5 or 10 x into it, and that's very exciting.

Well, you heard it here first. The only business that will remain after the robots take over everything else is observability in Datadog. So make sure that you're ready to be an SRE in your future. Oli, thank you so much. I really appreciate you taking the time here and sharing with the audience your experience and kind of where Datadog is going.

All right. Thank you.