Fantastic. Well, welcome Olivier back to your hometown Paris.
Yes, not quite. I'm from the suburbs, but you know, close enough. Yeah.
And big game tonight by the way, 10:00. So, I'm cheering for you as well.
So, for those people that don't know, basically Datadog is like a massive digital watchtower that looks at all your digital systems. They basically look at all the metrics that a website gives out, all this data information, and make sense of it, and it sends signals to all the engineers in the companies that pay for them, so they can figure out if something goes wrong, how to fix it. So, I think 3 billion revenue was the wrong estimate, right?
Yes, technically we're between 4 and 5 billion now, growing about 30%.
Well, pretty incredible. This company got started over 16 years ago, and we'll talk a little bit about what Olivier has learned about himself, but CEO co-founder for 16 years, growing at 30%, where has this growth come in at consistently 30% and such high revenue?
Well, there's lots happening right now, you know, I mean, you've heard about the whole AI thing, and it's a, you know, look, the way like in terms of problems we solve like we solve the problems of engineers having to deal with complexity. Like they have more machines, more applications, more services, more everything. And that keeps scaling up over time. Over the last 10-15 years, everybody was moving to the cloud, and that was generating a crazy amount of scaling and complexity, and everything was becoming digital, and that drove our growth. Today, companies are building and adopting AI, and that's adding another layer of growth that's touching pretty much every single part of our business. So, we've re-accelerated growth over the past six quarters, I think. And in great part, that's because there's just so much that's being built, so much that needs to be watched and monitored and secured and kept running, basically.
And what's super interesting I did a couple of sessions yesterday is that obviously with the AI growth is just more data. You have more metrics. You have more traces, more logs, more information. So, how are you adapting to the shift? And what's the impact on customers? So, you have a lot of CFOs who are saying, "Hang on a sec. My AI cost is going through the roof. Does that affect also your margins? Can you keep up that 30% growth?"
Yeah, I think all in all, it's a big opportunity because everybody right now across the whole customer base, from the startups to the AI natives and with all of the big enterprises like the big banks and train companies and whatnot in the middle, everybody is busy adopting AI. And everybody is seeing their cost skyrocket. And everybody is also finding that they're not in the position where they can really understand what's actually needed. Are they spending the right amount to get the right outcome? What are the outcomes, by the way? Like it's super hard. When you zoom in into specific parts of the business, you can see the outcomes, you can see value, you can see things. When you try to add everything up across the whole business, it's a lot more difficult. And I expect that to remain difficult for maybe a year or two, basically. But for us, that creates a huge opportunity because we can be there to help understand who's doing what with AI, what outcomes they're reaching with it, how much they're spending to do that, and how they can get better at it. One analogy I would give is right now we just invented air travel and we're sending postcards by buying business class tickets for them. That's the way we use AI models in general.
I think we can be so much more efficient. There's so many more things we can do for so much cheaper that there's a lot of room to optimize.
But the cost issue though, in terms of your margins, are you getting challenged from people saying, "Hang on a second, like I want more data, more data, more insights, and you've got to manage your cost. Are you increasing your prices or"
No, we're usage-based. So, if customers send more data or if they use more, basically that's built into the model. We're not seat-based like a lot of the traditional SaaS companies are seat-based and they're challenged in that maybe they get less seats because they have more non-human agents using their systems and maybe they need to bundle more functionality within those seats. We're usage-based. We're still bound to the same dynamics, right? At the end of the day, we have to show value. And there's only two reasons people buy software. They buy software to make more money or to save money. At the end of the day, that's it. So, we still have to deliver that.
But the business model itself is amenable to adding a lot of AI functionality and getting a lot more data.
Yeah. It's pretty interesting as well because recently you've made a pretty good acquisition as well. You bought an ML business that allows you to create agents that can also fix. Can you tell us a bit more about that acquisition?
Yes, we had acquired a company called Adaptive ML, which also happened to be headquartered in Paris. And Adaptive is what the product does is it helps you train your own models based on evaluations that you build with maybe some frontier models or other models you can already use in your product. And basically allows you through reinforcement learning to own the models and reach a different level of performance and cost. That's amazing. And that's something that we're actually not continuing that product. Like, we're not selling that product anymore, but that's technology that we're using internally to build our own models. Whether that's today models that replicate what you might do with some of the frontier models. So, basically thinking models, but specialize on our use case that can be 100 times faster and 10 times cheaper. So, that's definitely something we want. But in the future, we also have our eyes on building models that can be end-to-end world models for our problem space. So, instead of a post hoc investigation, say for example, I've had an outage and I'm asking, you know, Claude or ChatGPT to explain what happened and they're going to,
call some tools and reason and give you something. Instead, think of a model that takes your telemetry data as an input. So, say you send your metrics, your logs, your network telemetry, your code, your traces, and everything into the model. And then, that comes in on one end, and on the other end, what you get is, "Oh, this is what's going to happen in the future, and that's the problem, and this is what you should do about it." And the equivalent there is the I think of the self-driving car models. Like the self-driving car models are all end-to-end now. It used to be you had all these subsystems and those rules and everything else. And now, instead, what you have is models that take cameras and maybe lidar in, and then, what you get out is pedals and steering wheel.
But, that's a super interesting shift. So, basically, again, this concept of self-healing software. So, it watches all your systems, it figures out what goes wrong, and then, in the past, you would just watch it, but now you're actually fixing it. So, in a way, Datadog is both the judge, because you're judging what's going on, but also, you're the actor, because you're also now fixing it with your own AI agent. So, are you conflicted in any way? If not, it's a great way to make money. I'll create my own bugs, and I'll charge a lot to fix them myself.
Yeah, yeah. Well, I mean, that's a good question, because if you zoom out a little bit, the need to separate the actor from the judge is precisely why our customers buy us, right? Historically it was because our customers were using many different cloud providers, many different applications, and then they need one single way to observe, to secure, to manage everything. Today they're doing that with AI models. They're going to use many different AI models. I think it's pretty clear that nobody is going to be all in one model, all in one OpenAI, all in one Anthropic, all in one some open source. They're going to have a mix of different models. And they will need us to observe, manage, secure, and verify the correctness of all those models across.
And where do you see I think one quote was 40% of all cases can now be completely fixed autonomously if you want to use the autonomous car example. Do you see us get to a place where everything is 100% or do you still see human intervention coming in?
Oh, it's a couple of angles there. The first one is self-healing software is actually a lot harder than self-driving cars. And the reason for that is the roads pretty much all look the same. There's not a high level of variability. They don't change very fast. So if you think of software, imagine, if every year the road got more complex, if every year you added one branch to each intersection, or you had cars jumping on you, all these things that make the model so much more complicated. And by the way, the roads are changing so fast like software that you can't train a model and start using it for a few years and start using it afterwards because your model from a few years ago is obsolete. So self-healing software is a lot more complex.
So what self-driving car?
At the same time, look, self-driving cars were built like the technology largely works. Like we're in the very final fine-tuning of it, but if you climbed into a Waymo, for example, the technology works really, really well.
And what level of trust would you need to have for you to feel comfortable on an AI independent autonomous way to fix thing? 50%? 90%?
Way more than 90%. And the reason is twofold. You need to be right way more often than a human. Because
you get it wrong once, the customer is going to pull out.
Yes, exactly. Like, you will tolerate being sent the wrong direction by a colleague. That's fine. That happens all the time. By a machine, not so much.
