AI Agent 需要计算机：月环比增长74%、日运行85万次与全新 Agent Cloud——Ivan Burazin，Daytona

Ivan Burin

I've never experienced this: people literally call you if you do not give them access. They want access right now. So it's like, okay, they don't want this. The thing that they want doesn't seem to exist, or they have not found it, and they really, really want what we want. And then, when we understood that, we knew we were onto something.

When you think about the size of the market, the market for every single agent that will ever exist in the future is just like—what is that market? How big is that?

Before we get into today's episode, I just have a small message for listeners. Thank you. We would not be able to bring you the AI engineering, science, and entertainment content that you so clearly want if you didn't choose to also click in and tune into our content. We've been approached by sponsors on an almost daily basis. But fortunately, enough of you actually subscribe to us to keep all this sustainable without ads. And we want to keep it that way. But I just have one favor to ask all of you. The single most powerful, completely free thing you can do is to click that subscribe button. It's the only thing I'll ever ask of you, and it means absolutely everything to me and my team that works so hard to bring the Inspace to you each and every week. If you do it, I promise you, we'll never stop working to make the show even better. Now, let's get into it.

Shawn Wang

Okay, we're in the studio with Ivan Burin, CEO of Daytona. Welcome.

Ivan Burin

Thanks for having me, man.

Shawn Wang

Ivan, you and I go back—

Ivan Burin

Way back.

Shawn Wang

I don't even know how you found me. Did you reach out, or was it for Shift?

Ivan Burin

I reached out to you. The reason was that we were thinking about—I was one of the co-founders of Codeanywhere, the first browser-based IDE, and we were thinking for a long time that localhost should die. You had this article about localhost, and then I reached out to you because of that.

We talked, and I was actually at a different job. I was the head of developer experience, and you were quite well-versed in that. I reached out to you, among other people, asking, “How do we go about that? What are the key things?” You were nice enough to take the call.

I remember I was late for your call with you.

Shawn Wang

I don't remember.

Ivan Burin

I remember because I was with my then—I'm not sure if she was my girlfriend or wife at that point in time. It's the same person, so that's great. We were in Italy on vacation, and I was late for something. I felt so bad, and you were so nice about it.

Shawn Wang

The reason I'm nice is because I'm also late to other people. So it's like, who's without sin here?

For those who don't know, Infobip Shift was this whole thing that you did in the past. That was basically one of the inspirations for me starting AI Engineer. I have to thank you for giving me that push to be like, “Oh, you can build and sell conferences.”

Ivan Burin

Yeah, and I remember you asked me at the beginning to give you advisory shares. I was so focused on what we were doing that I said no, and I should have taken the advisory shares. I'm sorry, but anyway—

Shawn Wang

We're not venture-backed, you know. Anyway, I think what's impressive about you is that Codeanywhere is the thing that you've been trying to build. You put it on hold and came back after Infobip. Just give us the story—the origin story going into Daytona.

Ivan Burin

Sure. Really way back, my co-founder and I have been together. I've said this multiple times: it's like we were married, divorced, and married again. Some people actually ask me if my co-founder is my partner. They thought it literally—it isn't literally—but we have done multiple companies together.

To your point, we had this shift where we went from—

Shawn Wang

Codeanywhere to the conference called Shift and then back to Daytona.

Ivan Burin

We originally started stacking servers and doing virtualization in the early 2000s—routers and all these things at a foundational level. That was a services company, which we sold to focus on what my co-founder actually invented, which was the very first browser-based IDE.

I say the first, but before us there was actually Heroku. They did it for a very short time until they became Heroku. Outside of them, we were the only one, and it was called—

Shawn Wang

Cloud9.

Ivan Burin

There was Cloud9, which came out slightly after us. There was also Replit, which came out when we stopped doing it. Replit came out, and they have been successful since then, which is great. There was Nitrous.IO. There were quite a few that existed at the time, but it was too early.

The interesting part is that, at that point in time, there was no VS Code, for those who still remember. There was no Kubernetes, and Docker had just started when we began. I'm not sure if it was even public at that point in time. We had to build everything in the whole stack ourselves, and that was the key learning that we brought into—and that we've been using in—Daytona today.

About 3 million people used Codeanywhere. It was slightly more angel-backed than venture-backed. We ended up paying everyone back because it didn't have that sort of scale. Three years ago, we started something similar with Daytona, which was not what we are today. It was automating development environments for human engineers—the underlying stack of Codeanywhere—and then we did a hard pivot last January to sandboxes. So here we are.

Shawn Wang

Historic pivot. I independently invested in Codeanywhere, but also in E2B, and then both of you pivoted into the same thing. I'm like—

Ivan Burin

You invested in Daytona. You invested in Daytona, but you were the first. If we had not gotten your check, we wouldn't have done it.

Shawn Wang

No way.

Ivan Burin

No, it was like, “We have to get him on board first,” and you were that kicker that got us on the—

Shawn Wang

You were putting me on your pitch deck, man. I was like, “Man, this is a good trip if I don't invest.”

Ivan Burin

Well, that's because it was your quote. We did a bunch of research about “The End of Localhost” and who was interested in that.

Shawn Wang

Yeah. No, it's like, I wrote that blog post, and every single company in that field reached out to me. Then every VC who was receiving those pitches also had to call me and talk through it with me.

Ivan Burin

It's finally happening. It's finally happening with maybe nonhuman users.

Shawn Wang

Yeah, yeah, yeah. So what is Daytona today? Let's get a quick description. I'm wearing a shirt.

Ivan Burin

You're wearing a shirt? Yes.

Shawn Wang

It says—I think your branding is very good. It's very consistent. “It runs AI code.” It cannot be simpler.

Ivan Burin

Exactly. But we're probably going to have to change that because it's also a subset of what we do. Unfortunately, we really love “Run Code.” It's super simple, and people interpret it in different ways.

I think we've given out 5,000 or 6,000 of these shirts. People wear them with pride because it doesn't really market to us; it markets to the person itself. I think we did a really good job on that one, but it's also a subset of what we do.

When people think about “run AI code,” they just think about these small, let's call them isolated code-execution boxes: you send some code, and you get an output. Whereas what Daytona is today is essentially composable computers for AI agents. The market calls them sandboxes, which is misleading.

Yeah, exactly, because it can be misleading. People usually think about sandboxes as a demo or a test environment versus a production-grade environment.

What Daytona does, if you think of the laptop that you have in front of you, the computer that's over there, or my wife's computer—she's an architect, so she has a Windows computer with a 3D graphics card inside to do 3D rendering—computers are different compositions of computers.

Our belief is that agents today and going forward will need all these different compositions of computers to do different types of tasks. We offer that through an API.

Shawn Wang

To give people the aha moments and the wow moments, the market is exploding. You've been reporting 74% month-to-month growth, and it has been going like this for a while. It's not just you guys; it's every single compute provider.

I don't know if you agree with me calling you a compute provider or not, but it's organically PLG-driven growth, and enterprise is doing super well.

I want to rewind to January last year, when you did the pivot. You obviously called this market early, you were positioned for it, and you are now one of the market leaders. What was the insight that made you do the pivot?

Ivan Burin

The insight that made us do this pivot came the quarter before that. At the end of 2024, we basically did a demo with Devin. I don't think we discussed this as well. Devin was not public.

Shawn Wang

You actually gave me access to Devin at that time.

Shawn Wang

I did. Yeah. I don't think I was supposed to.

Ivan Burin

Yeah. Exactly. So it doesn't matter.

OpenDevin was available, which is now called OpenHands. We were like, “Oh, this seems to be a thing. This is not public. Let's take our platform for human automation of development environments and launch that as a SaaS.”

Not very many people signed up and used it, but a lot of people reached out who were building agents. They were like, “Hey, my agent needs a compute sandbox runtime,” whatever you want to call it. I forgot what it was called at that point.

Then we were like, “Oh, amazing. This is a new market. Here is our infrastructure. Here's our product. Go.”

And what we found really, really fast was that people did not like what we had built. It didn’t work. I remember talking to people at the beginning, when we were doing this—the sandbox we were building for agents. People were like, “Why is it different? It’s the same thing. We have EC2, we have VMs, we have all these things.” But we saw that everyone we gave it to—20 or 30 people—said, “No, this is not what we need. This sort of breaks.”

Basically, my co-founder and I didn’t know a lot about AI because we’re infrastructure people. We’re not AI people. So I took it upon myself to watch every single podcast that exists, including all of these and others, get up to date, read all the blogs, and understand what was going on.

Shawn Wang

Do you want to shout out who else was useful, just in case people are also looking?

Ivan Burin

Generally, I looked at a few podcasts in different segments and of different types. There’s you guys, No Priors, and Bill Gurley was great while it was around. 20VC is interesting from a different dynamic, and some have different dynamics.

Shawn Wang

But we’re not really about the compute market.

Shawn Wang

I guess you’re looking at the agent infrastructure market.

Ivan Burin

I was looking at the agent market and the AI market in general, understanding who the players were, what the perception was, and how that worked. Obviously, you complement this with going to conferences, going to events, going to meetups, reading white papers—doing all the things that you have to do to understand what’s happening.

When we sort of had an idea of what we had to build, literally on New Year’s Eve, I half-vibe-coded the first MVP—the first minimum viable product—of what Daytona is today. I went to sleep at around 3:00 a.m. I had just put my baby daughter and wife to sleep, said, “Happy New Year,” and went back to doing this. I sent it to my co-founder, my CTO, and he saw it in the morning. He was like, “This is absolute garbage. Do not show this to anybody at all.”

But the idea is good.

So he took 2 weeks and made it look like that. It wasn’t even like that—it was way worse—but it was a simplistic view of what it should be like. It worked, but it wasn’t ideal. He went and found the hole, which is his job as CTO, and he came back with this version.

We then called all the people who had said, “This is garbage,” a quarter ago. We set up these calls and just demoed it to everyone. All the calls went long—every single one. They were 15-minute calls, and they all went 25 or 30 minutes. Everyone said, “We need access.”

There was no login, just an API key, because this was a beta or an alpha. They said, “We want access,” and we said, “Sure, yeah, okay. Thank you very much.” But the next day, if we hadn’t sent it, every single one of them came back and said, “Where is my API?” Everyone wanted it.

We were like, “This is it.” I’ve never experienced this. The understanding, to your point, was that most people thought it was the same infrastructure for humans and agents. We understood a quarter ago that it wasn’t; we just didn’t know what the right primitive was. Then we came up with what that was, and we gave it to these people.

I’ve done multiple companies in my life, and I’ve never experienced this—people literally calling you if you don’t give them access. They want access right now. So it was like, okay, they don’t want this. The thing that they want doesn’t seem to exist, or they have not found it, and they really, really want what we want.

When you think about the size of the market, the market for human engineers in the enterprise is a very large market. Think GitLab or something like that. But the market for every single agent that will ever exist in the future is just—what is that market? How big is that? We were like, “We are all in on this.” That’s where we made the cut between the old product and the new one.

Yeah, but it wasn’t composable at the time. It was basically just a Linux box that you could change, where you could define the number of CPUs, disk, and RAM. That was what you could do, but you couldn’t have multiple operating systems, resize it on the fly, add a GPU, or do all those things. It was just the first variation of it.

And was it bare metal from the start?

It was bare metal from the start.

Which, you know, gives people the background: What is the normal path?

Most providers run this on top of VMs.

Firecracker.

Yeah, they run Firecracker on a VM. We also have Firecracker—we can get to that. We have multiple isolation layers, and we can do that.

The common way to do it is that the state of the machine, or the hard disk, is not part of the sandbox itself. The other thing is that they’re not meant to last forever. Most of them are preemptible; they can live only for a certain amount of time.

Our thought was that agents would be like humans, in the sense that you don’t want your laptop to be shut down until you’re done with work. You want to close the lid and open the lid, and have it be in the same state. Agents would want that: to pause and come back. They want those 2 things. But agents also really, really want speed, right?

When we thought about it, we needed something insanely fast. How do we make it fast, long-running, and stateful? It’s like combining a Lambda and an EC2—those 2 things together. We didn’t have any idea how others did it because we didn’t know there was much of a market around this. It was more like, “Okay, this is what we need and what they need.”

We looked at Kubernetes, but it wasn’t good enough for that. We looked at Nomad, but it didn’t enable that. Our history of rewriting our own scheduler at Codeanywhere is basically what my CTO came up with. He brought over the learnings from there.

The funny thing is that our third co-founder, when he saw it, was like, “Dude, what is this? This is like 2008. We went back in time.” And he was like, “Exactly.”

The reason why Daytona is super, super fast, and why you see this on benchmarks, is that we essentially run on bare metal. We have our own scheduler, and we use the underlying disk, CPU, and RAM of the underlying machine. That means your IOPS are insanely fast because there’s no network between it and EBS or something like that.

The snapshots—the point-in-time templates—are also preloaded on the bare metal machines. When you fire off a sandbox from a template or a snapshot, you’re essentially directed to the bare metal machine where that snapshot is based on the NVMe drive. It literally just turns on that machine, and it’s local. There’s no network latency or anything like that.

Those are the specificities we came up with when thinking from first principles about what a computer would look like for an agent. That’s what we came up with, and that’s what we created.

I should maybe—I don’t know if you endorse this—but there’s someone who does Compute SDK benchmarks. You guys do very well on there, with the TTI. Is this a relevant benchmark for you guys?

I don’t know, and it changes every day. So today, I’ve never heard of it.

But you are at least a third of the next tier of performance, and there are a lot of other better-known names that are very slow to scale.

Yeah, we’ve been number 1 by far for a long time. Now there are different definitions of sandboxes, different isolation patterns, and different other things.

Archield[?] runs it literally on S3—the data—so it’s a very different thing, and they spin up a container for that. It’s a different type of thing. The definition of a sandbox is something that we all need to get aligned on.

But yeah, we’re insanely fast at getting these things up and running. You can see even there that it’s 0.101.

Close enough.

Yeah. I mean, what else do you need, right?

The benchmark itself—I don’t think benchmarks equate to market ownership or revenue or anything like that. I’ve seen this with multiple benchmarks, not just in sandboxes, but in general benchmarks around this. It’s table stakes.

Exactly, but it doesn’t mean market ownership or revenue. You definitely have to be up there and competing so that people know, “Oh, this is definitely one of the top.” This is only 1 dimension of what customers look for.

There are other things, like how many you can spin up consecutively. There’s a feature set, support, and all different things that people look at. But you definitely have to be there on the benchmarks.

How many do people spin up concurrently?

There are 3 metrics that we look at. One is the time to spin up 1. Our time to spin up 1 is 60 milliseconds with network latency. So request, spin up, reply—60 milliseconds. The whole thing is 60 milliseconds. That is 1.

But if you want to spin up 50,000 at once, we are now at about 75 seconds. So it takes about 75 seconds to spin up concurrently 50,000. Some others—there’s public data around this—take 2,000 seconds, which is 30 minutes. There are different variations of that.

Then there is the speed of 1, the speed of multiple, and how many you can consistently have up and running. We basically have no limit right now to how much we can add because we own our own metal. Our biggest customer does about 850,000 every single day. They’re just shy of a million every day that they’re running. We do have a request for half a million concurrent, which is literally half a million CPUs running somewhere. So that’s interesting.

And they pay by vCPU-seconds, yeah?

Yeah.

The other thing is the sleeping and resuming, because it’s all the stateful resumption of these things. What kind of workloads are people putting through this? Do we measure by gigabytes in memory, gigabytes in storage? I don’t know—network, attached storage. What are the costly ones out of all these features?

The most expensive thing is CPU. Then it’s RAM, then it’s disk.

Which is snapshotting, right?

No, the snapshot is part of it, but basically it’s the size of your machine’s hard disk. Do you have 10 gigabytes, 20, 50, or whatever? And then there’s the transference of that. Currently, we don’t charge the customer for network at all.

Yeah, you’ve got to fix that.

Yeah, it is very much a larger and larger part of our bill, so we’re working around that. Obviously, that is the least expensive. The hard disk is the least expensive. So it’s basically CPU, RAM, network—which we don’t charge the customer for—and then hard disk. That’s how it’s spun up.

There are also different types of workloads. We basically split them into 2 types in Daytona. One is what we call background agents, or long-running agents.

Okay.

The other is basically RL and evals, which I put together. If you look at the usage of a background agent—and I’ll just name some companies—their usage patterns are similar to humans, which is follow-the-sun. Basically, noon is probably the highest and midnight is the lowest, and weekends are low, whereas weekdays are higher.

Yeah, OpenHands.

Yeah, yeah. Background agents—Cognition, Lovable, all these things, Harvey—these are all long-running background agents. If you look at their usage patterns, they’re similar to human usage, which is follow-the-sun.

Shawn Wang

That’s a fun question. How global is it? Is it very US-centric?

The US is a large part, but currently we have Asia, Europe, the EU, and the US. It’s quite global. We have it all. Our number-one city by users is—

Shawn Wang

Singapore.

Oh, wow. Which is interesting, right? Not by revenue, just by individual headcount. It’s just interesting.

Shawn Wang

Singapore is weirdly high in the adoption charts of AI for the population. It’s a 7–8 million population—

And it keeps showing up. No, it’s quite interesting. We were quite shocked, and I was like, “Oh, this is interesting.”

Shawn Wang

There’s a reason I’m doing this in Singapore.

Yeah. I mean, we’re there. We’re going to be there as well. It’s interesting that Japan is in the top, or Tokyo is in the top, which, in all the tech cycles, it has never been. It has never been in the top, so it’s quite interesting.

Shawn Wang

I think the Japanese just love AI.

Yeah. It’s that, and then there’s Brazil.

Shawn Wang

Yeah.

But Brazil has always been in the top. Even when I look at GitHub’s data, and historically with Codeanywhere, it was always the US, Western Europe, and then India, Brazil, and China. Those would be there, but Singapore was not. Specifically, Japan was never in the top.

Shawn Wang

So, actually, that helps you distribute your load through all time, you know.

Yeah. The interesting thing is that we have those kinds of loads, but if you look at the researcher workloads, they’re quite different. If you give them a concurrency of 10,000, 50,000, or 100,000 CPUs, whatever it may be, when they fire off a run, it just goes to 100% and then runs and runs and runs, and it stops. The usage pattern is basically squares, right? It’s also not follow-the-sun, because people will fire it off at midnight before they go to sleep, then wake up, so it’s very unpredictable and you don’t know where that is. The shapes of the usage are quite different from what we’ve had before.

What’s interesting is that with a follow-the-sun pattern, even if you have a high-growth company, you can predict your usage patterns and have enough capacity for that because it grows in a way you can project. When companies do evals and RL, they’re super spiky. They’re going to come in and say, “We’re going to use nothing. Now, can we have 100,000?” Then they go back down, and then it’s 100,000 again and back down. It’s very different.

You’ll rarely get a spike that is 10 orders of magnitude. You’ll get a spike—let’s say one of your customers has some exponential curve. What is that to? I mean, I’m using cloud as an example: 10%, 20%, whatever. I don’t have this data; I’m just assessing. It’s surely not 10x, right? It’s surely not something like that.

Shawn Wang

So do you want to lock them into commitments so that—

Yeah, we do. We have to lock them into some sort of commitments to have that capacity, because we basically have to have the capacity for peak.

Shawn Wang

Yeah.

Right now, Daytona’s mean utilization is 15%.

Shawn Wang

Oh, my God.

So it’s very low.

Shawn Wang

Because it’s very spiky.

But it’s very spiky, and we get up to 90%. What we’re looking at right now as a company is something similar to Cloudflare, where you can geo-move things around. That works really well for background agents, where there’s a follow-the-sun pattern, but this is a very different shape. Obviously, with scale you figure these things out, but that’s an interesting new problem that we have as a compute provider in the agent space.

When we were doing the conference recently, we talked to Nikita from Neon and Parag from Parallel. Everyone has the same problem: usage is super spiky. This is something that has not happened before. The amplitudes were never this high, so it’s quite an interesting use case and problem to solve.

Shawn Wang

I don’t know if we’re going to bring this up again, but let’s just talk about the conference. You had 1,000-something people at the Warriors game—sorry, where is it? What’s the—

Chase Center.

Shawn Wang

I went. It was very impressive. Obviously, you know how to throw a conference. What did you learn? You pulled together all these impressive names. What were you looking for?

My thesis behind the Compute Conference was: let’s bring together people who are building infrastructure for AI agents. When I think of what we’re building, the agent is the primary user. What are the ergonomics and usage patterns of agents?

What I found—this was a theory; it wasn’t proven—is that we all have these problems. As I touched on, we all have the same underlying infrastructure problems: spiky, unpredictable workloads that we’ve never had before in human compute or human infrastructure.

It was the same when I was talking to Parag, or when I was talking to Lin and Nikita. Everyone has the same problem. Lin especially—I was talking to her the other day as well. It’s a very interesting type of problem to solve.

I can touch on Cloudflare, because there’s a lot of talk about that recently. They have a bunch of geos, and as users work in different places, depending on your tier, they can move you around the geos. That’s how they get to higher utilization. You can sort of predict these things, and you’ll rarely get a spike that is 10 orders of magnitude. You’ll get a spike—let’s say one of your customers has some exponential curve. What is that to? I mean, I’m using cloud as an example: 10%, 20%, whatever. I don’t have this data; I’m just assessing. It’s surely not 10x. It’s surely not something like that.

Shawn Wang

So she also has the same thing?

Yeah. I know specifically that Neon had that issue as well. How are we solving these spiky loads and things like that? We talked about it, and the interesting thing for me to internalize was that, yes, everyone who’s building for agents first is going through this, and we’re all solving similar problems.

Shawn Wang

Let me double-click on this. For example, Neon—I happen to know that they’re very S3-oriented. They’re fully betting on S3, and you get to benefit from S3’s distribution infrastructure, so I would imagine that Neon doesn’t have to care as much.

Whereas Lin maybe has to care a bit more, because obviously she’s doing GPU inference. For listeners, we did an episode with her 1.5 years ago.

Shawn Wang

And you have to care, but, like, right?

Parag cares for sure.

Shawn Wang

And Parag is co-founder of Parallel, formerly CTO of Twitter. They’re the search company, for listeners who don’t know.

We can put it up on the screen so people can look it up if they need to.

And yes, they still have CPU and RAM allocation that you have to have running. So there are basically 2 ways to do it.

One is, you either overprovision and can handle the bursts, or you basically have—I don't know if this is a term—just-in-time compute. As your usage comes in, you can fire off requests for VMs or bare metal at other cloud providers and then get them up and running.

Shawn Wang

So this is if you go above 100%, right? Like your overflow—if your overflow, like spillage or whatever—you probably lose money on it, but it doesn't matter, right?

Well, you might. You might not. That is a more cost-effective way to do it, but it's a slower way to do it, because basically what you have to do is queue your requests, spin up this just-in-time compute, get it all ready, provision it, and then get your workload there. If the time isn't that important, that's fine and you can do that.

But if your customer—and especially for, let's say, the RL training runs—the reason why a lot of people come to us is because GPUs are more expensive than CPUs, right? So you want your GPU running at, what, 100% the entire time. When you're running runs on CPUs, when the CPU cycle is down and spinning up the next one, you want that to be instantaneous so that your GPU doesn't go down, right?

If you then have to go out and provision machines, you're essentially telling the GPU that it has to wait, and that's incurring our cost. So there are things that you have to try to solve for.

Shawn Wang

Yeah, let's talk about the different workload, right? You said that a few months ago you had zero RL workload, and now it's 50%.

It'll be 50%.

Shawn Wang

Let's talk about how different it is, right? I imagine, for example, a lot less dynamic code generation of arbitrary code. Here, it's probably all the same code; you're just doing parallel runs or something.

Yeah. So you'll have multiple—depending on the... For each run, you'll have a snapshot. For the most part, they actually do use our declarative image builder, which is like, “Oh, the agent wants these dependencies, these env vars.”

Shawn Wang

Yeah. Declarative image builder.

It's a very Modal-like thing. And so we build it on the fly, and then we propagate that snapshot. You can spin up as many sandboxes as you want against that snapshot. If you have to make changes, the model can do it, or it could all be automated. It's like, “Oh, now for the next run, we need to install these things or remove these things or whatever to get a task done,” and then it goes off and runs that. So, yes, that is something that it seems they prefer.

The number one reason I found—or, should I say, let's take a step back—is that what we are competing against in that environment is essentially managed Kubernetes.

Shawn Wang

Yeah.

So EKS, GKE, whatever. That is what the vast majority run on. Anyone who has tried Daytona versus GKE or EKS is like, “I'm never going back.” There are a few reasons.

One is the ergonomics. If you're using Kubernetes to spin that up, you have to essentially manage the interface interactions with that. Daytona, although it's a compute provider, is more akin to Twilio and Stripe from a consumption perspective than it is to AWS. You have an API and SDK; it's quite easy and seamless to get these things up and running.

That's one. The other is the speed at which we spin up, which we mentioned earlier, which is much, much faster, and the scale to which we can go. We haven't gotten into features, but an interesting feature is that it's very hard to have our sandboxes OOM, or run out of memory, because we can dynamically resize them on the fly, which is almost impossible on any other platform. There are some technologies that enable you to do that, but it's a very hard thing.

We actually saw this when the Terminal-Bench team brought us into this whole space. So thank you, Alex, and the team. They brought us into this whole space.

It is very, very, very rare that a framework would just say, “Guys, just use Daytona.”

Yeah, I think it says it somewhere.

Yeah, I was like, “What is this?”

There are multiple mentions, but they also mention a few other places.

Yeah. And so, Daytona specifically—we're just jumping on themes here—I don't know where it says Daytona.

Right. I don't know. There's a very, very strong recommendation, which is very unusual.

We do not pay them for this.

Yes, I know. They just like you.

Yeah, they like us.

Daytona has multiple isolation levels underneath. The customer doesn't have to know what they are, but basically we have Docker, which is a container that's hardened with Sysbox. So it's Docker's isolation that's security-equivalent to a VM, but it's still a container, and that is the default.

They especially, in these training workflows, really like that as an interface—to be able to use just a basic Docker container. We enable Docker-in-Docker, which for these RL runs—if you need to do a Docker Compose or Kubernetes—you can spin up k3s inside these things. That unlocks a huge amount of workloads that you can do that you cannot do on other providers.

That part is much more interesting. We went through that, showed them that we could do it, and they enjoyed that quite a bit. They being the Harbor people.

Do you know, are they a company yet?

I do not know.

All right. It's super obvious that there's a lot of excitement and success around these things. Tell us more. This is an exploding workload. Harbor adopted you, which helps speed things along, but what are you learning as this new workload comes online?

Sure. There are a couple of things that we learned, which we chatted about in the beginning, and this has led our story. As we mentioned, we talk to a lot of customers along the way, and we add more features and more toolsets as we talk to customers.

I think it's that the ecosystem is so small, or the models get smarter, where when we see one user come with a request, we know it goes on a roadmap if 3 to 5 customers come with the same request in that week.

It's very bizarre, and it happens so many times.

Because they're all friends. They're all in the same group chat.

Yeah, probably. Yeah, because they're like, “Oh, can you do this?” We're like, “Okay, this is interesting. We'll put it on a feature request.” And then the next one is like, “Oh, can you do this?” It's all the same, right? So it's the same.

What we try to do—and I personally try to do—I try to be on as many quote-unquote sales calls as I can. I'm in every Slack channel. We literally have about 1,000 Slack Connect channels, something like that.

It's interesting. There are so many interesting things you find out when you have all Slack channels. You can also see where people transfer between companies: you see them leave a Slack channel and enter a Slack channel. It's an interesting thing.

Also, I digress: I feel that Slack Connect is literally what LinkedIn should be.

Yeah. You have a list.

LinkedIn charges you to use your own connections, but Slack doesn't, right? Slack is like, “Do it for free.” It's more lock-in. It's great.

Yeah, it's amazing. It's one of—

You're going to pay Slack for life.

Exactly. You're there for life.

So that's interesting. One of the newer things we talked about earlier is that we made a big bet and put a lot of investment into computer use that has not seen the light of day publicly. We haven't gotten that yet.

Is there a thing I can pull up?

There is computer use there. It's right up a bit.

Yeah. Okay. Yeah, cool.

What we've talked about and what we've seen publicly is this theme now about the human emulator, where Elon Musk from xAI has talked about this publicly. If you think about the models today, they're actually quite sophisticated and can do a lot of work, but they still don't have access to all the tools.

I'm a strong believer that the most efficient way for an agent to work is essentially headless or through a terminal or whatnot. But if we look at knowledge work in general, there are about 100 million knowledge workers in the U.S., about 1 billion in the world, and their salaries aggregate to $10 trillion in the U.S. and $50 trillion worldwide, something like that.

And if we look at the 5 most important sectors of that—healthcare, government, financial services, and whatnot—that's about 56% of it. So let's say it's about half of that. Worldwide, it's about $25 trillion.

How much of it?

Our assumption is the following: in the RPA market, which is a similar market but not the same, 25% of these white-collar workers' work is automated. If an agent is more sophisticated, can go through more runs, and figure stuff out, let's say it's 40%, right? If you take 40% of that, you get to essentially $10 trillion.

A year. That's a T.

That is a T. So that's the TAM of the models, right? That's not essentially ours, but you get to that size. To be able to do that, you essentially have to give agents these computers with the legacy apps.

So, computer use—either Mac, Windows, or Linux. We obviously have Linux, and others have it, but Windows specifically is something very, very new. The only option right now is an EC2 with Windows or on Azure.

Both of them take anywhere from 3 to 5 minutes to spin up. We've created an actual sandbox, so it's a second instead of minutes, but you have point-in-time snapshots, forking, and all the things that you have from a sandbox. It essentially enables you to hopefully unlock all this value.

That's been our big push and bet. We've kept our ear to the ground to understand what the next things in the market are.

Yeah. Knowledge work and building, and sort of RPA—the next wave of RPA. I got very excited about RPA during COVID times. UiPath was IPOing, and it was like a very hard—it's Eastern European, isn't it?

It is Romanian.

Romanian. Yeah. It might be the only big Romanian unicorn.

Okay, yeah. I think there's a stage being set for the resurgence of RPA because everyone understands that no one wants to deal with these shitty apps, and no one's going to rewrite them. You just have to do a remote operation and programmatic operation of them.

My own setup was basically the following. I was doing a board deck recently—last month, whatever—and I thought, “Okay, let's just do this automatically.” All our data is in ClickHouse, PostHog, and QuickBooks, like everyone else's, and I basically connected that all to my Claude Code and said, “Here are the integrations. Go do that.”

It pulled out the first report, which was great. It connected to Brex and all these things and pulled everything out, which was great. I said, “Okay, now pull out this, this, and this.” I kept getting really McKinsey-style-designed reports, but the data said “partial data”—all the missing data, partial data. It couldn't access all the things.

I got so frustrated, and I got my Mac mini virtual sandbox with OpenClaw. I gave it its own account in our company, and then I went to all these services and created a read-only account. So it was literally like an intern in your company. I would say, “Now go and do this report,” and it would say, “I can't, via MCP or the API or whatever, get all the information.”

I told it, “Go log in,” and it would log in to the website, go in, export the data, and do the thing end to end. Even for things that have APIs today, not all of it is exposed. To get value—I get immense value right now—but it has to be computer use, unfortunately.

I spend a bunch of tokens just on that, but I get the job done. If even a startup like ours, using all the hottest tools, still needs a computer agent, what hope does Goldman have of having a headless agent, right?

Yeah, yeah. Why isn't Microsoft doing this? I'm pretty sure Satya had a post yesterday: “Every agent needs a computer.” I see, I see. So they have launched something.

Yeah, they have Microsoft Power Automate. I'm sure they're going to have their version of that.

And you're going to try to do yours. I always know there's demand for Mac, but I know it's tricky to host macOS sandboxes.

We will have macOS sandboxes fairly soon. The problem with macOS sandboxes is—I'm deep in this; I don't know how interesting this is—macOS has this problem.

It's a licensing thing.

Licensing thing. So, one, you're allowed to run only 2 parallel VMs per machine. Two, you can only license to a different user every 24 hours. If I theoretically want to charge you per second and I charge you for 1 second, I have to leave it idle for the rest of the day. I can't have anyone else doing that.

The pricing will be different in the sense that we would have to charge for 24 hours. That's not even the most difficult thing. The thing above that is, from a security perspective, they enable you to do memory snapshots, pause, and resume, but only on the same physical machine.

What you can do in the Windows or Linux world is move your snapshot in the background from one machine to another and manage load. If you want to do that, you essentially have to have your—

Yeah, snapshot your physical machine. You can't break it up. You can't move things around.

And all of that—that part—from a security standpoint, I understand the security aspect of that, but it disables you from doing these agentic, really scalable agentic workloads.

You need to do a vibe-coded clean-room implementation of macOS that you can then—it's like Clean OS or something. I don't know.

I guess so. I know because Linux was originally a clean-room rewrite of Unix, something like that, right? Same thing with macOS. Someone needs to do it. Someone will do that, and so we'll have some long-running agents for a few days to figure this stuff out.

But yeah, we're really close to offering something because people do want it, but the pricing will be different and the feature set will be sort of stringent.

Shawn Wang

Yeah, nobody's going to use this. I mean, the labs will, because they want to—

They have to. But the point is, with the RL part, if you do RL on macOS, then the next iteration of the model comes out and it will be able to use these tools significantly. Then you actually need to run those somewhere, so you're going to have to have that later on.

Shawn Wang

And if anyone at Apple is listening, I very much feel that they are shooting themselves in the foot in terms of the scale of the revenue from compute or licensing they could get if they would just enable a concurrency model similar to what you can get on Windows and Linux.

Yeah.

Shawn Wang

Yeah. I'm sure they've heard this before. They just don't care.

Yeah. And maybe they'll change their mind with the new CEO.

Shawn Wang

Yeah. We'll see. We'll see. High hopes.

High hopes.

Shawn Wang

Okay. But it's very clear that the market opportunity is huge in Windows, and you can go for a long time on just Windows. But your customers are going to want both.

Yeah.

Shawn Wang

It is interesting to me that this is the killer application for agents, right? How big was OpenClaw for you guys? Was there a significant bump, or—

Not for us. We're positioned differently. Although it's completely PLG and we have individual developers that use it, most of the users who use Daytona are sort of B2B2C.

It's either B2B or B2B2C. In the researcher world, it's B2B, so you're selling to labs and new labs and things like that. But on the long-running agents, from a scale-revenue perspective, it's mostly B2B2C, where you have an app-layer agent that uses you.

Shawn Wang

Yeah, yeah. Like a Manus-, Lovable-type—

Yeah. B2B2C is basically what I've been calling an agent lab. It's kind of like you're not a model lab, but you're making a very, very good wrapper that is a platform other people can sign up for, so they don't have to code those things.

Shawn Wang

Yeah, it sounds like a much better market than the direct OpenClaw market.

We've done multiple things. The Codeanywhere part of our career was very much an end-user developer product, and so that is great. You can get a lot of developer love, and I feel that we do, as a company, have a bunch of developer love. But it's a different type, whereas it's more akin to Twilio because you don't really run Twilio as a person.

I don't know how many people remember the “Ask Your Developer” billboards and whatnot. People really love Twilio, but they only used it inside of, like, “Oh, I'm building this app or service for a thing.” We're very much directionally aligned with that. You also know that I used to work for a competitor to Twilio, so it's kind of ingrained, I guess, in my—

Shawn Wang

People don't know Infobip is that big.

Yeah, it's like—

Shawn Wang

Because they're all American, they're like, “Whatever is in Europe doesn't matter to me,” but it's the same size or bigger.

No, no, it's about half the size.

Shawn Wang

Half the size, but it's still huge—multiple billions a year.

Exactly.

These are really interesting, large, revenue-generating, very sticky businesses. Whereas when your focus is the end developer, it is a very hard sell because they're very price-sensitive and very price-conscious, and it's very hard to scale. Your cap is the number of people who are willing to spin that up in the first place, and then spin up multiple of these.

Whereas if you're in the enterprise world, we know everyone's talking about how many tokens they're spending. A lot of companies today are like, “This is our company spend. Spend as much as you can.” Basically, that is where we're going.

If you think about that paradigm, where you're selling to companies that say, “Spend as much as you can to generate productivity,” versus, “I'm a single person. I have this much budget, and I'm doing this thing because it's fun or it's helping me out or whatever,” it is a different go-to-market strategy, I think.

Shawn Wang

Yeah, there's a lot of discussion. I'm just going through the mental list of things that are in your favor, which is, for example, MCP versus CLI. Obviously, you want CLI. It's been very good for you.

I feel like it's maybe a drop in the bucket, or maybe it's huge. I'm just checking whether these are big trends. I mean, those things work well in our favor, to your point, just because—

But they kind of jump in a bucket right now.

Yeah. I guess I think it's sort of all the things coming together. There are so many things that impact that, to your point. OpenClaw wasn't huge for us, but having the Agent SDK from Anthropic—Claude Code—was very interesting. The reason it was interesting is that a lot of, let's call them, app-layer agent companies—I don't know what to call them—essentially say, "Oh, I can create this new app, this new agent. All I need to do is use Claude Code and throw it into a sandbox, and then I have my interface to the human." That enabled so many more companies to actually offer this, and then they would pull on the sandbox.

That was interesting. And to your point, MCP versus the CLI: MCP is an interface against an API, whereas with the CLI, you can actually go do things. This is the difference between integrations and actually running scripts, data, or analysis against the thing. So being able to use a CLI very well enables the agent to do more things. Because of that, people will invoke a sandbox, run the CLI, and it'll do analysis on that data and then give you an actual result versus just pulling data from an API. It's a layer of indirection, basically. It's the same thing as agentic search versus RAG.

Shawn Wang

Just like—you just win whenever people put more agents into the workflow. So it doesn't really matter, but I'm just teasing out what else people have heard about that's sort of, "Oh, yeah, this is another sandbox use case. Oh, yeah, that's another one." Am I missing any big ones?

The thing that people talk about, which is the computer-use stuff, is probably the most interesting one. To your point, we've talked to so many people over the last year. It's like, "Why do you need a sandbox? Why do you need this? Why this?" And to your point, it's like, "Oh, I need a sandbox for this. I need a sandbox for that. I need it..." And so, "Oh, yeah, I need it for every single thing."

Basically, what I sound like a broken record saying is: You use a laptop every single day, right? You are one of one; it's just you. But now imagine how—and, by the way, the laptop, the computer PC market, the PC market, is about equal to the cloud market. So it's about $180 billion a year, something like that. Roughly, the 3 cloud hyperscalers are about equal to Apple, HP, Lenovo, whatever. Well, it's a little bit less, but it's sort of like that.

Shawn Wang

And now imagine—and that's just—how big is the addressable market? How many people are there in the world now? What's the latest estimate?

It's called 8 billion.

8 billion. And so let's say you can have 2 computers: one personal and one business, whatever. So it's double that.

Shawn Wang

And so that's 16 billion.

How many agents are going to be running in 2 years, 10 years, and 100 years? And for every single task, they will need one of these. Because that market is essentially, quote-unquote, infinite, you will get to the point—and Dylan Patel was at the conference talking about it. SemiAnalysis, which usually talks about GPUs, was also talking about how CPUs will now be a bottleneck, because they will be the constraint. You won't be able to grow, or we won't be able to have enough of these, because there won't be enough CPUs to basically do that.

Shawn Wang

Yeah. Well, I actually had a really good podcast with Doug O'Laughlin, who is president at SemiAnalysis, where they've basically been like, "Yeah, it's been a GPU shortage first, but then it's cascaded down to memory and now to CPUs."

And, I mean, what's next?

Shawn Wang

Sorry. Networking.

Yeah. Networking actually has been a shortage for a while if you're looking at just GPU networking. But, yeah, it's really crazy, the amount of computer use that's going on.

Shawn Wang

Yeah, cool. I guess the other question is—the one very big part is the open-sourceness, which you didn't have to do and your competitors don't do. I guess a lot of people are worried about keeping their projects open source because some competitor can just fork it. I don't know if you have any reflections on just being an open source company.

Yeah, there's a bunch. The original product that we did was open source.

Shawn Wang

Yeah.

Doing that was actually very good for us. There's basically a saying—what's the saying? Companies that are doing really well measure themselves against free cash flow. Companies that are kind of okay, it's EBITDA, and then it goes all the way down to GitHub stars. The original one was GitHub stars. That's what we talked about.

Shawn Wang

We're at the point where we talk about revenue.

So we've gone up the stack on that.

Shawn Wang

And so, profit?

Yeah, we haven't. We'll get there. But basically, at that point, we did stars on GitHub. What was useful in the original variation was that we split the core into its own repo, and it was Apache 2.0—very permissive—and then we basically bundled that on the enterprise side with a proprietary repo. So it was open core, but the repository was very clean.

When we did the pivot, we didn't have time to rethink this, and we had this open source community. It felt like a shame not to do that, but we still did want to add some restrictions. So in the new sandbox product, we added AGPLv3, which is a kind of shortcut way to do that, where you are open source and it is true open source in the sense that an enterprise can use it if it wants, but you essentially can't make a competitor without open-sourcing your stuff.

Shawn Wang

It's one of 3 approaches. There's BSL and some of the other sort of Elastic licenses.

Yeah, there are some others there. Pure open source believers agree that this is not full open source, and I totally respect that. That is absolutely true. But we did leave that, and Daytona, in its essence, everything outside of what's under a feature flag today—which is the Windows stuff, GPU stuff, and whatever it is—is open source. It is there. So everything is there, like our own scheduler; everything is there.

People have said, "You guys are actually open source—open source. You can actually see that." And people do like that, and it has helped a bit, but it's actually helped more in the consumption of our cloud product than in actually transferring people over. The reason is you can actually send the repository to your agent when you're integrating Daytona, and it just has more context: This is why this is happening.

Shawn Wang

You could equivalently just have docs that you can use. To be fair, it actually doesn't really help the growth significantly today. We've had this kind of conversation with investors and other people: How do you convert people from open source?

The open source business conversation is so all over the place, right? I'll just say, for listeners who maybe haven't thought this through: A lot of people say, "Oh, it's a free tier, right? If you run it yourself, but when you get serious, call us." And then, personally, because of my Temporal experience, it actually is the GTM into some of the largest companies where we wouldn't pass their review process, maybe because we're too young of a company or there are parts of the stack that just don't work with them. But because it's open source, they adopt it, and later we figure it out. That's the low end and the high end. I don't know if it—

No, no, no. Absolutely. That has been historically the thing that we have found in this AI transition. As we haven't talked about, Daytona's customers are everything from the single developer, the YC startup, to people—I'll say Fortune 500, Fortune 5—the biggest companies in the world.

Shawn Wang

And big labs—you told me about some. We'll keep them anonymous.

Ivan Goncharov

Enormous companies, right? And because the market pull is so strong, we're able to circumvent these processes. I'm not saying we don't—we pass security audits, we pass all these things—but as you mentioned, like Temporal way back in the day, in our old version of Daytona, it took us months, and usually at the end they would churn off because, just like, "Oh, you're too small of a company. We don't trust you enough."

Whereas today, we've had these large companies push us through. Usually, when you would go through procurement to become a vendor of large companies, it would take you 2 or 3 months. We could have done it in 5 days now. This is not saying that maybe we're great, but it's more, I think, a sign of where the market is today.

When you think about that, open source is something that we, from a go-to-market perspective, don't think about that much, because everything that we've created right now has been PLG through the cloud product: people signing up and just pulling us inwards.

Shawn Wang

This is a personal interest, and I don't know if you have an answer, but do you have problems with GitHub?

I do, a little bit. A little bit.

Shawn Wang

Yeah. Tell me, because I'm thinking about, okay, what would it take to replace GitHub?

So there's a lot of—I’ve thought about this, and I've tweeted about this, and I've looked at some. I've actually invested personally in some.

Shawn Wang

Is it Entire?

No. Okay, yeah. I've met Thomas virtually, and we've talked. I really think—and this was my reason for that—because we have a bunch of background, long-running agents. For a time, most of them were coding agents; everyone was building up a competitor to Lovable or Devin or whatnot.

What we saw from our customers was that they were all trying to figure out how to do versioning. Everyone was doing it in different ways, and there were some really weird ways people were doing that. The reason was that GitHub as-is was overhead: it wasn't fast enough for what they needed, and it didn't solve the problem they needed to solve. To be fair, GitHub is for after your inner loop, right? It's after your laptop.

Shawn Wang

Yeah. GitHub is the point at which the outer loop starts.

Ivan Goncharov

Exactly. People started using that for sandboxes, which is the inner loop, which is usually on your laptop, right? And so that is not what it's made for. We had everything from people—actually, the most interesting one is that we had one customer that would literally take the entire codebase inside the sandbox. I forgot what the time sequence was; they would just dump it all into a JSON and then push that to S3.

Shawn Wang

Yeah.

Ivan Goncharov

And that's it.

Shawn Wang

Make your own Git.

Ivan Goncharov

There's not even diffs. It's just the whole thing every single time, because it was super fast. Then they would go back and search and find what the file was, read and write and whatnot, because there's a text file there, JSON. They're very small, so the network cost is very low, and they didn't care; they just did it that way.

I'm like, if people are doing this, that means there needs to be a new solution to this problem, right? For me, it's quite interesting to look at who's building these types of new things, agent-first. I think Git as-is still exists in the future; maybe even GitHub exists, but there will be a whole new service.

Shawn Wang

Yeah, exactly. Git is like the deploy artifact to kick off CI/CD, but then there's a layer before that that's like the agent collaboration layer.

Ivan Goncharov

And so I think something's to be said there. But on the other side, another interesting thing is just CI right now. The amount of PRs being created is insane right now, right? In general.

Shawn Wang

Even for you guys, right?

Ivan Goncharov

Everyone's creating a bunch of PRs—everyone—and then all that has to go through CI, and then that's the bottleneck. Everyone's being bottlenecked, not just on Actions. Go to any CI provider; if you have a high throughput of PRs, you will not be able to keep up. There's one company we're talking to that does 1,000 PRs a day, which means they're just waiting; they have a queue.

Shawn Wang

What do they use, like Buildkite or—

I don't know what—

Shawn Wang

CircleCI? You know, technically, your tech can be used for CI.

That's the conversation.

Shawn Wang

Oh, okay.

Ivan Goncharov

That was the conversation.

Shawn Wang

Is that a serious conversation?

We'll see how that goes. We've had quite a few conversations around that. We're not a CI provider by any means, right?

Shawn Wang

But what is missing?

Ivan Goncharov

Essentially, you could use a Daytona Sandbox instead of whatever you use for your GitHub runners, essentially.

Shawn Wang

Yeah, yeah.

Ivan Goncharov

The only thing I would say is maybe CI machines are supposed to be very cheap. Maybe it's the low end because it's supposed to be nonblocking, or something like a background job. The urgency is not that important for CI performance, though.

Shawn Wang

Yeah. Performance, yeah.

Ivan Goncharov

Yeah.

Shawn Wang

Okay, that's interesting. Before we leave Daytona and go into broader founder takes and what have you, when startups evaluate you—you have all these names, and you have more that you can't even name. They see your wall of competitors.

Ivan Goncharov

Yeah.

Shawn Wang

And you have differentiation versus many of these, but what sells them?

The thing that we found that sells people the most—this is more of a day-two thing instead of a day-one thing—is responsiveness. We've seen this again and again. We have a bunch of case studies, and we have a bunch of them still coming out. They're all done by a third party, so we don't do the case studies.

It's actually interesting to watch those cases. I watch them—they're recorded—and because it's a third party, people are actually more open. They'll tell you, "Oh, we used this competitor," or "We like this competitor more," or whatever. The number one thing that people come back to us for is that we have insane responsiveness.

Shawn Wang

In terms of your team?

In terms of the team. Insane responsiveness has been, by far, the number one thing. We can talk about features, breadth of product, concurrency, CPUs, and all those things, but if all other things are equal, that is very much a differentiator, I found.

Shawn Wang

And is that entirely Slack, or Slack plus email?

There's email as well. There are calls, but the vast majority is on Slack. We've had customers say, "Hey, we have a problem. Can you get on a huddle?" We will get on that huddle in 5 minutes, literally. I've done this multiple times.

Shawn Wang

Wait. Okay. So how big are you?

25 today.

Shawn Wang

How do you do this kind of support?

We're insane. We don't sleep. 007. Have you heard the new thing?

Shawn Wang

07. I mean, I've met your team. They're very impressive and very dedicated. But how do you get a team to do that?

I have Slack exhaustion, you know.

Yeah, we all have Slack exhaustion. We're very, very tired. The thing that's unique—I don't know if it's unique about us, but I would say it's unique about any successful serial founder—is that you're able to pull in people that you've worked with before.

You can't do that as a first-time founder. I couldn't have done that. Of the 25 people in Daytona, I think about 13 of them we've worked with for 7 years or more.

You will if you ask any engineer, they're like, “You never sleep,” about me, and so then I do that as an example. I don't do an example; that's just how I'm wired. My wife doesn't appreciate that, I tell you.

The new segment that has come is almost everyone is sort of one degree of separation. It's like someone that someone has known, and so they sort of come into this org. We've had people that have not fit into the org as well. It's just that type of culture where there is a high expectation of being online and replying for these things, and I do that first.

Shawn Wang

My wife doesn't appreciate that. I told her about 996. She said, "I wish."

It's like these Chinese people are slacking.

Yeah, yeah. I think every company has its own culture, and that's something very deep for us. It's come up again and again, and every single day we're reminded about that. I didn't go out thinking that this is how I'm going to build it; it's just how I built these things.

I'll transition a little bit on the founder side. I'm very impressed by you in general, by your balance. You have a young family—

Two kids. Yeah, no, two kids now.

Yeah, two kids now. I think a lot of people I meet are like, "Well, I'm starting a family. I can't be a founder," and all that. What's your advice to those people?

My family is here right now, but I usually fly between Croatia and here a lot. A lot of our team is in Croatia. Part of our team, and a growing part, is here now in San Francisco. So I spend a lot of time away from my family, and that is hard. That's a sacrifice that you have to make.

Going in, people say that on your deathbed you're going to miss some of those things. That might be true, but going into this, I already said, "I know that this is going to hurt." Everything has to hurt. By the way, I'm very much of the feeling that everything has to hurt. Going to the gym hurts. Losing weight hurts. Everything has to hurt, right?

It is literally, but you actually have to enjoy the pain.

If you don't enjoy the pain, it's not for you. You get accustomed to that pain. I love the kids, especially. I have a daughter and a son; my daughter is the eldest. I love her and do miss her when she's not here, but that's what I signed up for. There is a plan and a target of what I'm trying to achieve.

Hopefully, with my wife, who does support me, we can get ourselves together more so it doesn't hurt as much. She takes a large portion of that. If you have a partner on the other side who is okay with that, then you can do that. But even if they are, you have to be okay with not being there, right?

Yeah. This is my vision for you. This meme.

Yeah, yeah, yeah.

So that's your kids in the future.

Yeah, yeah. I think so. Yeah. But we have to teach them that they're—

Not because Dad built the computers. Dad made sandboxes.

And built the spiritual successor to serverless and Kubernetes for agents.

Any other sort of hot topics or trends? You have a lot of hot takes. Actually, you are best known for—you were sort of in hustle-culture mode, right? Someone quoted you and said, “I haven’t even heard of you, bro. Just log off and take Christmas off,” and then your response was—

My response was like, “That’s why I can’t.”

Yeah. So, I mean, I think that’s very typical of you. I don’t have it here; I can’t bring it up, but I think that’s very typical of the culture. Any other sort of takes on the startup ecosystem?

Oh, the startup ecosystem. This was the recent one, which is—and this is general business—I feel that it didn’t come off well on Twitter. Some people always misread it, which is: the market is adding a premium to SaaS vendors that are reselling tokens.

Yes.

And I think that’s incorrect.

Why?

Why I think that’s incorrect is that, one, your pricing depends on what the price is, if it’s public market or private or whatever. You’re saying to the person reading that the reacceleration of revenue is equal to the old revenue, which it’s not—not even close—because, one, on SaaS you had typical SaaS margins, whatever it was, right? Say, stickiness and all these things.

Now what you’re doing is saying, “Here is my agent, and I have whatever the margin is,” and it’s way worse, right? Now you’re using Anthropic or, you know, OpenAI’s models, and we as a community are saying, “Now that is reacceleration.” One, I think that’s wrong, because first, it’s not the same; the makeup is not the same.

The other thing is—and go back to what I mentioned earlier, like the core and how I set up OpenClaw and whatever—I don’t want your agent, essentially. What happens right now is that we have a problem, and this has historically been the case: you have data siloed in, again, ClickHouse, QuickBooks; it’s all siloed. Now you’re giving me an agent that’ll give me the data, but it’s still siloed, right? So now I have to take that data and then get another agent—

Just expose the data. Just expose it.

And so I’m like, just expose everything and charge me for that. Charge me for consumption of the API. You’ll have your old seat-based pricing for humans—

Charge me for this.

The number of agents will skyrocket, and essentially you’ll have more usage and charge for more if your product has value. There are arguments; some of them do have value. It’s database, not databases—we can get into that—but some of them really do.

I was actually shocked that the first person to do this was Marc Benioff.

Salesforce, yeah. There was a tweet 3 days ago where he said every product in Salesforce has been exposed via API.

Everything. I’m like, now I understand why this person has built this. This is insane. Kudos to him. Amazing. It’s like, thank you for this world. I don’t know if you listen to me or someone else, but thank you for this world.

If you can get real reacceleration against that, against consumption of the API, that is actual revenue and that is actual reacceleration. That is where value will come from. I think there will be a cold shower when people understand that no one’s actually going to use and pay for these agents and tokens. That wasn’t actually real acceleration, but it’ll drop back down.

Yeah. I mean, look, obviously I think generally you’re correct, and I agree. But people are going to try to become an AI company.

No, no, absolutely. I have nothing against that, and this is not a downer on anyone who’s building this thing. Everyone has to get to the revenues, get to multiples, get valuations, and do what you have to do to get to the next step. Absolutely agree.

But we as a community are now saying, “Oh, this is the magical way to get out.” This is not like that. That is not what is happening, right? I think there was this kitchen-appliance company that put out some AI nonsense recently—

But it was also the sneaker. What was it called? Allbirds.

Allbirds is pivoting to GPUs. That’s fine. It’s like, I have some money left; I’m just going to do some lottery tickets. Would you go into offering GPUs?

Oh, yeah, we will.

But not for inference. What we think about is essentially the GPU sandbox. If you think of having a GPU in your computer, that is what you have—a GPU in the sandbox. There are workloads that do need GPUs. Again, I always go back to 3D rendering because it’s the easiest one to comprehend. But if you want to do any sort of RL on CAD or something like that, you will need a GPU in the sandbox. That’s coming now as well.

How about your own data centers?

Own data centers. We run on colocation providers and bare-metal machines. We technically can run on data centers or our own data center. That’s how we architected it today. From a gross-profit-margin perspective, it doesn’t make sense for us to get into that. You have to raise a large amount of capital and take on a large amount of risk for single-digit percentage points.

So today that doesn’t make sense, but we are fundamentally architected so that we can do that if we want.

Yeah. I mean, you’re a large customer of these guys now. Do you see any opportunity?

We will see.

Yeah.

We will see. Yeah.

Yeah. I see a lot of people trying to do the bare-metal thing. We talked to Railway the other day, and they’re also doing a very similar strategy. They think—I think—they’re building out something, or they have their own sort of data centers now.

Yeah, the majority of their own data centers. But I do think they still use Equinix and all those things.

So I think it’s just interesting that this model basically hasn’t changed. It’s basically a real-estate model. They manage the facilities, and then you do everything else.

I wonder how it can be changed for the future, because the AI wave is the opportunity to reinvent everything.

Yeah. Anything else? Cool. I think that’s about it. I didn’t have any other topics. I think this is as comprehensive as it gets. If you have any questions about the compute market and sandboxing in Daytona, this is the best place to start. Where does this go, man? We’re here in April. Things are going 75% month over month. Where are we going to be by the end of the year?

It’s an insane number. I’m sort of scared to say it out loud. It’s very big. Just the sandbox market alone—and we talked about this in general—the entire infrastructure market is growing 40%, plus or minus, month over month. Everyone is growing 40% month over month.

That’s also a hot take: if you’re not growing 40%-ish, it’s not that it’s just the market; you might as well not have to come to work. You’ll grow that amount, basically. I’m half kidding, but that’s where it’s going.

The thing I think about, at least from a CPU perspective—GPUs are even crazier—is that there’s a high probability that actually owning the CPUs beforehand will be a go-to-market tactic. It probably will, because you, as you do, probably talk to a lot of GPU providers; their growth is hindered by the amount of GPUs that they have right now, right?

It’s whatever NVIDIA decides to bless that day.

Yeah. That’s how much they’re going to grow, right? Whereas the CPU market in general—be it something like Railway, for example, or Vercel, or deployment, or the sandboxes—still has CPUs. Each is growing at the pace of its market and what their plus or minus of that market is, but it’s still not constrained by that.

My thought is that, for all of us in this market—and databases fall into that as well, because databases also run on CPUs—we all have to grow as fast as we can so we can get enough CPUs tomorrow from Intel or from NVIDIA, because they have CPUs now, and everyone else later on. So it’ll be interesting when we get to that.

Maybe one version I’ll phrase this is: are you the potential new Heroku, new AWS, or new Stripe? What’s the analogy that is most appropriate?

There are interesting analogies. There’s a new Cloudflare, but Cloudflare is Cloudflare. They’re actually doing a really good job.

And Cloudflare owns networking. No one can fight them, come on.

They’re doing—no, they’re doing really well. What I say is, in the sense of their whole agent portfolio, it’s actually really good. I should say there are some technical limitations, I think, around everything being constrained under Workers—Workers is their thing—but from a go-to-market vision perspective, I think they’re really, really good. I think they actually get it, unlike some other companies.

To your question, there will be an equivalent. Everyone says an AWS for AI agents, but it might look more like Stripe than AWS, in the sense that there will be a cloud built out specifically for agents.

That cloud will have sandboxes, web search, and databases like SQLite or Neon or whatever, specifically for agents, and other things. We are not at the end of the new infrastructure primitives for agents. There are more coming. So people think, “Oh, there’s nothing else to it.”

There are more. We have some ideas about the next ones. We don't have time to do them, but there are definitely more primitives being built out for AI agents, and there will be, I think, a cloud that runs all that.

Yeah, OpenAI has said AI cloud, Vercel has said AI cloud, and you are potentially also one of the other prospective AI clouds. I think it's a very big prize to win. Well, thanks for coming on.

Thank you for having me. It's been amazing.