Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO

正文

Warm welcome to Basil Chatha , our newest cohost from the Forward Deployed side of the AI Engineering world! We’ll be ramping up a new show with him and couldn’t pass up an interview with Modal’s CEO Eric on the occasion of their monster Series D!

Two years ago, we spoke with Erik Bernhardsson about his vision for truly serverless AI infrastructure: developers should be able to run code without provisioning clusters, predicting load, or paying for idle resources. CTO Akshat then joined us to talk about Modal’s infra primitives and stack to offer those primitives and agent experience.

A lot has changed in the past 5 months. Since their $5B series C, Modal now reports it crossed $1 billion in annualized run rate , raised a $750 million Series D at a $16B valuation , and has spun up nearly three billion sandboxes. In this conversation, Erik joins Vibhu to explain the infrastructure and economics behind that growth—and why demand for compute rarely looks like the contracts used to buy it.

Watch the full conversation on YouTube →

GPU demand isn’t rectangular

A typical capacity contract gives a customer a fixed number of GPUs for a fixed period. Plot capacity against time and you get a rectangle. Real workloads behave differently: they grow, spike, and fall throughout the day. Buy enough for the peak and you pay for idle hardware; buy less and you risk missing demand.

Erik describes Modal’s bet as a kind of duration transformation: pool demand across thousands of customers, secure capacity underneath, and let each customer use what they need when they need it. The smoother the aggregate demand, the more useful that pool can be. It’s a useful companion to our conversation with SF Compute’s Evan Conrad : Evan focuses on making capacity tradeable; Erik focuses on making it usable in smaller, elastic slices.

There is a financial risk on the other side. If a provider locks in capacity at today’s prices and GPU prices later fall, it can be left holding expensive supply. Erik says he doesn’t expect an imminent collapse, but thinks the possibility matters when building the business. Jump to the GPU price discussion →

The sandbox that almost didn’t make it

Modal launched sandboxes after Erik and cofounder Akshat Bubna noticed that language models could generate code—and that someone would need a place to run it. Early adoption was modest enough that the team discussed shutting the product down. Then coding agents, background agents, and reinforcement learning created new demand. Erik says sandboxes have grown to roughly a third of Modal’s revenue, a result he didn’t predict at launch.

That growth also changes what infrastructure has to do. An agent may create an environment, run code, inspect the result, and try again many times. The latency of starting and restoring those environments becomes part of the experience. In our recent interview with Akshat , we explored this shift from developer experience to agent experience; here Erik adds the product and business story. Jump to the sandbox discussion →

Fast startup is an infrastructure feature

Elasticity only works if capacity can arrive quickly enough to matter. Modal’s serverless GPU write-up describes the machinery: a buffer of available GPUs, a custom filesystem, process checkpointing, and GPU memory snapshots. Those pieces reduce the work needed to bring another model replica online. It is a concrete example of how infrastructure decisions show up in an application’s responsiveness and cost.

We also discuss Erik’s path from Spotify’s recommendation system to Modal, how the company buys compute, alternatives to NVIDIA, forward-deployed engineers, and why code review is optional at Modal. Jump to engineering culture →

Show notes

• Modal — the AI infrastructure platform Erik founded.
• Erik’s writing — his essays on infrastructure, markets, and building companies.
• Modal’s Series D announcement — the company’s latest reported growth and sandbox milestones.
• How Modal achieved truly serverless GPUs — the startup and scaling techniques behind elastic inference.
• Modal’s resource solver — how it selects capacity across changing prices and availability.
• Scaling to one million concurrent sandboxes — an engineering account of sandbox scaling.
• Our first interview with Erik — the original serverless infrastructure thesis.
• Our interview with Akshat Bubna — infrastructure and developer experience for agents.
• The Programmable Data Center — Evan Conrad on GPU trading and physical settlement.

Timestamps

• 0:00 What Modal does
• 1:25 From Spotify to building Modal
• 4:19 Founding before the generative AI boom
• 7:43 First customers and fundraising
• 10:00 Building for inference and autoscaling
• 13:21 Where Modal gets GPU capacity
• 15:39 Renting vs. owning infrastructure
• 19:45 The risk of GPU prices falling
• 24:18 Is AI demand infinite?
• 26:12 Alternatives to NVIDIA
• 28:25 What Modal can learn from neoclouds
• 30:26 GPUs as an asset class
• 31:55 Modal’s products: functions and sandboxes
• 36:03 The case for smaller training workloads
• 41:21 How agent sandboxes are changing
• 46:48 Anthropic and sandbox security
• 49:29 Using agents to test security
• 52:51 Developer experience in the age of agents
• 57:45 What Modal’s forward-deployed engineers do
• 59:55 Modal’s approach to product management
• 1:03:29 Optional code reviews and engineering culture
• 1:05:56 Cloud, on-prem and bring-your-own-cloud
• 1:09:25 Capacity, fundraising and growth
• 1:13:07 Scaling the team
• 1:17:28 Erik’s tweets: ephemeral sandboxes
• 1:18:31 The problem with rectangular GPU contracts
• 1:21:20 Inference providers vs. model builders
• 1:23:25 Where models improve next

Transcript

Automatically transcribed from the revised edit; names and wording may contain errors.

0:00 Thanks for coming on, Eric Fern-Hartson from Modal. Could you explain the business a little bit and maybe how you got here? Yeah. So we build, the easiest way to put it is we build AI infrastructure. So we offer a compute platform and a storage platform that enables a lot of customers like Dacagon, and Suno, and Runway, and Cognition, and many others to essentially build AI applications. So what that means in practice is a lot of inference, a lot of sandboxes, sometimes training and fine-tuning, sometimes batch jobs and other things. So we have basically built a software SDK that enables customers to move really fast and build this. We manage all the capacity underneath, make it easier to deploy and scale and run these things in production. So you were on Light and Space like two years ago, a lot has happened since then. How big were you back then, like whatever numbers you can share and comparatively, how big are you now? I actually don’t remember where we were, but it’s probably like at least 50X, there’s revenue and users and users. Yeah. I use it actually for a side project that I have. Nice. Yeah. It’s like, yeah, turn 2D images into 3D models. Yeah. So yeah, just hosting a couple of models here. What do you think? Any feedback? I mean, I love it. Nice. Good. I mean, yeah, that’s why I use it over any competitors or anything. Good. So before you started Modal, you were at Spotify and also like better.com, I think.

1:30 That’s right. Yeah. Yeah. What were you doing at Spotify? I actually did a lot of different things at Spotify, but my main contribution was I built a music recommendation system. The first version I joined after school in 2008, 2009 maybe, and spent seven years at Spotify. And most of that time was building basically from scratch and music recommendation system and built up a small team and we launched a bunch of features like related artists and Discover Weekly and a bunch of other stuff. So a lot of, a lot of matrix factorization, a lot of compute intensive stuff. And arguably like a lot of what sort of later became the genesis of Modal was just building AI applications and realizing like the infrastructure fucking sucks. Like, there’s got to be a better way to do that. So a lot of sort of, you know, the rationale for Modal was like kind of building a tool I always wanted to have throughout my years at Spotify. Yeah. So Modal was more of an offshoot of the stuff you learn at Spotify, not much from Better.com. Better.com was a little bit of a side tour. It was kind of a roller coaster, a long story about the company. We talked about it separately. But I was a CTO for five years of better and grew to a tech team from, I think, you know, two to three, four hundred. So I learned a lot about management and building teams and hiring and also learned a lot about developer productivity. I thought a lot about tools and things like that. Did a lot more like full stack stuff. But yeah, I kind of came back to my, was close to my heart, which is infrastructure and AI and machine learning after that. What was Better.com doing again? Mortgages. Okay. Yeah. How did you get into that?

3:00 I mean, we don’t have to spend too much time on that. Yeah, no, it’s fine. Like this guy, Zach Frankel, who I didn’t know at that time, ended up recruiting me to Better as a CTO and the company was 10 people. And, you know, for whatever reason, you know, the founder of Vishal, very interesting guy, very, very smart, a little crazy in mostly positive ways. We got very well along. So we, yeah, we started building it in 2015 and I spent five years growing that company and then, you know, eventually turned out that doing mortgages is pretty hard and, you know, interest rates went up very quickly after the pandemic and that cost a lot of issues at the company. But by then I had already left. That was like 2020, 2021 when you left? Yeah, something like that. Yeah. So, OK, you left before you had the idea for a model or you left and then you started? It’s hard to say. I had a lot of random ideas when I left Better and I, I actually wasn’t sure if I wanted to start a company. I was kind of just, you know, during the pandemic, I sort of almost took like a sabbatical at two kids. I mean, I still have kids, but I mean, at that time there were there were I think two and four and so I decided to hang out and took some time off and was kind of working like part-time, like hacking on certain things and wasn’t sure maybe I should do some consulting and I started building stuff. And I just got like super pumped about this idea that eventually became model. But actually both, you know, model was like, I think the 30 I built that I got, you know, that so there were like a few attempts before then that went nowhere.

4:30 But my model was like the first idea was like, this is kind of cool. Like I should do this. This is like, this could be a really cool platform. Were the other ideas like also in a similar vein like developer tools? The first thing was like some sort of web scraping thing, which went nowhere. But actually, you know, in hindsight, maybe could have gone somewhere. And was this before the chat GBT launch? Yeah. Yeah. 2021. Oh, OK. So chat GBT was late 2022, right? Then I started building, so maybe actually to rewind a little bit context, like when I worked at Spotify, I released an open source framework, Kaluigi, which is like workflow orchestration. So sort of the same categories, airflow and pre-effect index or sort of dated categories to some extent, like people don’t talk about as much. But for a while, it was like, you know, got a lot of attention. Anyway, I never turned it into business. But a couple of years later, when in 2021, when I was thinking about business ideas or things to work on, I kind of came back to workflow orchestration. And I started hacking on a bunch of ideas of, you know, running code in the, you know, like basic orchestrating code running in the cloud. And what I realized was like, you know, at the end of the day, like core puzzle pieces, not just like orchestrating the computer, it’s like executing the computer. And I was like, that’s kind of a bigger problem actually to solve. So in a way, I actually like, you know, left it on the shelf and then said, like, actually, I’m going to like go to this other thing over here and I’ll go back. I’ll get back to the workflow scheduling a little bit. But, but like, you know, I never got back to long story short and ended up building modal, which is like the focus is more like, let’s just figure out a way to make it easy to run code in the computer, in the cloud. And the focus was initially like fairly general purpose.

6:00 And it actually in a way like it’s always, we’ve kind of gone full circle more recently, but, but we focused on so many different applications. It’s kind of the idea was like, let’s just make it super easy to run code in the, in the cloud. Could be GPU code, AI, machine learning. It could also be like data stuff. And focused a lot on, you know, how do we scale it up? How do we make the developer experience awesome? Realized super early that like in order to do that, we’d have to go very deep in a stack and built on file system, built on container runtime and did a lot of, you know, solved a lot of hard engineering challenges. And then you mentioned chat, CPT, that was actually not quite the killer app. But like, if you remember, like that fall, another thing came out around the same time, which was stable diffusion and stable diffusion turned out to be a killer app for our, for what we’re building at that time. We were only four or five people at that time. So we started seeing a lot of people coming to us and wanting to run a stable diffusion model. And then we were real, that, that, that made us realize like maybe gen AI inference is actually a really good use case for us. And how are you finding it? I’m just like compulsive tweeter. I was even more like now, like I’m too distracted by building a company, but I always like tweeted a lot. I was blogged a lot. So I think I had a little bit of a following. And so, you know, in the early days, I like a lot of what drove the growth was just like me just compulsively just like tweeting about model and like what I was building and getting people excited and writing long blog posts about where the cloud is going and all these things. And I call it a bit like founder led marketing. I always like feel like you kind of have to lean into like what you’re good at

7:30 and like what I’m good at is like, I guess like just shit posting and being terminally online. And so, you know, like work, it’s still like, you know, I think a lot of our growth is we’re just kind of good at social media and just like being, you know, online a lot. Nice. How long until you had your first customer? It took a while. Also, did you raise like early on? We didn’t raise for a while. I actually self funded it for a while, even with up until the first few employees, which in a way was like kind of nice because it may be really like frugal about money and like wanting to see like ROI. I wasn’t taking out a salary for a while for a long time. Actually, at some point it became a problem because I needed the money, but we didn’t raise for a while. And until early 2022, we raised the seed rounds. So at that point I’d already been working on it for about a year. Akshah joins the CTO and co-founder in about five years ago, about now August, maybe 2021. But then it took until like basically like late 2022 for us to see any meaningful revenue. I guess that’s why I’m proud of you. Yeah. And stable diffusion then in our case was the main thing that’s sort of, so we went from like, you know, $100 a month in like, I think like November, 2022. And the next month we did like $2,000 a month. And then like January, 2023, we did like 30,000 all of a sudden. And, you know, so finally I felt this like massive, just like anxiety thing off my chest. I’m like, maybe this is actually like we’re onto something. It feels like we’re onto something. And then 2023 we grew very fast.

9:00 And I guess we’ve been growing very fast since then basically. That 30 K a month moment, was that just like a couple of big customers or was it like very diverse? It’s just one. And I don’t know if I ever talked about like who that was and if they even know it, but I think they suspect it. It was a company in New York called Mage. They’re doing like a stable diffusion company. So at some point, you know, in the first, I think first quarter of 2023, I think there were like 80, 90% of our revenue. But then after that, you know, it started diversifying a little bit. And late 2023, we got soon as a big customer. They were a big chunk of our revenue for a while too. But then, you know, it’s over time, like we actually have really good revenue diversification now. Yeah. For the stable diffusion stuff, I remember using it for like comfy UI. I’m assuming a lot of people were using comfy UI, like running comfy UI workflows on it. Some people were, but I think the people that were using model, they tended to be more like low level stuff. Like they did like complicated stuff. What they like with models, like it’s always been pretty hackable. So they deployed their own custom workflows with adapters and custom inference and a lot of optimizations and all the stuff we had done in order to get the developer experience to be really good, turn out to be really good for also running production workloads. Cause auto scaling turns out to be actually a very similar problem. Like you want to scale up containers up and down very fast. So we just happened to be like very well positioned to capture a lot of those workloads in the early days. Did tools for auto scaling like not exist? Like why did you have to build those yourself?

10:30 In general, like my thesis is like, you know, every new sort of era of like, you know, technology kind of runs up against limitations of tools. You know, whether that’s you look back, you know, like the web or the cloud or, you know, mobile or PCs or, you know, AI or whatever. So I think, you know, it’s often the case that tools exist, but, but they’re often not like particularly good. And often you kind of need to rethink a lot of those tools in order to reap the benefits of, you know, and really turbo charge and start building for this new world. And you saw this with cloud, you know, we have to invent things like Kubernetes and Docker and things like that. And in a similar way, you know, like with, with gen AI, you could build these things on the sort of prior generation of tools. Like you can, you can build these things on top of Kubernetes and Docker and in these things, but, but, but those technologies are built for a very different type of workload. They’re built for a very sort of steady work, steady back end workloads. They have an even, you know, steady number of requests, like running a pretty low utilization because CPUs are normally pretty cheap. And so people don’t really worry so much about cost. And, but now we’re entering this new world and you have, you know, very expensive GPUs you want to manage capacity and utilization. You have very large models that you need to load over network into CPU into GPU. And, and it turns out like a lot of these tools, like, you know, like they were just not built for that. So we felt that, you know, we had to rebuild a lot of that, you know, building our own, I mentioned the file system, but we also spent a lot of time over the last few years, like building, you know, some really cool stuff around like CPU

12:00 snapshotting, GPU snapshotting and audit, we built on autoscaler and scheduler and, and, and in all kinds of capacity management tools. And we kind of feel like you have to do that because like you can’t build, you know, for this new generation of software on top of, you can’t build like a wrapper on top of the old stuff in order to do the new stuff. You kind of need to start over a little bit from scratch. So I guess like the auto scaling, like did you, did you see any like weird behaviors on the platform? Like was there anything that like broke while you guys were scaling? I mean, capacity is always from time to time been, been a challenge. Like how do you get like, you know, hundreds and thousands and tens of thousands of GPUs? So that’s, that’s always challenging. We had to go multi region very early. So I think that was like kind of a fateful decision in hindsight, like doing that and kind of just biting the bullet and building it. Like we always thought of like our capacity management as like one big pull. Like it’s like one big pull over the entire world. Like we don’t think of it as like separate clusters running in each region. And so that was like one of the like very consequential decisions we made early to kind of think about it that way. So I don’t know if there’s like been like, you know, specific issues. I mean, we’ve obviously had bugs and, you know, outages and I hate, you know, as much as I, you know, would love for a model to be a hundred percent uptime, you know, and I think recently we’ve basically hit four or five nines, but throughout the history of model, of course, we’ve had, you know, some run up against scaling issues. Yeah. The GPU market and well, actually before that. So you guys use AWS like GCP, like all these cloud providers underneath.

13:30 Like do you also buy any infrastructure? No, let’s talk a bit about like cloud capacity. Because I think it’s actually changed a lot. Like in the first few years we did, we got almost all the capacity pretty much from, from the hyperscalers spot and on demand capacity, mostly spot capacity, running in all the different regions, you know, using initially AWS, but also GCP and overtime measure. We started renting some capacity from Oracle pretty early because we wanted the eight tens and eight 100s and we got really good rates on those. And over time we added a couple more clouds, like doing reservations, but, but like still up until like actually end of last year, like most of our capacity was either spot or one year reservations. Spot means you’re just like renting it right there and then. Exactly. Yeah. And it’s also preemptible, which we had to deal with. But, but it sort of works in Agria because you can sort of just, you know, aggregate capacity from many, you know, hundreds of regions basically. So that, that, that model worked really well until we started running up again. I mean, first of all, like modal started growing like very fast as we needed more capacity, but fundamental also like the cloud market started changing a lot. So in the last year, basically now in order to get the capacity we need, we had to work with a much larger set of clouds. We actually work with almost 25 clouds at this point. And in many cases, in order to get the capacity we want, like we, we have to go into, you know, three plus year deals and you have to pay a meaningful amount of money upfront. And you’re locking in pricing on a three year term, right?

15:00 So, you know, like, and, and, and so we have to spend the entire, you know, last six to nine months, like, you know, kind of rethinking how we get capacity. We have to build a whole procurement like desk, like we have like four guys working full time basically, all they’re doing is like procuring capacity, calling Neoclouds, you know, doing deals, vetting the capacity. We’ve had to build a whole engineering team, basically 10 people just like managing like bare metal capacity from all over the world and dealing with network issues and bad drives and GPUs falling off the bus and, and all these things. We don’t own capacity. We rent all the capacity so far, um, you know, on, on, on typically this three year deals. And that’s like a big decision. Like, should we buy, should, should we, you know, or should we keep renting? I’m actually like not like mega dogmatic about it. I think renting is nice because there’s a certain amount of capital efficiency to it. If you’re buying, you know, like, I mean, first of all, there’s not this like operational complexity of it. Like you have to, you know, deal with vendors and suppliers and, and, you know, power and, and, and cables and all kinds of stuff. But there’s also, you know, financial complexity to it. Like, how do you fund it? Like, should you raise equity, you know, and, and you can raise a billion dollars and go buy hardware for it. That’s a very dilutive way to do it. Um, you can raise debt, but you know, if you raise that, you know, you typically want to have like locked in long-term revenue to, to, to pay the interest rates. That’s also not trivial. But you know, like we look at the economics, we look at, uh, the operational

16:30 complexity, we look at the risks. Uh, it’s a complex trade-off and, and we’re open to, to, to buying. We’ll, we’ll sort of, there’s other constructs too. Like we’re, it’s to get the capacity we need for, for 2027. It’s going to take a lot of work on, on, you know, going a lot of different avenues to, to lock it up. So do you have to hire a lot of like finance people to do all this stuff or not? Really? You know, I’ve had to spend a lot of time on this and we’ve had to, to, to spend a lot of time on this. And I do think over time, I mean, there, there is a lot of, you know, these are very expensive assets, uh, the hardware. And, you know, I think, I mean, it’s kind of staggering. You starting to see like AI investments starting to become like several percentage points of, of, of GDP of the U S and, you know, starting to see like interest rates in, in America, shooting up in the last few weeks. And I mean, I like, at this stage, like, you know, and you go and look at the Neoclouds, like these are very, very sophisticated financial engineering companies. And so yeah, we, we look at that and, you know, we look at all the options and, and it’s possible we might have to do similar things ourselves. I wouldn’t want to like, I would rather focus on being, you know, like a very lean software company, like doing the things we’re good at. I don’t think it’s good in general for software companies to own hard assets. You know, in the same way as like, should you own your office? Like I don’t think that’s necessarily the right solution for most companies. But of course, like in order to get the capacity you want, like, you know, everything is on the table. Yeah. And I guess the supply is going to be constrained for a while.

18:00 So you kind of have to do that. Like, do you, do you agree with that? Or do you think this is just like a moment in time? I don’t necessarily agree with that. I mean, like Nvidia is delivering, you know, several hundred thousand black walls, you know, a month, right? And, and I don’t want to say like the market is like 100% efficient, but like, you know, in general, like if you’re willing to pay a high price, you can get GPU capacity, right? Like if you’re willing to pay more than everyone else, and, and of course, Anthropic and OpenAI is buying up as much as they can. And, but fundamentally, like there’s constantly supply coming on the market. And so I don’t necessarily think it’s like a fully cornered resource. Prices are going up. And I do think it’s worth kind of reflecting on that. It’s like the normal state of hardware is like, it’s like out of like gravity. Like eventually, hardware prices always go down. So we’re a little bit in like a weird sort of period of time where like hardware prices are going up. And, you know, it’s hard to say, you know, how long this is going to last, you know, could be another six months, could be another 36 months. Like it’s really hard to know how long. But I fundamentally don’t know if like it’s true to say, like, you know, you need to buy capacity in order to me, that’s a little bit to sort of almost like the great toilet paper shortage of the pandemic or whatever. Right. Like, you know, at least this like, you know, run on the bank type situations. And, but that being said, like, you know, capacity is going to be tight for a while. And, you know, prices are going up. And, you know, luckily, I think, you know, it’s not even necessarily bad for us. If prices go up, because, because, you know, scarcity also means like we can

19:30 charge more from our customers. And, you know, and the fact that we have a lot of reservations locked in at a rate means like, to some extent, we’re effectively like long GPU prices already. So I don’t necessarily worry too much about prices going up. I do worry more about actually prices collapsing. I don’t think that’s going to happen anytime soon. But I think that would be a quite bad thing for a lot of the players in the market. Why would that be bad? I think it would be bad for if, for anyone who’s locked in a lot of supply. Yeah. Because they paid way too much and they paid too much, you know, so there’s a risk of buying at the top, right? And again, like, I don’t worry about that much. Like, I don’t think it’s going to be like, you know, it’s going to happen anytime soon. But like, you know, like I said, like hardware prices in the very long run, they tend to go down. And so it’s, it’s worth, you know, reflecting on that, right? Right. Like right now, I think, I think your, your capacity is constrained and, and that leads to sort of, you know, in economic terms, there’s a scarcity rent, right? Like who captured the scarcity rent? I think the primarily goes to two places, mostly Nvidia, but of course, but also to some extent, Neoclouds, like Neoclouds and in general, like, you know, anyone owning capacity has better economics. Like that’s undeniably true today with capacity being constrained. Long term, you know, in the very long run, I don’t think that’s necessarily entirely true, right? Like there’s always going to be a little bit of a premium you can get from owning GPUs, but is it worth the risk? You know, is it worth the risk of like holding this like very expensive

21:00 assets on your balance sheet? I’m not so sure, but you know, we’re going to see, like it’s going to take a few years to get through this. And, and, and in the long run, you know, we might own some, we might rent some, maybe a combination, maybe do some other crazy stuff, joint mentors, who knows? Maybe some complex SPDs, whatever. Um, everything is on the table. How do you compete with open AI and enthropping and buying GPUs capacity? I’m assuming they’d want to like buy up literally everything that’s available. They are. And, and I think what you’re seeing is, um, you know, a lot of that, the, the price is going up. Is it a reflection of that? That being said, you know, NVIDIA is shipping a lot of stuff. I think also like to some extent, like NVIDIA is like kind of prefer a little bit of a diversification. Like they don’t want like, you know, small amount of large buyers in much, in the same way as you see, you know, hyperscalers getting a little bit of a constrained allocation. Uh, I think, you know, also NVIDIA prefers to see, you know, a lot of the neoclouds and a lot of alternative clouds, like getting more allocation. And I think a little bit, what we see in the market is also this is more speculation, but I think we see, you know, smaller clouds that are a little bit more sort of risky to work with being more of an opportunity for us. Like where, you know, maybe NVIDIA, sorry, enthropping and opening AI, my feeling is like they prefer to buy from like the, the, the try and through, uh, part of it is also because they had to do a lot of training runs. So they need to have like very large blocks, but there’s plenty of capacity out there, you know, you can get 32 nodes in Armenia.

22:30 You can get 128 nodes in Indonesia. Like, you know, like we love that kind of shit. Like, you know, so we can go out there and like get a lot of, you know, smaller capacity all over the world. Pretty good prices. Is it pretty much just like if someone offers you like a certain number of GPUs, like that’s all you care about. Is there nothing like, oh, you care about like how reliable it is, you know, is it just some like janky warehouse that someone has these in? I mean, sometimes it’s kind of a janky warehouse. It feels like, you know, but, but like, we do have like a spec sheet, basically. Like when we talk to vendors, the first thing we do is we send them a spec sheet. We say, we need this number of IOPS. We need, you know, like the CPU needs to be at least, at least, you know, have, have this many, you know, whatever the pass mark scores or whatever it is. We need to see like, you know, minimum, you know, North, North, South, bandwidth, egress and ingress. So we do have a list of requirements. Often we try to do like an acceptance testing period where we get like one node and, and runs and tests over time, you know, like to some extent, it also involves a little bit of a leap of faith. You know, sometimes we sign very big commitments, you know, 10s or 100s millions of dollars and we start working with the vendor that we still don’t fully know. And, you know, over time, like, you know, we’ll see, you know, and, and, and I think over, like, in a little bit of, like, there’s a little bit of sort of just, you know, our own sort of, like, you know, how like semi analysis has like a cluster max, I kind of feel like we almost like have our own, like modal max score internally. Like, I know which vendors I like more than others, you know, like, and, and,

24:00 and we sort of prefer to do additional deals with the vendors that we’ve seen, you know, very good reliability, not just that also, like it’s not just reliability, it’s also, are they very responsive when, when we have, you know, support issues, do the fixed issues we ask for to turn around time. If things go down, you know, how quickly do they fix it? Like things like that. Do you agree that the demand for intelligence is basically infinite? And if so, like, why wouldn’t you just buy up as much as you, as much capacity as possible? And then, like, if you over buy, you can just like sell it to somebody else. I think it’s very easy to say when there’s like a scarce resource, like, let’s just buy as much as you can. But I think, you know, first of all, like, if we bought it today, we couldn’t utilize all of it. Like we could use utilize a lot of it, right? But like, But you can sell it off to somebody else who wants to use it, right? I think the challenge is like, if there was like something in the market that happens, there are a lot of correlations that happen at the same time. If, you know, prices going down and utilization going down and the market for reselling capacity goes, vanishes at the kind of the same time. I’m sitting here and like, I’m extremely aggressive, like buying capacity every day and like raising money to buy, you know, so overall, I’m, I’m very bullish, but like, I do want to point out and like, you know, like there are tail risks here. And when those tail risks happen, like it could be get very bad. Part of it is like, maybe I’m, you know, I’m just old and I’ve seen this in the past, like I actually had my first job in 1999. So I saw the dot com crisis and I’ve seen the great financial crisis and things like that. And, you know, like, I want to build a company that, that is around for like 20

25:30 to 30, 50 more years, right? Like, and so, you know, I think we can certainly go out and like buying an insane amount of capacity right now and trying to just corner the market. But, you know, do you know about the Hunt Brothers? It’s like a legendary story in trading. They try to corner the silver market. It fails spectacularly. You know, so it’s a cornering the market. I think in general, like it can be not, not a good idea. And, but, but we are in general very, very aggressive, but we’re ramping up, you know, like at an aggressive plan that basically matches our demand growth and matches where we want to try to get to at the end of next year in terms of run runway or run rate. Are you only buying up NVIDIA stuff or also like, you know, maybe a capacity from other types of companies like training them and we’re looking at it. I think the alternative accelerators are starting to get very interesting. It’s, you know, ironically, like every, every year in the last five, maybe even 10 years I’ve been saying, you know, like next year, it’s going to be the year of like non NVIDIA accelerators. And then I’m always wrong. But, but I do think 2027 might be the year of non NVIDIA accelerators. Finally, so we’ll see if I’m wrong this year too. But so I, you know, I am these are starting to get like really good scores on certain things. I think, you know, TP use are obviously used very, you know, widely internally at Google. I seem to work really well. I don’t know much about training. I don’t feel like I see quite as, but, you know, the team is very good. I know they’re working on a lot of the software stuff.

27:00 You know, hardware is hard in a certain way. Like software is like easier. And that makes me feel like, you know, if we just fix the sort of software layer for these accelerators, like, I think there’s going to be a lot of demand for them. And I think fundamentally, you know, with NVIDIA being such a bottleneck, like, we’ll probably see, you know, more alternative accelerators. That being said, you know, there’s bottlenecks further down the pipe. I mean, you know, there’s like TSMC and ASML, maybe like a deeper bottleneck to all of this anyway. But, you know, long term, I think monopolies don’t always last. And so I think you’d be healthy for the market to have more alternative. As much as I love NVIDIA, like, to make that clear, like, I mean, they’re obviously an amazing company and within the amazing product. I do expect to see a little bit more diversity in the market as well. Also, things, you know, like Cerebras and Grock and these things. Yeah. What are your thoughts on like some of those providers? I think they’re pretty exciting. And, you know, and you see them doing really well on certain workloads. And, you know, there’s some interesting, you know, people doing these like hybrid, you know, disaggregated inference where they run, you know, NVIDIA for the pre-fill and Cerebras for the decode and things like that. Oh, I haven’t heard of that actually. Yeah. And, you know, Grock too, like, because a lot of these things, you know, certain things are compute bound, certain things are memory bound. So you can sort of mix and match. And I think that’s quite exciting. It starts to get very complex when you’re deploying these things in practice. But I do think there’s a lot of, there’s a lot of, you know, opportunity and these sort of, you know, more complex hardware setups. Are there any learnings that you get from some of these Neoclouds, like, I don’t know, like Nebius or whatever, and like the way that they’re doing things versus

28:30 like what you’re doing? I think one of the things that, you know, is very interesting with Neoclouds is like, how can they expand so quickly? I think actually the most crazy example is like N scale, which went from like nowhere to like, you know, having like, you know, whatever, hundreds of billions of dollars in the pipeline, despite barely having any capacity. How can Neoclouds grow so fast? I mean, there’s like an, you know, and basically the answer is like, you know, and this might be obvious, but like what they do is like, they go out to like, you know, enthropic or open the eyes and some big buyer or whatever. And basically say like, Hey, do you want to buy like five years of chips? You know, sign this contract, we give you a good rate. And then they take that contract and then they go to, you know, Wall Street and say, look, we got this like, you know, big buyer, you know, they committed to $10 billion worth of chips. We just need to monitor it right now. And then they go on to raise the money. And then they like, you know, go to like Super Micro and DAL or whatever in the video and like get the hardware and then they install it. And then, you know, now they have like, you know, delivered to the customer. Right. So you have this like very nice sort of financial construct where you can basically sort of pre-sell things that you haven’t even bought. And that means, you know, you can scale up extremely fast. And so that’s what we’ve seen with these Neoclouds is, is, is they can scale very, very well. I think modal because we’ve always had a little bit more of a PLG focus and, and not the sort of, you know, ability to lock in long-term contracts in the same way, it’s a little bit more challenging from a financial perspective to scale that fast, even though I think, you know, in the long run, like we’ll get to the

30:00 same point. But I think that’s been kind of a fascinating thing to see with these Neoclouds and especially like CoreWeave and those big ones. They’re very, very savvy with, you know, how they raise money and how to, you know, think about GPU as collateral for loans and all these things. What’s the relevance of, like, I forget who it was, but like some like financial company was like, oh, we’re going to make, oh, I think it was Nvidia. Actually, they were like, we are going to turn GPUs into an asset class. Is that just because like they need to raise money and they’re like trying to figure out where to get the money from? Like, what’s the relevance of that? I mean, there’s a bunch of startups. I don’t know if you were thinking about that. There’s like Silicon Data and computer exchange and like Orn and, and, and a few other ones that are like, basically like trying to get to like, you know, creating price indices and indexes and trading derivatives and things like that. I think it’s interesting. Like, I think with compute, like it does, there are a number of challenges with compute as an asset class. Like it’s definitely not as homogenous as other asset classes. You have different network and different locations and, and, and different disks and CPUs and all kinds of stuff and reliability. I mean, that being said, like you can say the same thing about oil and like we kind of created like, you know, relatively commoditized, you know, tradeable assets out of oil. It remains to be seen. Like I do think, you know, at the scale where we are with the sort of AI, like data center and GPU deployments, like we do need to be a little bit smarter with like risk hedging and, and, and how we share risks and, and, and, you know,

31:30 having more like sort of insurance and, and things like that. So I, I, I’m very happy that people are thinking a lot about this. That being said, like I’m, I’m not entirely sure if we’re going to see, you know, GPUs becoming sort of, you know, whatever it is, sort of Henry hub, like, you know, Brent crude type, like, you know, super commoditized, like tradeable asset. Yeah. I want to shift the conversation to modal sandboxes. Yeah. I guess like before we do that, what are the different products that modal has? It’s, I guess like the main product obviously is like the, the GPUs. Then you have modal sandboxes. Is there anything else? I think of the like the two big products that we have as basically number one is modal functions and number two is modal sandboxes. So, so with modal functions, we launched that, you know, back in 2021. It was the idea of like serverless functions that are very easy to deploy and very easy to scale and we managed all the capacity. That turned out to be really, really good for inference and it turned out to be really good for managing GPUs and running on things like that. But we actually never had like, you know, dedicated like inference product or like a GPU product per se. Like we never thought of modal as like a neocloud and I think there’s like a risk of like getting sucked in that direction and becoming a neocloud, like just selling, you know, sort of what I think was like rectangular, like here’s like a three year. You get this many GPUs. Models has always been very good at like you deploy something you can kind of scale it up. You have this like very cyclical or like bursty stuff. Like we manage all the capacity. And the beautiful thing is like, because we have many, many thousands of customers,

33:00 we can sort of aggregate that capacity and get very smooth workloads, you know, very smooth aggregate demand, which we can manage very, very efficiently. So modal functions was the core. We launched that in 2021. Like I mentioned, we started seeing product market fit in 2023. And now, you know, it’s about two thirds of our avenue and empowering a lot of, you know, in practice, it’s mostly inference, but you can use it for all kinds of other stuff. Modal sandboxes, we launched that in 2023. And it’s actually kind of, you know, took a while to see product market fit similar to functions. Why did we launch it? We realized a lot of people we talked to were using LLM to output code. This is like me and Akshar talking over lunch one day. We’re like, maybe they need a place to run that code. And we realized like, actually, we kind of have like a lot of the infrastructure already to just take code and run it. Because we’ve already built the primitives for like managing containers and isolation. We already had, you know, G visors and isolation primitive and things like that. So we get kind of excited about that idea. And Akshar wrote a prototype and launched it in basically one day or two days or something like that. Zero. Like we saw like very little like growth. I don’t want to like trivialize it. Because we did see some growth. Like there were like some sort of, you know, And it was 2023. So yeah, it was 2023. And like, you know, like some people picked it up. Like we had some early users who kind of enjoyed it for certain things. But like overall, like it was like a pretty insignificant source of revenue into going into the summer of 2023. So two years later, we even discussed like, do we even need this product?

34:30 Should we shut it down? Then it just started exploding. So 2023. And I think what really happened 2023 was like, LMS got really good at generating code. In 2023. 2023. I mean, it was like around that time. Like I feel like that was like when they went from like being able to like refactor functions and like write snippets, so like generating like full apps. Right. And so we started seeing a lot of growth coming from like vibe coding, you know, companies like Loveable and other ones. And then, you know, companies like ramps started using model for background agents. And then we started seeing tremendous growth coming from a number of Neolabs doing things like reinforcement learning on model sandboxes. So sandboxes is now about a third of our revenue. You know, we considered it like basically like our second product. And to me, it’s like really exciting because like even back in the early days of model, I always knew that model, you know, is about like creating a whole toolkit of products. Like we always wanted to be almost like a cloud, you know, you have the storage and the compute and a bunch of different ways to work with it. We always wanted to be a multi-product vendor, but that wasn’t necessarily true in practice until like now. We’re like, I think now it’s like kind of clear. We have functions, we have sandboxes. We’re working a bunch of other stuff too. We’re working on training, working on batch, we’re working a bunch of other stuff. So to me, that sort of evolution into becoming like a multi-product company, I think it’s been amazing because that was always so core to the vision. Even in the very early decks, I had a bunch of slides on this. But it wasn’t necessarily true to the external world until recently.

36:00 Do you have anything in the works that you think is going to be like that next product? Any predictions maybe about the market? Anything like that? I mean, I’m very bullish on training. I think training is like kind of an underserved market. Obviously, there’s like a lot of focus on like very large training runs. But we actually see a lot of demand, the sort of smaller and mid-sized training segment. So I’ll be talking a little bit about that later this week. We’re organizing a conference. Batch is I think like a really interesting market. I think we see a lot of demand for sort of, you know, customers that need a... You know, we had a customer the other day that needed to transcribe, you know, 5,000 years of audio. They had like, you know, millions and millions of audio files. And we had a lot of capacity. So we let them do that, you know, and took a few days. I think there’s some really interesting use cases there. Very bullish on like low latency stuff. I think there’s going to be a lot of, you know, once we sort of get the audio, like the speech-to-speech models working, like the voice models. I sort of expect that to be very interesting market. Then of course, like all this like agente stuff, right? And in a way, like I have like mixed feelings about like the word agente, because I feel like in certain ways, like, I think all software will become agente. So it’s almost like to me, like, what does it mean to be agente? Like I think all software in a certain way, like will become like AI powered over time. But like clearly there’s like a set of like building blocks of like, you know, doing more sort of LLM powered software that seems to be sort of, you know, common across all the different use cases.

37:30 And we’re like, you know, the way I think about models, like we’re in the business of building building blocks, you know, we want to give you like amazing, generalizable, battle-tested, you know, general purpose, composable building blocks. And I think there’s some really interesting stuff also, like, you know, with the agent use cases. That’s interesting. You bring up training because I think in another podcast, maybe like a year ago, maybe like two years ago, you said training you were not focused on like specifically, like what changed in the market? I mean, we’ve been building a training product for multiple years on and off. So I wouldn’t say we’re like, we’ve never been focused on it. But it’s like, it’s like one of those things we’re like, I mean, now we’re like 150 engineers, but, you know, a year or two ago, we were a much smaller number of engineers. And so, but just virtue of, you know, being smaller, we had to focus a little bit more. And so inference became like our sole focus and was our sole focus for many years. But we’re bigger now. Like we, we, we do have opportunities to also place bets in other markets. I’m also really interested in, you know, kind of continuing the list of things. Post training, reinforcement learning. I think there’s some really exciting stuff around that. I think there’s also many other like stuff around like storage and building data warehousing and feature stores and real time streaming pipelines and all kinds of like, I don’t know. I think the longterm view of, of modal is like, basically like, you want to build some cool AI shit, like, you know, we got you covered, like, you know, we got all the building blocks, like comp tasks and like, we’ll help you, you know, take all these building blocks and compose them.

39:00 And that’s the beauty of software is like you, you can, you know, build almost anything if you have the right building blocks. So that, that’s very much the sort of company and the product that want to build. When you say training, are you focused on more of like the bigger enterprises doing huge training runs? No, no, they’re going to go and like, you know, do these like, you know, big, hyper-scaler or rather Neo, Neo cloud mega deals and get, you know, 100,000 GPUs or whatever. But, but, but there’s like a little bit of an underserved market of like, in my opinion, you know, like researchers iterating or like people training smaller models, there’s, you know, Texas beach and speeches, text models, computational biotech is actually a market that we see a lot of demand from. We have a lot of customers in that space training relatively, you know, modest models. You can train on 64 GPUs or whatever recommendation system, a lot of sort of traditional machine learning. There’s a lot of things like you don’t need a, you know, 100,000 GPUs. That market, I think is like, we’re not going to like focus on that for a while. But, but I think there’s a lot of stuff in the, you know, slightly smaller markets that that are relatively underserved. It’s very hard to set up training clusters. Yeah. Have you noticed like since, I don’t know, like GLM 5.2 or like, maybe like, can we 2.5 that, you know, a lot of people are using those and like fine tuning them? If anything, I’ve actually seen the other direction. I feel like fine tuning the value prop a couple of years ago was like, yeah, you know, open source models, they’re not quite as good. But, you know, if you take, you know, a small model and you fine tune it on your specific

40:30 domain, you can actually reach the frontier in terms of, you know, quality and cost. Today, I would say like the open source models are so good that like you can basically just like swap them out and just run, you know, instead of, you know, the frontier ones, you can run, you know, Kimi K3 or whatever, DeepSeq 4.1 or GLM 5.3. And they’re like basically almost as good as like the frontier model at a much lower cost. So we still see a lot of demand of for post-training. And I do think it’s a really interesting space and we’re working on some products in that space, you know, both sort of SFT and but also reinforcement learning and things like that. I would also say, you know, fine tuning is no longer the thing that you kind of have to do. Like it’s very optional. Like you can open source models like work really well just off the shelf too. Back to sandboxes actually. I forgot to ask you. So like what did you have to update about the sandboxes that you guys were offering that like existing sandboxes did not did not allow for? Not a whole lot. I think the thing we’ve always been very, very good at, I mean, we’re infrageeks, right? Like we’re very good at like, you know, scaling and reliability and doing a lot of stuff. And I think that’s why that explains to a large extent like why we have, you know, in my opinion, I don’t have numbers, but like I think we have by far the largest market share in this space. Like we’re just very good at running, you know, hundreds of thousands or even millions of sandboxes. Yeah. I will say we’ll have to, we’ve had to rewrite our systems, you know, a bunch of times in order to handle that scale.

42:00 But we were always like very good at it. There are other things like over time. I mean, the sandbox space in general, like, you know, like it’s like not the most like differentiated space, but you know, we’re working on some really cool stuff. Like over time, like we’ve had to add support for VMs, like virtual machines. So we previously bought our sandbox product on GVisor, but now we offer full virtualized machines as well. Turns out a lot of for certain use cases, you kind of need full Linux compatibility. There are other things we’ve had to build, you know, better primitives for like more interactive, you know, piping data back and forth between sandboxes, working with file system. We’ve had to spend a lot of time on snapshotting. I think a lot of the things that over in the long run will differentiate will be like, you know, really cool things like, you know, forking sandboxes, snapshotting, resuming snaps, like live migrations. In general, like I think sandboxes, the novelty of sandboxes is always, you know, this sort of ephemerality, like you can spin off sandboxes for a few seconds. And that’s something I like, that’s a string I want to pull more in. I think there’s a lot of really cool stuff you can do, you know, kind of decoupling the agent and the harness decoupling it from sandboxes, having agents control many sandboxes, forking sandboxes, you know, snapshotting things like, you know, like running basically like very short lived, maybe one per command, like things like that. So I think there’s a lot of cool stuff there. We’ve already spent a lot of time doing that, you know, around snapshots. But I think there’s a lot of cool stuff to be built there as well. You brought up snapshotting a couple of times.

43:30 Like what do you see as like the use case for snapshotting? Like what would that give us that we don’t get right now? I think the big thing for me is like today a lot of sandbox workloads are kind of like lift and shift. Like people think a lot about sandboxes almost like, you know, I have my dev box, now I’m going to shove it in the cloud, I’m going to have an agent control it, maybe I’ll run the harness inside the sandbox. I don’t know if like long term that’s like the right way to think about, you know, on like compute for agents. I think a much better model would be you’re going to have this like very long running agents, which almost like to me are like kind of durable execution workflows. Like they’re going to run, you know, 24 seven, you know, every once in a while they’re going to run commands and, and you know, having the ability to like spin up a sandbox in, you know, 100 millisecond, run a few commands and then like shut it down and like snapshot it. I think the faster you can get that like, you know, wake up, you know, go to sleep cycle, the more you can sort of create sandboxes that only run for like a very brief period and that will make compute much more efficient. And it will also enable you to do some really cool things, you know, like forking and moving things around and, and, and, you know, having like imagine, you know, like Claude code instead of, you know, having sort of a very linear flow, like sequential flow, like kind of branching and, you know, running hundreds of agents at the same time. I think a lot of like CI is like another like illustrative example, like for most of the, the history of CI, you know, talk about like GitHub actions and things like that.

45:00 Like you had a very sequential flow, like, you know, you slowly install a bunch of stuff, then you run the test suite and then, you know, generates some reports or whatever. And it’s like, you can break it up into shards, but it’s kind of annoying. Like what I would love to do is like take your entire test suite and just run every single test in its own sandbox, right? Like, and just paralyze the shit out of that around like a thousand sandboxes. Why can’t that be done? I mean, the problem is like there’s too much, there’s too much overhead around like, you know, starting and stopping sandboxes. So I think the more like you create like, you know, frictionless, like starting and stopping, the more you can enter a world where like you can do this like extremely bursty, like large scale, like run as much as you think you want in parallel. So, you know, I’m an infrageek, like I said, like I get very excited about these things. You know, once we figure out all these primitives and the right building blocks, I think it’s going to unlock a lot of really cool stuff that, you know, we previously could not do with software. Why hasn’t that already been built? I would assume even previously people would want to parallelize as much of their CI as possible. You can’t, but the overhead, it’s like Amdahl’s law or whatever. Right? Like the overhead of parallelization right now is very high because like right now, like you can get up actions, for instance, like we parallelize in the modal test suite, right? But like what that basically means is like now if you take your test suite and you split it up in two, like now each of those shards both need to do all the installs and the test setup. And that takes a long time. Like it can take minutes, but let’s say you can do the test set up once

46:30 and then you snapshot the container and now you clone it and now you can run the test suites independently. I think a lot of the benefits cannot be, you can’t materialize those benefits until you get the friction and the overhead down, the transaction cost. I think you guys have a partnership with Anthropic and they’re using sandboxes for something, like modal sandboxes. Yeah, I mean the managed agents with Anthropic, this is sort of bring your own sandbox integration. So you can use modal for those, yeah. How did that come out? Like come to be? I think Anthropic is like trying to create like a thriving ecosystem and community and so I think they make money from inference, right? Okay. I think it’s very much in their interest to promote, for everything else, as much of an ecosystem as possible and they want to have the largest possible surface area as long as you use Anthropic for inference. So I think it makes like very rational sense for them to partner with us as a sandbox provider. Also, I feel like we have to bring up the hugging face incident. Are you okay talking about that? Yeah, yeah, I mean it’s all there. Yeah, yeah. So yeah, could you just briefly describe like what happened and I think they were using the modal sandbox but it wasn’t your, you know, it wasn’t like a fault of the sandbox. Yeah, exactly. I forget all the details of the exploit but I mean, you know, I think what you’re referring to is the fact that like modal was involved. Yeah. And so, you know, just to comment on that, like basically the agents running inside,