Modal CEO が語る推論の未来、GPU市場の混沌、サンドボックス賭けの原点

How much of like an organized marketplaces compute these days versus like the wild west? I went to GTC last year. It was like an open-air drug market, but like GPUs. Walking around like talking to people, trying to get GPUs and meeting people in hotel lobbies, exchanging numbers. It's very far from an efficient market. I mean, some people are working on that. Compute comes in many different shapes or forms. I also feel like structurally, it will be a little bit like sort of unwieldy for a long time. Like, it's just not something you can like sort of buy and sell in an exchange because it's not like a fungible to want. Welcome to Unsupervised Learning. As you can probably tell from my voice, I am not Jacob. I am Patrick Chase, GP at Redpoint. So awesome to have Eric on Unsupervised Learning for the second time. He is the founder and CEO at Modal. I have learned so much from working with him over the last three years. He is one of my favorite people to jam on AI ideas with. He has an incredible ability to move seamlessly from the lowest altitude of developer shipping code

to the highest altitude of where the market is going and building what is missing. We talk about the mobile journey over the last three years, the crazy technology they have built to make it all work, inference, sandboxes, and what's coming next. Hope you enjoy the episode. For you guys, there's obviously a ton of AI compute demand and lots of people going out and trying to reserve, you know, GPUs and CPUs. What's it like for you guys finding compute in the market today? It's pretty rough. I mean, like, you know, not rough in the sense that, like, we can actually find compute. I think there's a little bit of a misconception that, like, we're sort of, we are certainly scarce, you know, in terms of compute in certain ways, but, like, there is capacity you can get. You just have to look a little bit further, accept higher prices, accept longer terms, et cetera. and we do that a lot. I mean, we probably do like two compute deals every week at this point, if not three. And so we keep, you know, stacking up, you know, more and more compute. And I really think that's the beauty of modal

is we can aggregate all this demand and manage the capacity in aggregate much more efficiently. And the value problem of modal has always been like, don't think of my capacity. Like we handle all of that, like annoying stuff that's sort of the day-to-day knife fighting of getting capacity and working. We work with 25 different clouds at this point and adding more all the time. How much of like an organized marketplaces can be these days versus like the Wild West, like random people that somehow have capacity and, you know, late night text messages? It's a lot of late night WhatsApp groups and like it's pretty unorganized. Like, I mean, I went to GTC last year, or so earlier this year in San Jose, the NVIDIA big conference we were doing. And it was like an open air drug market, but like GPUs, like you're just like walking around, like talking to people, trying to get GPUs. and, you know, meeting people in hotel lobbies and, you know, exchanging numbers. And, you know, I mean, at some point we, I think we got connected with someone who had GPUs because I think our VP of sales cousin was golfing with someone and we got connected.

There's many such cases. Like, it's very far from an efficient market. I mean, some people are working on this and I think over time they get a little bit more efficient. But it's very heterogeneous. Like, compute comes in many sort of different shapes or forms. So I also feel like structurally, like it will not, like it will be a little bit like sort of unwieldy for a long time. Like it's just not something you can like sort of buy and sell in an exchange because it's not like a fungible contract. Yeah, but the key difference is being like, I guess, the kind of actual powered shell or size of a cluster and like the chips themselves. To your point, there's a lot of people have been trying to build like, hey, you know, compute as this fungible single asset that you could build a marketplace around. What are some of the kind of the key differences you observe? I think a lot of it actually boils down to reliability I mean, first of all, it's like, you know differences in like disk and networking and of course CPU and memory and things like that like all these things vary and region, like location like they're all over the world like we have GPUs in the US of course, like all over the US

but we also have GPUs in like Indonesia like all over Europe, etc. So location is one of those big factors but also reliability. And that's another thing we do exceptionally well, is we aggregate all this capacity, and then when NeoClouds die, or when they have failures, which they often have, because some of these are former crypto miners who haven't really thought about running a 24-7 operation, we can just take it out of rotation. Or when GPUs fall off the bus or whatever, in many cases, we just file a ticket automatically and someone goes and looks at it. Have you found the data providers, the companies that provide pricing for GPU and that sort of thing, Have you found them to be helpful? I think moderately helpful. I think the pricing less so than just their ability to connect with a lot of providers. I don't fully trust the pricing today, but I think some of these providers are actually very good at, like, you know, knowing all the NeoClouds, and we work with them sometimes to get connections with cloud providers. The pricing, I think, over time will get more accurate.

Like, our internal pricing does not match what we see publicly when we buy capacity. But I think over time, hopefully, like, we will have more efficient, like, external analysis. I think that would actually be very useful. You've been on this, like, rocket ship journey these past years. And so lots of things we'd love to dig into for our listeners on the business side. Maybe starting, you know, I think one of the interesting things about the history of modal is, you know, for a while it felt like you, you know, publicly said, I think on your first unsupervised learning episode, that you were less focused, you know, on these LLM inference calls, specifically like a simple API call for a model. And you kind of recently released a product there. Maybe talk a little bit about the journey and kind of that initial reticence and what made you decide that now is the right time to really double down on this. Yeah, so we started building this in late 2021, basically, or early, somewhere around that time. And the initial product was this serverless function. And we thought it would be really useful for a wide range of stuff around data, AI, machine learning, broadly speaking.

Kind of looked at a lot of different use cases. But then ChatGPT came late 2022, and Stable Diffusion also came around at the same time. It turned out we were super well-positioned to host a number of Stable Diffusion applications. So that was the first year of growth, was all Stable Diffusion. And then we started expanding into adjacent use cases, like audio and video and stuff like that. And it was like early teams that, you know, they were deploying JNI inference. They didn't want to deal with deviling infrastructure. They wanted to move fast. Like we started aggregating a lot of different GPUs across a lot of different clouds even back then. We always knew we wanted to be multi-product, but we started with the serverless, like the sort of compute primitives. We felt like it was so general purpose. And we focused a lot on like kind of what I think was like high code use cases, like focus on like machine learning engineers and researchers who want to write code and focus a lot on being kind of lower in the stack, right? And pretty early, we built also a storage layer, a file system, to store data, in particular, container images.

Then we launched the Sandbox product in 2023. And the reason why we launched it back then, we started seeing people messing around with LMs to generate code. We figured actually a lot of stuff we built on the infrastructure side is really well suited to run that code. And they're going to need a place to run that in an isolated way. But it didn't immediately, it wasn't like an overnight success, but like basically 2025 it kind of blew up because I think models got really good at writing code back then. It's about a year ago. So that was sort of our venture into like kind of becoming like a multiproduct vendor. Like starting about 12 months ago, like Sandboxes has been in a tier and grown like, I don't know, 1,000x almost or something like that. It's about a third of our revenue right now. Standalone is the 300 million run rate business, Sandboxes. so that's been incredible in terms of becoming a multi-product vendor we just launched also a training product so sort of trying to lay out all the different building blocks and what does that suite of tool look like and what are the toolboxes in this

new world of AI applications they're all going to need a bunch of new primitives and we want to offer all those primitives and what does that look like so LM inference was something we were a little bit on the sidelines for a while until actually I would say the DeepSeek moment. So that was about a year and a half ago, like early 2025 when DeepSeek came out, which was like the first Chinese open source model. It was like actually really good. And then I was like, fuck it. Like we got to get in here. And like we started building up a team to do a lot of inference optimization. Models has always been really good at LM inference, but we didn't necessarily have a product for it. In fact, we didn't even have an inference product until very recently. We're just like really good at hosting inference, including LM inference. But it was like kind of custom. and a lot of the teams that want to run LLM inference a year or two ago, like they had to deploy it themselves or like we would work with them. But we started seeing a lot of demand with the DeepSeek moment over the last 18 months, started seeing a lot of growth. LLM inference is now about a third of our revenue too. So we've seen incredible growth with that.

And a few months ago, we shipped a number of LLM-specific features because we think, you know, we kind of laid this like foundational layers, like the storage, the compute, the serverless functions, the files, et cetera. I think there's a lot of opportunity to move up in the stack. So we started moving up in the stack now and launched a dedicated inference product. So you can get like single tenant models, like all the state-of-the-art Chinese models or a bunch of US ones too, like one single click or CLI command. You can just deploy on modal as well as a bunch of shared endpoints. Like you can just call models off the shelf without having to deploy them, fully token-based pricing. And so I think those are the first out of many more products to come, like slightly higher up in the layer. We think there's a lot of opportunity because we're sitting on this very solid foundation in the compute and the storage. Maybe you guys will start at the bottom of the stack and work our way up. But going back in time, you touched on it, but you guys invested in some really hardcore technology in the early days.

You built your own file system, container runtime, those types of things. How does that show up in the product today? when the experience developers can have. Yeah, I think a massive source of alpha in this market is just not being afraid to do hard things. So we went all in when we realized in order to deliver on the developer experience we wanted to have in the very early days, we realized we kind of have to build our own file system to do container images really fast. So we started building that in the very early days. And that became, over the next few years, you know, it still is like a very massive source of alpha. It's been a longer journey for sure. Like we've had to go a much longer route, like building our own file system, building our own container runtime, our own schedule. Like we don't use Kubernetes and Docker. Like we use like very little off the shelf software. We use GeoVisor, which is a fantastic Google product for isolation. Recently started using also KVM for certain tasks.

But like decided pretty early on that like a lot of the stuff we're going to have to build ourselves. And I'll just do some really cool stuff. I think, you know, one of the best examples where that shows up is we can do cold starts extremely fast. Turn out, like, the initial developer experience that we wanted, like, very fast feedback loop when you're launching stuff in the cloud, that is also very helpful when you're building autoscaling applications that need to scale up and down very fast. So, first of all, loading the containers very fast. But over time, we also launched a couple of really cool features around CPU snapshotting and GPU snapshotting. So we can very fast load data from the disk or from network straight into the GPU and fire up even these large models in a few seconds, which is really cool. And over the last six months, we built this incredibly complex system that basically uses RDMA over peer-to-peer between the nodes to copy data very fast. So I think a lot of this stuff is like kind of, you know, accumulating and really starting to pay off.

Like we built some crazy stuff. And now that's like, you know, we're really kind of, you know, reaping the dividends of that. Those sort of investments that we made for many, many years was a longer route. But, you know, now I think we have some incredibly unique technology that's very hard to replicate. And then kind of moving up to inference, what do you view is you guys have made, you know, tons of investments there as well. and optimizations. What do you view as the unsolved problems in the inference world or in the inference stack? I think we're still pretty early with inference. I think, you know, a lot of stuff is like, it's like one of those like very empirical sciences. Like I think people are trying different things and seeing what works. I mean, over time, like people have kind of figured out a lot of stuff, you know, take something like disaggregated pre-fill, right? Like it's a very non-obvious idea. Like you sort of, you know, do the pre-fill on a set of nodes and then you copy the data over RDMA and do the decode and kind of swap. There's some, like, very non-obvious things, but, like, we've invested a lot in those types of things. The KV caching, there's a lot of complexity around, like, KV caching.

Of course, like, running more efficient kernels, that's something people do a lot. There's a lot of these things that, you know, we have a team of, I think, 20, 30 people right now who are extremely focused on just, like, LM performance optimization. And, you know, it takes a lot of work. You have to stay up with research. You have to read papers. You have to look at all the software coming out all the time and try and see what works, what doesn't work. It's still kind of an empirical science. Over time, I think, you know, patterns will emerge. I think also the models themselves will change very quickly. Like, now we have, like, a mixture of experts and more complex things. Over time, like, who knows? Like, maybe we'll swap out the transformers for something else. And, you know, that would change how, like, KV caching works. Like, there's all these things, like, that could, you know, change overnight if someone figures out, like, a very different paradigm for all these things. And maybe they will. but maybe they won't. I don't know. There's also obviously a lot of stuff around quantization and other things. So, yeah, it's a wild world out there. But, yeah, we're playing around with all these things.

But it feels like there's so many folks that are working on, you know, different parts of improving inference. And obviously what's, I think, been really interesting in the last two, three years is a lot of those optimizations and improvements have been in open source, right? And, like, you know, for the broader community, maybe shared via research. as a CEO, how do you think about like the level of what the kinds of problems you want your team to go solve versus, you know, it's like, hey, we're going to benefit from just the entire community, like pushing this face forward Hi I Anne Maris and I the producer of Unsupervised Learning Thank you so much for listening to the show And if you enjoyed the conversation please subscribe Subscribing helps us grow which helps us continue to lock down some amazing guests Thank you so much for your listenership and see you next episode. As a CEO, how do you think about like the level of, what kinds of problems you want your team to go solve versus, you know, it's like, hey, we're going to benefit from just the entire community, like pushing this face forward. We actually love open source So I think a lot of our alpha is in the proprietary side, in the infrastructure layer, like all this cool stuff around autoscaling, for instance.

So I think on other inference providers, we've actually leaned very heavily into open source. We work very closely with both the VLM team and the SG9 team. We contribute very actively to both those repos and do a lot of optimizations with them. And so I think this is maybe a little bit of a different approach than maybe other inference providers. Of course, there's going to be certain things that we keep internal, but almost everything we push to open source. And I think it's because you have so much alpha in the infrastructure anyway. We might as well just work with the amazing open source engines that are out there that maybe weren't state-of-the-art a few years ago, but now they are. And that's an incredible thing that we'd love to leverage, frankly. You know, you were talking about heterogeneous compute and the different places where you're getting compute from and also how the, you know, kind of inference stack is changing. How important are, you know, the long tail of chips to you guys, the non-NVIDIA chips and things like that, as you think about your supply or kind of what your serving stack looks like over time?

I mean, I think NVIDIA is still like the king here. Like, they have the best chips. There's way more demand for NVIDIA versus anything else. Like, you know, we're curious about some of the other chips. Like, you know, we made some experiments, like, taking a look at it. But NVIDIA definitely has the best chips. But, you know, like, looking at AMD, Terranium, TPUs, et cetera, like, starting to be, like, pretty good. So we're always curious about that. And then, of course, there's a lot of, like, older chips, too. And I think the challenge has been, like, for a lot of the LMs, like, you're kind of constantly pushing, like, you know, how many parameters can you fit? So typically the latest and greatest LMs, they will always require a full node, maybe in the future even multiple nodes, who knows? So like eight B300s. They will require the latest chip and like eight of them, because that's how many you can fit on a node. But there's plenty of models outside of that, like some older LMs, smaller ones that are optimized for performance, might be fewer parameters. And then outside language, there's a lot of stuff around like voice

and audio transcription or speech synthesis or pharma, like bio stuff. A lot of that stuff, actually, they don't require the latest and greatest GPUs. They actually get more cost-efficient serving on hoppers or even mPairs, or different types of L40X or RTX Pro 6000, stuff like that. So we actually, I mean, we are ramping up very, very aggressively on Blackwalls, primarily GBs now. but we see actually like plenty of demand also in the older chips which I think is quite exciting. There's just so much stuff going on. People forget about the hoppers but like they are our most common type right now still at modal like H100s. We run some very like large production use cases based on hoppers. How efficient like is model to chip matching today like you know in the industry at large? Do you feel like most people are using the like optimal you know chip or from a price perspective for what they're working, or is there still a lot of room for improvement there? There's a little bit of room to improve.

I mean, one of the things that we want to do over time is kind of automate that for people. I think in practice, it's less of a problem than it might seem like because often you're kind of memory constrained. So knowing how much memory you need for a model, you basically pick the cheapest CPU that fits the model. That's usually the heuristic that works in almost all cases. So if your model is 80 gigs, like you can either fit it on a hopper or an older a100 80 gig i guess in that case it depends on the flops um but but in any case it's like whatever whatever gpu like is the smallest that fits your model i think over time like there's a bunch of stuff we've been thinking about that where like basically you just deploy something a model and then we manage like different ships and maybe automatically give you discounts because like we can you know the more options we can have to like run it on a lot of different ships like that lets us aggregate capacity in a better way and run things on higher utilization because then we can switch between different GPU types as there's underutilization on those.

So that's something actually we're quite interested in doing over the next six months is giving people an option to deploy things in a way where we automatically switch types underneath in a way that's still consistent with your dollar per request or dollar per token or whatever. I think there's incredible cool stuff to do there. Super interesting. And one thing I imagine is so interesting in your role too is like you're obviously predicting for like what your compute fleet should be. And obviously you see what the workloads are today and you probably have some sense of, you know, what models run well on what chips. Then you're also trying to predict like, you know, hey, a year or two from now, like what, you know, what kind of workloads will we have? I think there's lots of, you know, there's lots of different debates about whether we're going to, you know, most tokens running will be really large models or actually, you know, we've saturated most tasks and these models will get smaller and smaller. Like, you know, as you think about kind of the evolution of your fleet in the next year or two, like what, you know, what's your perspective? on where you'll see relatively more different kinds of workloads and less. I think there's been kind of basically a trend of larger and larger models over the last three, four years.

You won't bet against that. I definitely feel like that trend is going to continue. We're going to need – if B300s had more memory, I'm sure a Kimi K3 would be bigger. People are kind of training the models in a way that they fit on these chips. So I think on the, at least on the LM side, like on the, you know, LM's are in a way like kind of a crude type of world model. Like you're trying to like encode as much information as you can in these LM's. Like you need like very big, like almost like lookup tables of facts. You know, anyway, it's kind of a conception I think about it. And that sort of pushes like the memory requirements a lot. I mean, at some point, like it might taper off, like who knows. But like for now, like people will make use of every possible memory. Like the more memory, like they'll just train more parameters for sure. You obviously sit in some ways in the eye of the hurricane of a very exciting, large, and competitive market. And so I'm curious as you think about what ultimately will differentiate players in this space and win out.

I think there's so many true believers in the size of the overall inference market, and I think lots of folks are asking what actually wins out at the end of the day. And I know you have a good thesis on this, so I would love it for you to share with our listeners. I mean, I think a lot of the inference market will be quite rough, right? Like, I think the sort of straight down the middle, just like, you know, serving tokens, you already see sort of open router basically commoditizing the underlying providers. I tend to think, you know, like, the real alpha or like the real point of differentiation, like there will be a lot of differentiation just in like more efficient serving. But it is a little bit of a race to the bottom. And I think there's going to be a lot of opportunities to sort of stay ahead of competition. You're going to have to, like, put a lot of effort into that. But I think the more sort of durable differentiation will definitely be in, like, you know, providing the entire platform of a lot of different things, which is, like, I think something we are quite well positioned to do. We have the sandboxes. We have reinforcement learning. A lot of customers doing also normal fine-tuning. A lot of customers, obviously, you know, deploying other types of models.

We have batch processing. I think my view in general is like customers, they don't want to buy like single trick ponies. Like they want to buy like a multi-product set of tools. So that to me is like ultimately like probably the biggest source of differentiation is like, you know, how do you provide that entire set of tools? How do you provide this like very general purpose thing that lets engineers like iterate very fast, deliver value? And then like this with day-to-day, like we're still going to have to like go and do the knife fighting of like, you know, staying ahead of your competitors in terms of performance and use the latest kernels and the much more efficient KVCash. And there's going to be a lot of that stuff for a long time. There's just so much money in it. But I'm kind of pointing out, I think that's sort of, that's like less durable as an advantage. I think that, you know, because over time, like people, these tricks will sort of spread and people figure out how to copy them. So the more durable thing, I think, is like being that sort of portfolio of a lot of stuff around inference, not just the inference in itself. of having all the models and a lot of stuff like around reinforcement learning maybe or other things.

You know, the biggest modal customers are on multiple products. Maybe can you touch on a few of the more interesting ones? Yeah, we have some incredibly cool customers like Cognition is using us at large scale both for inference and sandboxes to do a lot of reinforcement learning on modal, to use sandboxes for all the rollouts. We have customers like Ramp using this for background agents, for the internal coding agent that they actually launched as a standalone product today that they're selling. We have customers like Suno and Runway that use this regenerative video and audio. Decagon is another customer that uses this for a lot of customer support stuff, voice and text. DoorDash uses this for a bunch of internal applications, mostly LMs, but also all kinds of other stuff. We actually have a lot of customers in computational biotech using this for various types of drug discovery, using molecular simulations and sequence alignment and all kinds of stuff. We have a bunch of customers in robotics, 1X, and a few other ones using us to train robot models, like doing a lot of training, but also batch processing of video and things like that.

A lot of customers in the data generation space, so it's the customers who are building environments and running benchmarks on modal, often using sandboxes. I mean, one thing you kind of mentioned is portfolio of products, and I think that's like, you know, one of the most interesting parts of the business you're building. And I imagine, like, honestly, one of the hardest parts must be just like prioritization of all the different things you could build. There's probably so many new things that come up of really exciting tools you can build for developers. And, you know, I guess I'm curious, obviously, you kind of, you know, with sandboxes, you have, you know, incredible products around, you know, coding agents around, you know, reinforcement learning. You know, how do you think about kind of building down that vector? There's also kind of like the fine-tuning or all as a service stuff that some of your competitors seem like they prioritize. Like, how do you think about, I guess, you know, where to prioritize within that portfolio? I think it's an incredibly hard problem at, I mean, most startups, but especially multi-product companies, right? Like the sort of balance of like, you know, first of all, like when do you go multi-product and then where do you put your bets?

Like we're 150 engineers or something like that today. And I struggle with it every time, like every day. Like, you know, I wish we could, you know, do that and, you know, whatever. But I kind of know that even if we had 1,000 engineers or 10,000 engineers, like it's all, you know, there's lots of dollars. Like, you know, I wish we could build a rocket and go to the moon and whatever. Like, it just never goes away. So, like, you kind of have to be just, like, ruthless with prioritization and really think about, you know, what's the highest ROI and, like, where do you draw the line? And you have to be sometimes, like, pretty aggressive also saying, like, actually, like, I know we started on this project. I know you're really excited about it, but this other thing kind of came up. So we're going to have to pause this for a second. We'll get back to it later. We've had to do that many times, too. It sucks when it happens. But I really think that sort of prioritization and how do you allocate resources. I mean, resource allocation is basically the number one thing that I think a CEO has to really be ruthless about it, a company. And so much value is being created, like allocating the work to the right resources.

The modal roadmap in each board meeting is extensive and very daunting, and I'm always impressed by the amount of software that the team can ship and hardcore software while keeping the system running. How much has, you know, you guys were building the company pre-ChatGPT and then, you know, post-ChatGPT moment. How much has that changed, you know, maybe on the product side, you know, what you guys were doing, but then also when you were thinking about building and the velocity that you can ship software at? Yeah, I mean, we've always benefited from, you know, being quite well positioned at a time where, like, AI kind of blew up. and that's been incredible to see. And we're powering so many different AI applications today that I had no idea would exist four years ago or five years ago when we started Modal. So this has been kind of incredible. And like you said, software in itself has also, how we build software has changed a lot. People write, I mean, I think probably at this point, like 95% of the lines of code that gets merged at Modal

is probably written by agents to some extent. Infrastructure still, I feel like relies a lot on, agents are kind of good at it. I've actually been on a tier the last few days, or the last few months, actually, writing an insane amount of code using agents. It's kind of incredible what they can build. You kind of still need to supervise them quite a lot. Infrastructure has a lot of tail risks around reliability and making sure that things scale and stuff like that. So at least for now, I think software engineers are pretty safe, and I think we still benefit from having incredibly good engineers that can supervise these agents. and I think probably for the next few years. Another thing I think has been kind of interesting to see is like, actually like, I think the adoption pattern is actually the opposite of what I would have expected. Like I thought it would be the junior people that would adopt agents the most. I've actually seen almost the other way around. It's like the pretty senior people because for them, it's like giving them like 10 interns. Like they've been kind of constrained by like typing speed for a long time, but they have all these ideas in their heads. And I think using agents now they can get it like,

they can get it out like much faster. They're really good at building a test suite, like, you know, give the agents like, Here's an extensive test suite. It implements all of, you know, the POSIX standard for file systems. Now go write a new file system. And it just goes on, like, just writes a file system. And it takes, like, you know, three days to do it. So we've seen, you know, some of the very senior engineers, like, go, like, complete agent maxing. And I don't know. I mean, it's kind of incredible, like, what you can do with agents. You can move very fast with them. At the end of the day, it seems like humans are still safe as software engineers for the next few years. I don't know. We'll see. How have you found the velocity of shipping software has changed what you guys need to do with code review or CICD or deployment? Have those pieces of your development lifecycle changed at all? I think a little bit I think we coming up against some of those constraints Like we had to spend a lot more time on CICD Like always even long before agents I was a CTO for many years I was always obsessed with CICD as very, very important to figure out. So in that sense, I don't know if it's changed my view as

you really need to have automated deployments, automated tasks, all these things. Now, by the way, agents are pretty good at writing tests to it that weren't a year ago. So I don't know if we've necessarily yet run up against the foundational limits, But, I mean, Anthropic, like, you know, a couple of days ago, they announced they're moving off Git, I believe, or at least GitHub. We might run up on the end source limits at some point when we have, you know, 10x the amount of commits every day. But it's interesting to think about kind of the, you know, how different the world is with these really powerful coding agents. And obviously, I think, you know, part of what I believe your co-founders talked about is kind of the shift even in who the consumer of modal is over time, right, from a developer to, you know, to agents themselves, right? And I'm curious how you think about what that means for modal as a business and some of the decisions you might make differently as more and more of folks interacting with modal are either doing a VM agent or purely end-to-end agent. It's actually changed a lot less than I think I would have expected. It turns out agents, I mean, in the early days, we built modal in order to give the engineers the better developer experience and the ability to move faster.

It turns out agents kind of want the same thing. like they want fast feedback loop they want like good error messages, good documentation a lot of the stuff like that that engineers want too so most of the stuff is actually quite similar I do think it's like certain things are a little bit harder like one of the things we've found for instance is like deprecations is much more annoying because agents are trained often on like a lot of old code that exists online so removing features is much harder for us like we can't just rely on like you know remove it from the documentation and wait three months because, like, it's in the training data. People will use, you know, LMs to generate modal code, and that will be, like, you know, using deprecated features. And that's, like, quite hard to change. So we've had to change our deprecation window for quite a lot. Another thing we've started doing quite a lot is also setting up a lot of automated pipelines to basically have, like, LMs continuously write modal code and making sure it compiles, or not compiles, but it builds and runs, trying to just, like, catch, like, things that agents struggle with.

And then maybe another thing we spend a lot of time on is more, like, kind of, like, agent optimization stuff, which is actually more, like, similar to, like, search engine optimization. Like, if you, you can go and ask, you know, Cloud Code, like, something like, hey, like, I want to do something with inference. I want to deploy this model and not say anything about modal and just, like, see, like, which tool does it recommend. And that, obviously, is something we wanted to recommend modal as much as possible. So that's something we track internally, make sure that, you know, over time, like, and actually think that's like something that will change software engineering a little bit. Maybe I was like to do worse. Like, I feel like in a way, like, because of the training data and the training loops, like, there's more like, you know, like incumbency power in a way. Like, there's more staying power in software today. Like, if you get inside the training data and you get the models to recommend, you know, a certain framework. I've heard people say like, oh, I'm just going to use React for everything because agents are really good at it. I'm not going to try these like new fancy front end frameworks.

Right. Like I think there's like something similar with like a lot of the frameworks today. It's like that same power is actually very, very valuable now. So we obviously want to make sure that agents recommend modal as much as possible because that creates like a little bit of a flywheel. And one thing I'm struck by is you obviously had this like very prescient bet in 2023 on sandboxes. Right. It sounds like you guys had kind of this instinct of like, hey, this is going to be something that's required for coding agents. And then, you know, it took a few years for it to really take off. And then, you know, it became this really important primitive. I'm curious, is there anything else that you guys are working on today that you feel like you're a little early on, but like, you know, you're equally as excited? And how much of your product portfolio is like those kind of bets versus like, okay, we know this is a hair on fire, like here and now problem. Jacob, are you fishing for the next thing we should be investing in? 100%, 100%. I'm nothing if not shameless. I feel like Soundboxes, they did create a little bit of like a high bar for us. Like everyone's like, what's the next thing? you know what's the secret sauce you're working on and like what's cooking in the lab you know and and i i wish i had like an easy answer like yeah like 100 confidence you know this thing's

would you have answered sandboxes in 23 or was it just like part of a bunch of things you were doing no because what sandboxes represented back then was like a little bit of like an asymmetric bat like sandboxes was not like a very hard thing for us to build back then like to getting out like a prototype of sandboxes took like upshot like a few days and then i spent a little bit of time you you know, like, improve InvestEK and some other people, like, started building. And then we started seeing some customers. We rewrote it. And then in 2025, when it started exploding, then we had to rewrite the system, like, five times just to handle, like, more and more scale and make it faster and more reliable and things like that. But the initial investment was actually pretty basic. So in that sense, like, it was a little bit of, like, a sort of asymmetric bat, like, kind of a low downside, high upside bat. I don't know if there's, like, anything similar today that we're working on. So I am, I mean, like, just kind of called out, like, a couple of, like, you know, things that I'm quite bullish on. And these probably represent, like, way larger investments, like, you know, larger risks and, you know, in terms of, like, that size, low latency.

Like, I think, you know, there's going to be a lot of voice-to-voice, a lot of real-time video type stuff. That requires us to rethink a lot of how we manage to compute, like, globally. How do we get down the number of hops? Like how do we scale up and down in different regions? Like almost like building a CDN for inference. So I'm pretty excited about that. I think training is actually kind of an underrated thing. I think there's a lot of opportunity in training, like mid-sized models and reinforcement learning. That's been a big bet for the last couple of years, like getting all the RMA, Infinity Band type stuff working. I think a lot of VM stuff is quite interesting. I think supporting full-blown operating systems like Mac and Windows and Android and iPhone could be kind of cool. I think there's a lot of upside there. I have a lot of those ideas in my head, and we're building a lot of those things. I don't know if there's anything that represents the same explosive potentials like sandboxes. Maybe if we have this podcast again in three years, we'll kind of look back.

We're going to make it recurring at this point, so we'll look forward to revisiting that. You know, I guess, like, obviously, you know, sandbox has become this fundamental primitive for, you know, coding agents, and it feels like that then brings you into this world where there's, like, you know, I imagine you look at the current state of coding agents and you're like, God, there's so many more tools that would be helpful for these agents to make them more effective for people that use them. How do you think about, like, you know, whether you further build into that world or, like, the things that are interesting for you there? We do think about it, but I think we think about it more from an infrastructure perspective. I think it's very easy to kind of get lost in, like, you know, whatever is, like, the latest today. When I look at, like, the architecture of, like, agent harnesses and, like, how we're building agents and stuff like that, I still kind of feel like we're so early and these architectures are evolving so fast. And, like, what our job, to some extent, is, like, trying to figure out is, like, what are, like, the sort of core building blocks that we can put out today that, like, will last, you know, like, people will need them for the next five years. And that, to me, is, like, less obvious. Like, we are interested in, like, sort of agent orchestration as a category.

I think there's a lot of interesting stuff there. We're interested in, like, you know, evals and model gateways and all kinds of other stuff around, like, you know, the whole, like, you know, life cycle of learning from agents and innovating on agents. But I think it's, like, less clear exactly, like, how the architecture. My view is, like, it will be more cloud-centric and it will be probably more, you know, like, durable objects, like kind of workflow engine type driven. and we have some ideas of stuff to build in that space. But I find it harder to lay out exactly what agentic development will look like in five years. Part of it because I also think to some extent all software will basically be agentic. I kind of hate the word agentic in certain ways because I feel like it's like, okay, so it's using AI to make decisions. Well, shouldn't all software do that? So I almost feel like if you look at it on a five-year horizon, I think all software will become identical. And what does software look like then?

I don't know, but I hope we can build a lot of the building blocks to support all kinds of aspects of, in general, just how software uses AI. It feels like one of the hardest parts of building an infrastructure company. Across the board, for many people building AI infrastructure companies, what actually persists versus a pattern is like RAG from two years ago or whatever it is. RAG is a good example. Exactly. Like if we had made a bet on like RAG like two years ago, like it would have been kind of a bad thing to do. So it's hard. But I would also say like one of the benefits of building an infrastructure is like you can take a longer view and you don't have to like listen to all the noise. I mean, you have to listen to the noise, but like you kind of have to like ignore a lot of it and trying to figure out, OK, what are the patterns here? You kind of alluded to some of these other, you know, other workloads that like low latency workloads like voice to voice and real time video. So I feel like there's a lot of chatter in the air about, like, multimodal inference, robotics inference, these other categories. I'd love to hear you just riff a little bit on, like, to what extent are those problems, you know, similar to the set of problems the inference markets had before?

To what extent are they, like, different and new? I think they're all quite similar, but all, I mean, they're like snowflakes, unique in their own way. Robotics is going to be a lot of it on device, of course. So that's a big difference. And, you know, I don't sort of envision like we're going to support, you know, most of what robotics looks like because a lot of it is going to be on the devices themselves. But I think we're going to be able to support a lot of the testing and development and training of robotics and a lot of the simulation environments. So that's something I'm very excited about. On the multimodal stuff, I mean, there's audio and music and image and stuff like that. Like people don't talk about it as much, but that keeps going like crazy, too. I feel like LMS have kind of dominated the airspace in the last few years. but I'm just incredibly, you know, excited about all the different applications. Like, why are, like, voice assistants so bad in 2026? Like, I don't get it. Like, they're just horrible. Like, I have a Google Home, and, like, every question is just bundled. I just don't understand what I'm talking about. Like, you know, so I just go to ChatGPT. Like, why can't you just connect all of it, like, end-to-end and train, like,

an assistant that can, you know, do, like, speech-to-speech? I think there's a lot of we're going to see a lot of sort of incredibly cool applications in that space I always think about you know you look back like you know four years ago three years ago whatever you know could you guess that like things like instinct or Suvino or like cognition or like none of those things like are very easy to guess like you know like three years ago so I kind of feel like also like if you look down the road like three years like most people don't have an idea of like what the world is going to look like in three years. It's so hard to know. But what I do know is kind of judging by the past three years, it's going to be like, there's going to be some really cool applications. I just don't know what it's going to be, but they're going to be very cool. And hopefully they'll use modal. I was looking back at an email you sent me with the modal series A pitch deck, and this is what you said. This was after we had signed a term sheet. By the way, we never actually reviewed a modal deck, but here's something we made earlier that I just updated

with the latest numbers. Could be good to review at some point. When you sent this, you were at roughly 2 million in run rate. You've crossed a billion run rate, which is insane, in three years. What's been the biggest change to supporting that kind of scale in just a couple of years? First of all, I actually realized the other day, I said on another podcast that I've never made a pitch deck. Then I was actually thinking about it. I was like, maybe I did make a pitch deck after Series A because we had a couple of angels that came in and they were asking about it so i was like oh fine i guess i'll put something together so i i i realized i did lie like i did make a pitch deck but it was actually but you never showed it to us so i don't know that it counts yeah so i don't know if it counts like i never like sent a pitch deck to an investor i sent it to a few angels actually uh but and that's the only pitch deck i've ever made actually um but uh in terms of the scale yeah i mean we've grown you know 500x i guess in terms of revenue something like that since then uh over the last three years so a lot of growth I think in a way like

I don't know maybe I'm like just you know my experience has mostly been at hyperscale startups like I was at Spotify for seven years and some other places so for me this like fast growth is kind of all I've seen like you know there's like every six months you have to rewrite all the systems you have to kind of rethink everything you're doing you have to kind of change all the processes like you always feel like you're behind. I don't know. I kind of like the mass. I guess we'd love to, you know, since we have your mind on the podcast, we'd love to kind of zoom out and maybe talk a little bit about the broader inframarket and some things that have been in the news of late. Maybe most top of mind these last few weeks, I feel like there's been tons of discourse about neoclouds, right? They like, you know, have the debt's trading, they've traded down, they're kind of back up a little bit now. And I'm curious, like, what do you think happens to this landscape over time? I think the neoclods are like fine business models. Like, I think people have been, like, confusing it, like, a little bit. They're like, wow, there's so much demand. Like, revenue is going to grow, like, so fast.

And then, you know, there's a little bit of sort of a – almost like a VC mindset of, like, oh, like, thinking in, like, revenue multiples. But I think all of these, like, these neoclubs, like, growth is actually quite expensive for them because they have to go and raise a lot of debt. And so I don't know if necessarily, like, you know, just because, like, they grow a lot, like, you know, like, it's very, like, capital inefficient. like these companies at their core like financing vehicles they're doing leasing of hard assets and that's like inherently harder to scale the capital now that that market is kind of saying well we deployed a lot of capital into neoclouts like maybe hold on for a second like you know we don't have that much more capital to lend and so the interest is starting to go up but also by the way like these companies have incredibly high leverage so of course they're going to have more volatility like the residual equity value is going to go up and down much faster on the stock market with GPU prices and stuff like that. But they're like fine business models. Like, you know, they kind of locked in customer contracts and raised debt against it.

And like, you know, now they're paying off that debt. And in many cases, it turns out the residual value has been higher than people expected. Lifetime of GPUs has been longer than people expected So they can resell once they paid it off like at a higher than people thought So they going to make a bunch of money from that It takes a few years to get there But I think in that sense, NeoClass will do well, but not particularly differentiated. So long term, I think they're going to start with that. You work with a lot of them. Is there any differentiation you see? A little bit, but at its core, it's relatively undifferentiated space. I mean, my view is Like, you know, I mean, Elon was like tweeting the other day, like, all of your clothes are going to die. Like, I think that's kind of bullshit. I do think Elon had a point when, like, basically, like when he implicitly was saying, like, SpaceX has a much lower cost of capital. I don't think there's a coincidence that if you look at the three successful hyperscalers, Amazon, Azure, or AWS, Azure, and GCP, that they're all attached to companies that have very solid cash flow, and they're all public companies that have very low cost of capital.

because that at its core makes the life a lot easier if you want to build something that's very capital-intensive. So I think in general, that is something that Neoclass will struggle with. They're going to have a high cost of capital, and it's going to be hard for them to grow fast because of that. So we'll see about the market. But they're decent models. I think there might be an opportunity to buy the dip. I don't know. You had an interesting tweet that was a lot of the Neo labs could become Neo clouds, you know, if their model doesn't work out, but they have locked in these contracts. I definitely could see the world going that way. If, you know, if you have valuable compute that continues to appreciate because there's so much demand. Yeah, I mean, Mr. All kind of did that, right? And, you know, like, it seems to be fine. I think a lot of people are making these, like, very large buying decisions and they kind of rationalize it. Like, they're like, well, if you don't need the GPUs, we can just resell them, right? Like, that's almost like a form of insurance. And you have to, you know, assume with, like, all these Neolabs, you know,

there's so many Neolabs out there that have each raised $1 billion. They always raised exactly $1 billion for some reason. I don't know why. I guess they like that number. And then buying GPUs. Very precise financial modeling that points me out. Yeah. So some of them, I think, will probably work out really well, you know, but I imagine in the grand, you know, like, some of them will not work out. And if they don't work out, well, now they're sitting on a ton of GPUs. Like, you know, let's resell those GPUs, right? Like, especially if there's ever, like, a scenario where, like, VC funding pulls back. Because I think what's kind of a scary thing about some of these commitments is, like, you know, effectively, like, what happens today when you make a GPU commitment is you only have to pay about 30% up front. Sometimes a little bit more, sometimes a little bit less. But what that means is, like, you can raise a billion dollars. You can go out and get GPUs for $3 billion. and only pay a billion dollars up front. I mean, not exact math. But the problem is, like, you know, in a year, now you're on the hook for another few hundred million, and, like, every month, now you have to pay.

So I think that creates, like, a little bit of, like, implicit leverage in a lot of these companies and a dependence on subsequent fundraising. And, you know, that works as long as they can go and raise VC dollars to fund the GPU bills. But at some point, it might not work. And in that case, I think you're going to see a little bit more of a secondary market where some of these Neolabs might pivot to doing a little bit more of the sort of NeoCloud style, like we have a bunch of GPUs, or just resell it back to the cloud. I don't know. I'm very bullish on AI as a whole, but I always look for these sort of structural risks. So overall, I think it's going to be an incredible application and incredible demand. But is 100% of the Neolabs going to succeed and be a trillion-dollar company? Probably not. I feel like on the AI Infra side, one big category that has gotten a lot of buzz is like this URL as a service market and kind of a bunch of folks, you know, almost forward deploying and like helping companies build their own models. And, you know, you've had folks like Satya saying, hey, every company is going to have their own model. Like, do you subscribe to that?

And like, what do you kind of think of this, of that overall market? No, I think some companies will have their own models. I think that like weighted by dollar revenue, like I think most models will probably be post-string, right? Like, but like weighted by company, like in terms of numbers of companies, most companies will not have the need to do their own post-training, right? But the large ones, like Cognition, is training their own code models and Cursor Date and all these companies are. And those are the ones that also serve a lot of tokens. So I think when you get reinforcement learning and post-training, certain types of parts of it is very attractive, right? Like you can work with companies and serve, you know, better models and do it at lower cost and higher throughput and all of this stuff. Some of it looks like it's kind of like just a service business. So I think a lot of this sort of tension over the next few years when you look at post-training is going to be, can you turn this into product? Like can you sell that product to these companies? This is something we think a lot about.

It's like how do we create this like amazing product you can just kind of sell and not turn it into just like a service business. And then, you know, I guess the last one before we move to the quickfire round would just be like, you know, I think there's this question about, you know, how many tools live under like a single infra provider in the AI world versus, you know, our separate categories. I've always struck when we talk to you about like just how much you want to go build and all these kind of building blocks. I'm wondering if there's categories of AI infra, you know, that you think will be unique and distinct from the stuff you're building that you're optimistic on and that will kind of like live as still large standalone opportunities. yeah I mean we're not going to go into model training I think type safe with JAV is going to be there's probably going to be a lot of these custom models I'm sure in other verticals too other modalities like audio and video and music and obviously on the application side there's going to be a lot of incredible companies and then there's probably going to be a lot of companies focusing more on software development and stuff like that And, you know, I don't know, like my general view on these things is like there's always like more boxes on the market map than you expect.

But I also do think that overall, every company that is successful in the end, to some extent, has to become a multi-part company, too. So I don't know if that's a tension or not. Like, I think, you know, we obviously want to find, like I said, a cluster of like very adjacent things and do all of it in a way that like, you know, maximizes the revenue and the shareholder value. but I you know by no means you know are we going to try to expand into everything I think people kind of make this like false assumption also like you know you look at like Anthropic or OpenAI and they're like why don't they move up to stock and launch their own law firm and launch their own whatever like they're very good at like serving these models and doing that at scale that's how they achieve scale right like and they're very good at doing that like I don't think it makes sense for every company to do everything and so similar for us like we're not going to do everything but we're going to do a very large cluster of adjacent things and do it exceptionally well. And I think there's going to be a lot of room in the market to create a bunch of very large $100 billion or $1 trillion opportunities in the market.

We'll end on our quickfire round where we stuff probably a similar set of questions, what we asked you three years ago, maybe a little bit of an evolution. To start, maybe I'd love just one thing you've changed your mind on in AI in the last year. Well, the opportunity side of the sandboxes has obviously been kind of crazy, right? Like, I didn't think it would be anywhere close to, you know. I thought you knew that in 23, and that's why you made the big bet. Yeah. So that's been kind of incredible. Coding agents have changed a lot of what it means to be an engineer. How, if at all, has it changed how you think about hiring? Not much. I think we're still going to need humans for a long time. I think the sort of demand for humans is just going up the more productive they get. I think we've seen this for the last 30, 40, 50 years of software engineering, whether it's compilers or high-level programming languages or CICD. Every time you make engineers more productive, the demand just goes up. And so I think we're going to need a lot more engineers.

I do think maybe on the short term it's going to be a little bit harder for new grads, except for that very high side of the spectrum. But in the long run, I'm very bullish on still, you know, computer science. Actually, ironically, I didn't study computer science, but I'm very bullish on studying computer science in general. Like, if my kids wanted to do it, like, I would definitely encourage them. What's the biggest mistake you've made in building mobile so far? Oof, I don't know. I think one mistake I made in the past is that I think sometimes, like, you kind of want to sequentially do things. You're like, well, first when you do this, this is the blocker. We'll get to this other thing later. because this other thing is blocked on this. But in a world where everything is moving fast, one of the things I've learned is sometimes you kind of need to do both at the same time. You need to assume that, oh, we're doing this thing to unblock it. You might as well do this other thing at the same time just in the interest of moving fast. That's something I'm naturally not mega aggressive, which is kind of ironic because we've run 15x in the last 12 months.

I sort of had to retrain my brain to be a little bit more aggressive and saying this is a massive market opportunity. We need to be even more, you know, like hiring, deploying capital, like whatever it is, like massively just like paralyzing everything we can. Sweden has a ton of interesting startups that have emerged. What are your thoughts on why that is? And has the Swedish house mafia been replaced by the Swedish tech mafia? Did you know that like one of the Swedish house mafia members was my classmate? It's a fun trivia. Really? I did not know that. back in the early 90s growing up in Sweden when I was 8 he told me my bike was ugly I was very upset yeah I mean we've had a lot of tech in Sweden, Spotify, Klarna, you know a bunch of gaming studios more recently obviously Lovable and Agora, stuff like that I think my thing about a lot of this stuff is like success kind of just creates success It's like, you know, a lot of like what makes people go into tech is like they know about some cousin's friend who made a lot of money in tech or something like that.

So a lot of it is I think it's like a little bit of a flywheel effect. I do think that Sweden as a European country has always been pretty entrepreneurial and commercial. So I do think there's a little bit of that. Like we've always been a little sort of maybe going back to the Vikings, like kind of let's go conquer the world. In the past, maybe it's been through music and furniture and stuff like that. But now increasingly with tech. So, you know, there's that. I think there's sort of a mentality in Sweden. Stockholm is funny because everyone in Stockholm thinks they're this mega global city that's part of the world. In a way, I actually think they aren't. But I think the fact that they think they're so global is kind of good. They have this vision that everyone cares about Stockholm. But I actually kind of like that because it means they're out there building things and they're saying from day one, we're going to conquer the world. We're going to do this startup. It's going to be awesome. And I like that, the sort of megalomania. I have this vision of, like, your elementary school reunion and just what a wild group that would be. That's quite the cast in what I imagine was not a particularly large class. Yeah, yeah, it's kind of a random fact.

I guess it's a shameless question. Do you think GPE rental prices 12 months from now will be higher or lower? I think higher. You know, it's funny. I actually changed my mind on this in the last 48 hours, and I'll tell you why. I read the Anthropic S1, and I actually had an agent summarize it for me. But it was kind of incredible how profitable that company is, to a degree, but I hadn't anticipated. I thought, like, in my mind, like 48 hours ago, my argument was like, oh, Anthropic, they have really good gross margins, but they're spending so much money on training that, like, as soon as they go public, they're going to have to, like, slow down in how they grow training and post-training, and they're going to maybe grow inference as much, but not the other stuff as much. But then I realized they're basically a profitable company. So they have zero incentive to slow down on capacity acquisition. And I imagine there's a similar dynamic going on in OpenAI. And so OpenAI and Anthropic has obviously been the biggest two buyers of capacity out there.

And so the fact that I don't think they're going to slow down buying because clearly financially they're doing very well, So I think it will mean that, like, there's going to be a lot more demand. And I think for that reason, prices will probably go up in the next 12 months. I've heard you say that you've always thought of modal as your 20-year project, you're five years in. What do years 10 through 20 look like? Wow. I mean, hopefully I'll hire a bunch of people that can do a lot of those sort of, you know, I'm too in the weeds of, like, everything right now. So I, but I don't know, like, I just like love this stuff. Like, I love to keep working on this. Like, I think, you know, year 10 to 15, we're going to look more like a conglomerate. We're going to have a lot of different products. We're going to have probably some like crazy bets going to moon and Mars or whatever. But I still hope to be quite involved in this sort of the high level, like resource allocation and what are the things we're betting on. And I'll probably be less involved in sort of day to day, like, you know, let's get a

contract from this GPU vendor, like whatever. But, like, I think that, like, sort of, like, high-level, like, resource allocation and stuff like that, to me, that's the most fun. And I hope to stay involved in that for another 20 years, hopefully, maybe longer. Yeah, I mean, if we keep current pacing, at least five more unsupervised learning episodes in that time frame, right? Yeah, exactly. Well, Eric, thank you so much for joining us. We really appreciate it. We'll give you kind of the last call to action for the audience. where can they go to learn more about Modal or anything that you would direct them to? Well, you should obviously go to modal.com. You should also run, keep it installed Modal in Terminal and play around with it. You know, there's a lot of people maybe listening to the show who are building startups or thinking about building startups and, you know, would love to see all the cool stuff they're going to build. And, you know, hopefully at Modal, we can help you with those startups. So let me know. And feel free to reach out, by the way. Like, I always love talking to customers. But feel free to play around with Modal and looking forward to see what everyone's going to build. I'm Jacob Efron, and this has been Unsupervised Learning,

a podcast where I get to talk to the smartest people in AI and ask them tons of questions about what's happening with models and what it means for businesses in the world. As I hope is clear, I have a ton of fun doing this. It's the nights and weekends project, in addition to my day job as an investor at Redpoint. But our ability to get these incredible guests on really comes from folks like you subscribing to the podcast, sharing it with friends. It's really what ultimately makes this whole thing work. And so please consider doing that. And thank you so much for your support and listening. We'll see you next episode.