Modal Labs raises $355M Series C at $4.65B valuation as sandbox product grows 2x per month

Level.

Yeah, I love it.

Amazing. Well, thank you so much for taking the time. Great to meet you. Great to meet you.

Make sure to share uh your announcement next week with us and and we'll cover it on the show. Have a great rest of your day. Talk to you soon. Goodbye.

Bye.

Up next, our last guest, Eric from Modal Labs with some big

It's been a minute. It's been a minute.

I'm excited to catch up. How you been? Good.

You've been busy. What did you do?

Tell us what happened.

Uh we just uh announced a C round today which is very exciting.

How much?

How much?

How much? 4 uh 65 post value.

Wow. Value.

Fantastic.

Racing uh 355 million. So general catalyst led it together with Redpoint.

Fantastic.

What was the specific catalyst that led to this round? There there's been a lot of stuff that's been going on. I I think one particular exciting moment in the last 12 months or really the last 6 months has been uh we've we now have a a product called sandboxes that's just growing insanely fast almost 2x every month uh for the last 6 months. So sandboxes if you don't know basically the idea is like you can take typically LM generated code and execute it in a safe environment. So it powers a lot of reinforcement learning, powers a lot of um vibe coding apps, a lot of background agents. But yeah, I mean in general we've seen tremendous growth also with inference. Moto's always been kind of we started a company 5 years ago with the idea of building pretty general purpose infrastructure. We felt that a lot of infrastructure wasn't really built for AI. And so when you look at all these companies out there building sort of AI applications, they all need different things. Some some of them need sandboxes, inference, training, bath jobs, like all kinds of stuff. and and our goal is to provide all of that and and really kind of build a new sort of almost like a new cloud in a way that that supports all these types of very cool applications.

Very cool. Uh I feel like it's been almost a year since you were on which feel which is like you know 15 years uh 15 years uh

October well there there we go feels like feels like feels like a year ago even though it's maybe uh maybe less. Um what yeah what have been the key kind of inflection points since then sandboxes sounds like sounds like one but I imagine a lot of this is uh the space is moving so quickly if you get a signal from two or three customers hey we have this problem can you build a solution and then you can just see tremendous growth in like a a very short period of time and so in some ways like you have an existing portfolio products that are all you know growing but are you trying to constantly ly be making you know new bets uh based on customer needs or is it is that maybe not even the the approach?

I I I think one of the benefits of building infrastructure is like you can sort of you know look at customers of course and see what they're doing but but I think with infrastructure you can also sort of argue a little bit more from like first principles like what are the building blocks that people are going to need in terms of compute. So sandboxes was actually something we launched 3 years ago and it really started taking off like nuts in in in uh in the summer last year. Uh and and and similarly I I think with with you know in France it was sort of similar. We launched it almost uh four years ago. Uh we're working on some other really cool stuff uh around training and some other products. Uh and I think one of the benefits of infrastructure is you can have a little bit more sort of a first principle like you know let's build the right building blocks and then let our customers sort of figure out what to use them for. How how do you think about uh planning like planning planning demand?

Add a zero to last.

Yeah. Add two add two zeros to everything. No, but yeah, it just feels like the you know the best this is the hardest like demand planning challenge that really any business set of businesses have ever faced in history. Uh

yeah,

it's rough. it like I mean for us it's like we basically look at like the last couple of months figure out you know we're growing 30 40% every month you just you know take you know take the power of three that's how much you're going to need in 3 months that's how many GPUs you go out and get those GPUs

you can typically get GPUs with like about three years of sorry 3 months in advance

sure

uh but yeah that's a big part of how we think about the business today is like we need to get the GPUs now that we're going to need in 3 to 6 months which is a lot

yeah in some ways it's like you know not not too dissimilar to the challenges that uh you know consumer consumer packaged goods founder is facing. you're growing super quickly, but if you want to stay on that growth curve, it means you need to be committing to things, you know, now and and uh committing capital. Uh but but

what's exciting to you across

luckily we have capital now.

Yeah. Well, what what's exciting to you across the uh like semiconductor e ecosystem? There's a whole bunch of AS6 startups that have been on the show. Cerebrus IPOed. Uh every hyperscaler is working on uh different chips at this point. um what's most interesting, what's under discussed, what's on your road map or or or uh what have you already sort of uh sunk your teeth into?

Yeah, I mean there's some like really cool alternative accelerators. I'm quite bullish on on TPUs, AMD, Tranium, like all these things.

We see zero demand from our customers for for any of those things to be clear

because only the really big labs anthropic can actually go and and figure out how to run it. Is that

exactly? Yeah. I I think the cost today of rewriting your software to run on those stacks is just like very high. Sure.

So while I remain like very bullish on the sort of you know two three of Horizon Yeah.

You know I do also want to like temper the expectation a little bit like this is

it sort of has a fixed cost to rewriting for that chip stack. And if you're not doing billions a month in revenue it's hard to advertise that cost.

That's exactly right. It's just not worth it unless you're you know operating at a very large scale. So but but I think that that cost is going to go down over time. We're gonna have software that basically lets you take existing CUDA compatible stuff and run it on other alternative accelerators. So yeah, I think it would be good for everyone to have a little bit more competition in the space, but we also love Nvidia and and that's what our customers want today.

That's great. Uh do you have a view or any opinions or predictions about uh the compute futures market that's been talked about uh any I I've seen like these price charts of like B200's going up and down. everyone is trying to read the tea leaves, understand, you know, where we are in the various AI cycles uh based on it. But, uh, is that something that's relevant or or important at all? Do you do you look at that data? Do you have your own internal data set that's more relevant to you?

We we we look at it a lot. I mean, also like I feel like we're very plugged into the market. We talk to NeoClass all the time. We get capacity all over the place. So, we have like a very good like pulse of the market. I I think the market's going to remain quite tight.

Yeah. I I think fundamentally like also like that's big part of what we do is like we offer you know that as a product like don't think about capacity come to us instead.

Yeah.

And so that becomes now our problem is like managing that capacity for thousands of companies at the same time like we built a multi-tenant product to sort of aggregates all that demand. So if you need a thousand GPUs you can come to us and we'll give you to that give you those GPUs like often within minutes because we have like a very big pool we can sort of tap from.

Yeah.

But but yeah like I I you know I think GPUs are going to be tight. You look at the prices, they keep going up. At some point, obviously, I think it's going to normalize like most markets do, but it might remain tight for the next year or two.

When I think about the big applications of uh the big pool of compute, I think uh training and inference of LLMs, coding models, agentic workflows, I think image, video, audio generation. Uh what's next? What do you think the next uh big driver of demand is? Is it world models? Do those fit in? Are those structurally different? Is there something else that you're tracking where you're seeing customers come to you with sort of like a a different product but it it fits in the same shape so you can work with them as a business partner?

Yeah, we have a very wide range of customers. So we have everything from Sunno which generates music using AI cognition is training uh coding models using reinforce reinforcement learning.

RAMP uses us for background agents. We have a lot of vibe coding platforms like lovable. We also have you know drug discovery companies like Chai using us to simulate you know molecular dynamics and we have weather forecasting companies, robotics. So so so we we we have you know very general you know we look at across like a lot of different verticals. Um obviously like the the the big application in the last couple years has been LM inference and that keeps growing a lot but we've also seen some really cool applications with the fusion models. In terms of what I look forward to, like I I think the probably personally the coolest thing I kind of expect to happen in the next couple years is speech to speech, like imagine just like talking to a computer. In order to to figure that out, we need to get the latency down, you know, for for all the three components of speech to speech, you know, doing the LM and the, you know, text to speech and stuff like that. So, that's something very cool, but I'm also really cool about I I think it's incredibly cool if we can, I don't know, cure cancer or something like that.

Yeah. Yeah. It's interesting because uh I'm I'm with you on that on that timeline. Speech to speech being uh you know somewhat impressive, but you're often not actually hitting a real reasoning model. There's so much speed up. You know, we could be we probably need to be 100x faster, a thousandx faster on the response time to really have a breakout moment. And then you watch Google IO this week and you see with Omni a glimpse of okay, it's more like a FaceTime call with the AI and it's generating a video in real time and you know how long it takes to crank on video models to get an 8-second clip. It's minutes and minutes uh and and and they're still have errors and stuff and we're still not even at the the super the superhuman element. So lots to do, lots of servers to rack, lots of GPUs to set on fire, I'm sure. Yeah. Would at what use?

Yeah.

Any any plans to actually, you know, go full stack, get your own, you know, land powered shells, etc. Is that like not the highest and best use of your guys's, you know, talents?

We we think of our values like adding a big software layer on top of the underlying computer on the underlying comput layer. So, we think of ourselves as almost like a cloud like one layer up from the existing clouds. The existing clouds are very good at running computers, you know, in the cloud and and offering that through an API. We we would love to keep tapping that as much as we can. If we can't get the capacity we need, we might have to build it ourselves. Ultimately, it's like, you know, it's a kind of a economics question or, you know, practical question. I I wouldn't rule out any sort of move. uh the idea of, you know, racking computers and plugging in cables and dealing with, you know, fires and