So, I'm Erik. I'm local here, based here in New York. Lived there for 15 years. Built technology at various companies. I started a company called Modal about 5 years ago. We do AI infrastructure, is the best way to put it. More broadly, really compute anything, compute infrastructure. So, we do probably a lot of use cases in AI, data, machine learning, computational biotech.
We have customers like Suno for AI-generated music, Cognition that trains models using reinforcement learning for code generation. We have customers like Ramp, Harvey, robotics companies, mini bio companies like Chai, Akira. So, that. So, we're about 100, 120 people, which is close to C round. We announced it yesterday, led by General Catalyst, Redpoint. Been scaling very fast. There's a lot of demand for AI infrastructure today. And so, why did I start the company?
By the way, with Akshat, who's a co-founder, on the previous slide. Basically, what I realized about 5 years ago was that today's infrastructure, and I'm really looking at the clouds, is just not really built for AI applications, right? And so, part of that was inspired by my time at Spotify. I built a music recommendation system at Spotify. This is more than 10 years ago. I ended up spending 90% of my time building the infrastructure rather than the application logic. And so, you know, like, how do you deploy things to production? It turns out it's kind of annoying to work with existing infrastructure. You have to think about containerization. You have to think about getting the GPUs. Like, where do you get them? In many cases, you have to go reserve a big block of GPUs. Maybe you don't want to do that. Maybe you don't know how many GPUs you need. You have to write a bunch of YAML. It's really not good. And it really makes it slow to iterate. So, one of the things I wanted to build was basically a tool I wanted to use to build complex AI applications. And we found really good inference, really good product-market fit in particular with inference. And I think a little bit like why today's infrastructure doesn't really work is like all this stuff with GPUs and working with large models and GPU shortage and GPU capacity issues. Like, this made a lot of traditional infrastructure just kind of break. It's kind of annoying to work with.
But when you're launching these models, you have like very unpredictable workloads. You need to deal with GPUs all over the world. And you need to iterate fast. So, we sort of realized, what if you just rethink all of that? What if you just kind of build a software layer that runs on top of all the clouds? We manage all the capacity for you. Don't think about capacity. We kind of take care of it. If you need a thousand GPUs, we'll give you a thousand GPUs. And then, you know, then it becomes our problem. But from a customer's point of view, they can iterate fast. They can deploy models and not really think about the underlying infrastructure that often gets in the way of delivering value.
Actually, how many people here write software? Just curious. Okay, so like a third. How many people have deployed models on GPUs? Okay, so kind of a quarter. Okay. Just curious. So, Modal, we focus on people who are building these models primarily. Like, we focus on researchers. We're focusing on people who either deploy models. Now we have a new product around sandboxes. So, part of it is inference, part of it is sandboxes. Which basically the idea is like you have a lot of LLMs that generate code today and you want to execute that code in an isolated way. So, we have a product for that as well. We also have a product for training. So, the idea is to provide this, like, work with software engineers and build a tool that they want to use to iterate with all these things. And so, whether that's training or sandboxes, we have a lot of people using us for batch jobs, like scaling up things and running at 1,000 GPUs, or creating these sandboxes for safe code execution.
We have a lot of customers running millions of sandboxes at a very large scale for reinforcement learning or for background agents, or for vibe coding platforms. We have a couple of vibe coding platforms like Lovable and another one is using us at a pretty large scale. I think at the limit, like, what we end up thinking about it as, like, part of the product is packaging all of that in a great developer experience to make engineers and researchers iterate very fast and write code and not have to think about infrastructure.
So, that's the sort of developer experience. And part of, you know, what I realized, like, optimizing for the developer experience is actually a really hard thing to solve. Like, we ended up building our own file system very early. We ended up building our own container runtime. Like, we ended up having to build a lot of very deep infrastructure in order to support this beautiful developer experience, which today's cloud offerings don't really do well. Like, it's kind of annoying to work with. I mean, actually, I love AWS. I used it for 15 years. I love it for all that it enables me to build. It just never had good developer experience and it's very painful to work with. And so, we kind of wanted to rethink. And by doing that, we had to build all these things. Like, we basically threw out Kubernetes, we threw out Docker, built our own scheduler, our own container runtime, our own file system.
The other way to think about Modal is we end up, in the limit, sort of aggregating the world's computers in certain ways because we run on 15 different cloud vendors today, probably more in the future, hundreds of regions. We can aggregate all that capacity. And the beautiful thing is that resource pooling is kind of a free lunch. If you take a lot of customers, we have a lot of customers running on many thousands of GPUs, you know, thousands of customers. If you aggregate all those customers, even though they all had sort of unpredictable workloads, you know, you end up like, basically the variance goes down with a sort of inverse square root, or not quite because there's a little bit of a skew in the distribution, but the general idea is that the more you pool, the more smooth, the more predictable the overall workloads are. And a big problem today is that GPUs are just not very highly utilized because it's very hard to predict workloads. And it's pretty hard to manage capacity.
But because we can, like, you know, aggregate all that demand and manage that in aggregate, we can actually drive much higher efficiency, too. So, part of how I think about Modal is we're sort of becoming this very large compute demand aggregator that can sort of then go and find capacity on the other side and work with a lot of neoclouds and hyperscalers and make it easy to take our customers' code and run it in the cloud, whether that's for inference or training or sandboxes or other things. So, we're basically managing compute all over the world. I mean, this is just like one screenshot from a dashboard, but like, we built this quite complex system to... It actually ends up being a mixed integer programming problem to solve this bin packing problem of taking all these tasks and placing them on the worker. We do that multiple times per minute and manage that compute in aggregate. The objective is basically to minimize cost on our side. Customer pays a fixed fee. And it ends up working really well.
So, today, kind of going back to what I said, there's like two ways to think about Modal. One is the sort of capacity pool. We aggregate all this capacity and manage stuff for customers. The other way to think about it is like we make it really easy to write code and iterate very quickly. So, it's basically to supercharge developers and researchers and make it possible to build way more complex AI applications. And ultimately, that's like kind of the goal here. It's like all these companies out there building cool AI applications, they shouldn't have to think about infrastructure. So, we manage that infrastructure for you.
So, today, a lot of GenAI inference, video, audio. We have customers like Runway here in New York doing generative video. We have a lot of customers running big batch jobs. LLM inference, of course, is a very large use case for us. Forgot to put Sandboxes on the slide here, but that's actually a third of our revenue, growing very fast. A lot of reinforcement learning, background tasks, vibe coding platforms, etc. So, we've always had this ambition of building this very general-purpose platform you can use for a very large amount of different things. So, almost like building kind of a new cloud, actually, but like a layer up where developer experience is better, you don't have to think about capacity. So, here's just some examples of customers. Suno, Runway, and Lovable. Many other customers. Meta is using us for reinforcement learning. Example is here locally. DoorDash uses us for LLM inference. These are just eight out of thousands of customers.
And why do you use Modal? Typically, it's builders want to run custom stuff with their own code. Modal is high-code. We're not trying to simplify behind a UI. We work with software engineers who want to deploy things or scale things up. We want people who want to iterate very quickly. A very big value prop with Modal is because we aggregate all the capacity and demand and manage that very efficiently, we can charge people based on the time the computer is running. So, if you're running, you know, scale up to 1,000 GPUs, scale down to zero, you only pay with Modal for the time the GPUs are actually running, or CPUs. Which is a very nice value proposition, as you don't have to think about capacity planning. You don't have to go out and reserve like 3 years of GPUs. You can sort of just pay as you go. It's a very flexible capacity model.
How do you try out Modal? It's very easy. If you're a software engineer, pip install Modal and you're trying Modal, it'll install it, get you up and running. Typically, it takes a couple of minutes to run some Hello World and get a feel for it. Obviously, if you can build very complex things, it takes a little bit longer. And that's basically it. If you want to stay in touch, I'm on Twitter, or X, I should say. Feel free to ask me. I'll be around for another hour or so. Thank you.
