Fal raises $125M Series C and crosses $100M ARR on AI-generated media platform

products. Brainstorm, design, and build with your team at figma. com. And we have our next guest in the studio. Looking sharp. Looking sharp. How are you doing? We caught up yesterday, said he was going to come prepared, but you're you're blowing us away here. You look fantastic. Uh, give us an introduction.

Give it explain the company. Give us any news that you got to share with us. Of course. Uh my name is Garam. I'm one of the co-founders of FAL. Uh FAL is a generative media platform for developers. We provide fast readytouse APIs for image, video and audio models.

And then application developers in enterprises like Shopify, Canva, Adobe, they they use these APIs to build AI enabled products in their in their own product. And then I do have some news as well. Some news. So two weeks ago, we announced our CC announcement. to be raised $125 million.

It's not announced until it's announced on TV. We just broke the news. We just broke the news. I have other news as well. So, this is great. What else you got for us? Hit that soundboard, Jordy. And and this month, more news coming in. Okay, give it to us.

cross 100 million run ra we have 10 more days but yet we still we still crossed it so that's crazy it's been an incredible summer for us and the whole industry to be honest yeah congratulations that is absolutely insane amazing amazing uh what what uh what's the key to the high growth uh what marketing channels are working uh do you have are you are you buying an incredible amount of steak dinners is there some bottom up adoption is self-s serve option.

Uh are you are you secretly brothers with a hyperscaler CEO or something? What what's going on? How you how you selling this thing? Yeah, I would say all of the all of them. You are brothers with Tim Cook. Interesting. Except that one. Uh that's great.

Uh but yeah, we are taking enterprise sales really, really seriously. We work with some of the biggest enterprises, but also like our developer experience, the way to get started is so easy. We are getting a ton of like indie developers. Level zio for example is is a big customer and you know amazing.

So we are getting a lot of enterprises and indie developers and indie developers they go work at bigger companies and they bring fun with them. So both of the sides of the goto market strategy is working. So what trends what trends are you seeing on the uh on the cost side for your customer?

I'm sure people are saying, "Oh, yeah, you're making this much money in revenue. What does the gross margin look like? " We saw amazing news from Google today that they dropped their inference cost by 33x. Yeah.

And the other thing on the media side with with with textbased prompting, it's it you can use a a non-frontier model and get a great result if you're just using it in sort of like a Google search function. I would consider last generation video generation to not be on brand for a lot of use cases.

Um, so you got so we we see you're absolutely right. Some people they want to use the best newest model and they're willing to pay whatever it takes for that. But we we make more money actually on like the cost effective models, people who are who want to use it for higher volume but the model is good enough.

So that that's a very interesting sweet spot that we actually make more revenue. But you're right for for a certain type of customer, they they're always looking for the best and the newest. Maybe they are trying something. Maybe they just want to get it out the door so the the cost is not a problem for them.

Are you seeing opportunities on distilling models for specific use cases? I could imagine um like a an image or video generation model that's just good at really really good at logo animations or text or just product photography or just beverages and then that being just cheaper to inference. Is that a crazy idea?

Is there something there? No, I don't think so. I I wouldn't say it's cheaper to inference, but it's maybe better to inference. So for image models, fine-tuning is actually a really really big use case.

not enough people pay attention to it because I don't think it works very well for LLMs but for image model for the past year. Um we've done like really good work on fine-tuning research and you know you can basically fine-tune on a product and you now can generate better images of that product.

It's actually a really big use case for us and same same thing is going to happen for video. Uh we we are just getting open-source video models that are fine-tunable.

Uh so a lot of companies are working on post training strategies to create different camera camera angles, different styles, you know, different angles of a certain product or a person. Uh so I expect a lot of developments on the video fine-tuning side going forward as well.

What are you seeing out of ch what are what are the best uh image video uh audio models coming out of China on the open source side? Yeah, video hasn't really had a deepseek moment, have they? I wouldn't say so. Like there is the best open source model is is video model is Chinese right now.

It's it's coming out of Alibaba. It just came out couple weeks ago and people started using it right away. I wouldn't call it the deepseek moment. I don't think the deepseek moment for the whole industry hasn't happened yet.

um you know we haven't really hit that inflection point that video generation or image generation is mainstream it's it's hidden inside products if you go to Canva Adobe all these products it's still not the main feature it's like an additional thing we still have some time to hit fully mainstream but yeah we have great models coming out of China and we have and also just for I mean this is the classic story of enterprise software like there are open source databases And there are still companies that build amazing businesses on top of them providing enterprise tooling and and support and infrastructure and all the different pieces to actually make a product useful instead of saying, "Okay, yeah, let's go provision some servers to do this ourselves.

" Uh, fascinating. Um, George, do you have another question right now? I I No, for now. I'm just blown away you guys are cooking. Yeah, I I do I do want to talk about uh hybrid deals or hybrid models.

I somebody had me demo a AI image generator that uh I I had it take a picture of me turn me into a professional bodybuilder.

It was the most remarkable thing because typically when you go to CHP and have that do that it kind of gets the face a little bit wrong but very clearly and I could tell this from the grain pattern. What it had done is it had cut out my face and then only operated on the body.

And so the face was a pixelto pixel perfect representation. And for a certain thing, oh, you just want to change the logo on the shirt. You don't want to change anything else about the shirt. It seemed like a good value. Now, I was being gaslit by the founder who's like, "No, we're not doing that.

It's all in one network. " And I was like, I don't care. I just want good images. It's cool. Um but but where do you see the different like value added like maybe less sexy less pure endto-end AI but value creative opportunities in in image and video generation.

Um I could imagine like a multi-layer diffusion model that has a specific a specific layer just to add text on top of images. So you're not trying to confuse those.

I don't know if that's actually how these models work, but what do you see as as valuable in terms of like merging multiple models to create a better pro end product like the way you know 03 Pro has a web browser, a Python shell, and then also the ability to search the web and and pull stuff together and reason. Yeah.

I mean one of the reasons why uh you know our business has higher margins than some LLM insurance providers is because of this this fragmentation of different kinds of models different kinds of use cases. People are trying to chain these models together and create workflows.

Um both image and video models they haven't really generalized yet. So you can't write a long prompt and expect uh you know the the image and video to edit itself. It's still a lot of model chaining and workflow generation. So like the use case you you just said like putting your face onto something else.

It's like chaining couple models and we have a lot of uh that behavior on the platform. The models are getting bigger and better.

So some of that is actually happening you know by by the model itself but but still unlike LM where it's like a single prompt and you expect it to you know write a book but also generate code and like generalizes very well it doesn't it doesn't happen for image on video yet.

Yeah I can imagine that there's tons of interesting enterprise like features and harnesses on top of the models that will be extremely value creative. So, seems like you're in the good in a good business. But, uh, everyone already knows that from the revenue and the fundraising. So, congrats on all the progress, Jordy.

Anything else? We'll have you back on again soon. We'd love to have you back. This is fantastic. We'll talk to you soon. Legend. Bye. Let me tell you about Julius, the AI data analyst that works for you. Connect your data, ask questions in plain English, and get insights in seconds. No coding required. Sorry, Tyler.

You're out of the job. I'm going to be using Julius. Try Julius for free. Uh they're loved by over two million users and trusted by individuals at Princeton. And you can connect Google your Google