Burkay Gur on Why Generative Media Will Be Bigger Than LLMs | 2026 Upfront Summit

0:01 It's been two years since we first invested in the company. Yes. It's been a wild ride. I'm going to give a a quick snapshot on the company and I I want you to kind of walk us through what generative media is and and what Val is doing these days which I think is super super interesting especially for the LA community uh as well as um all the industry here. Um so number one Val has created uh a foundational platform for developers, creatives, enterprises in generative video, images, AI voice. Um but interestingly 2 years ago when you launched the the the product and service, you've now grown to uh uh you've just passed 300 million AR.

0:41 >> Yes. >> Uh in two years. So you you've beaten um our our previous fastest growing first two years company was Uber. So you've beaten Uber on the first two years of revenue growth. So kudos to you guys. >> Um uh we we'll talk about GPUs and what that takes to grow in AI. Um but maybe if you can uh walk us through like the original story, the the inception story of where F came from and your vision behind it.

1:06 >> Sounds good. >> And then a little bit about generative media. >> Awesome. Uh nice to see you all. Uh I'm Burkai, co-founder CEO at F. Um, and yeah, two years ago, uh, when I was pitching to Steve, uh, this category did not exist. There was no such thing as generative media. Um, I remember me and my co-founder actually in a room talking about like what should we call this space and we came up with generative media and we've been calling it that and so as many others in the industry now that are building these these models. uh but basically to to summarize generative media is um all the AI models that are capable of generating uh content. So this can be uh generating images, videos, uh audio, 3D and and all sorts of like content type of assets. So you know when when we had the chat GPT moment uh obviously was all about LLMs but just just side by side there was another big uh technological breakthrough was happening which was these models also turned out to be able to generate content and the it was actually pretty pretty much a coincidence like maybe because there there are really two different technologies like under the hood the that image generating models and text generating models are completely different and they just happened to happen at the same time. Um and back then all the attention was going to LLM right so these are like model companies infrastructure companies and we found this like very interesting niche we thought that uh this space was going to grow a lot and that's where we decided to double down and and here we are uh 3 years later it looks like everyone's everyone's aligned with that with that same idea and we're seeing a lot of adoption >> I think when we first met there were something like uh 15 20 models and they were all open source at the time.

3:09 >> Mhm. >> Uh about 2 years ago. How many models are on FAL today? >> Yeah. Yeah. When we first met, uh there was like maybe one or two models as base models and then there was like custom fine-tunes um and other customizations people have built on top. Right now we actually have thousand different models on the platform. And I want to highlight that like this is very different than language models. in language models there's like um there's obviously like a very rich uh open source ecosystem uh but there is like these like large labs right um and they mostly dominate and they suck up all the air so uh and then the the diversity in the models are also not much right so there's like search and personal assistant type of models uh there's coding models um and some of these are in one model some of these exist as separate models but in the content realm uh there is hundreds of models there's there's hundreds of companies thousands of models and and different fine tunes. So that actually as a platform where we we host all of these models like it it is it is very much like it's it's our reason to exist uh is that is this um diversity in the in the landscape and yeah every week we're adding like close to 10 10 new models and these are from different labs or different uh workflows people are building and different finetunes people are building and you know by the end of the year we'll probably have 2,000 models >> and what's what's really you know those early days it was all sort of like open source models and maybe some small research labs but today it's uh uh everything from deep mind to open AI >> to um uh bite dance >> and uh a long tale of venturebacked research labs now providing models for the file platform. Yes. um how many developers and enterprises and designers are are using the foul platform for their work?

5:11 >> Right. Um so we've we've had like 3 million signups uh since since we started. Um our main focus is enterprises. So most of our revenue actually comes from you know the top thousand customers. Um [snorts] but there's a lot of value also that that this long tail is getting. But the way that like we actually engage with enterprises is usually it's like one or two people from this company. So they will actually start playing around and then they will slowly find their use case and then let the rest of their team know. So so we have this like product that growth. So for us like this top of funnel is very important but ultimately we make we make money from from enterprises.

5:55 >> So it's one part is a marketplace of models and model labs that create them. Then on the other side, designers, developers, enterprises, all of that. Um, what are some of the the top use cases that you're seeing out there that are happening in, you know, enterprise verticals or or consumer? >> I think the most obvious one and the one that's growing the fastest is marketing and advertisement. So these think like product photography, right? So when you go shopping to any any kind of e-commerce website, uh those images are now more and more edited or generated by AI um to to make them look look a lot better. There's also more engaging content like uh AI generated video with those products. So there's there's a very big category uh it's probably one of the fastest growing um and and then I would say following that is the creative like industries like entertainment. you know, these are like shorts or or even like short series. Some people are even trying to build um long form like movies and and we're we're seeing more and more of this and this year um there's been a lot of breakthroughs in like the quality of video models as well. So like we're going to see a lot more, right? And we're going to see it in mainstream platforms, right? like like Netflix and and you know on YouTube this is going to be pretty pretty normal and there's going to be mainstream adoption.

7:20 >> You were telling me a little bit about um some of the retail uh customers and what they're doing. Uh if you could share that anecdote uh that was really interesting, >> right? Um yeah, so retail has like two use cases. One is on the e-commerce side, so these are like product photograph product photographs and the other one is on the marketing side. And we we've actually been invited to one of these like large studios where they're doing a lot of um you know manual work and these are like very professional setups with like very expensive gear and you know for like a e-commerce platform retail platform that has a lot of SKs it is like a extremely difficult um uh you know manual work to to get all of this done. So, so yeah, and like you know, we've been working with these kind of companies for the last year and a half and the how fast the the uh space is evolving is it's it's kind of ridiculous. Um, we went from like, hey, let's try to optimize some parts of their workflows to like let's take some photos with an iPhone and we can just like AI edit it to look extremely professional. So, that's that's like how far this space has grown. I think all of the retail companies are actually aware.

8:37 You may not hear this from the outside, but like this this is actually happening very very rapidly. >> Well, let's talk a little bit about the the adoption. So I know you can't name names on someone, but you can describe who they are. Um uh uh I think in the last board meeting you said uh 60 of the Fortune 100 uh companies out there are are users and customers of FAL. Yeah. >> Can you talk can you maybe what what names you can mention? Can you share some of those and how they're using it?

9:06 Yeah, I think um the biggest use case that's like widely deployed right now and it's not even to its fullest fullest extent is u in in marketing >> and um we're basically seeing like really large retail platforms um essentially able to automate all of their like product photography and ad creation and deploy this to uh very very like personalized ads. So, you know, the old world, you generate one ad, you spend like $100,000 on it, and you show it to >> And these are video ads, commercials on TV.

9:45 >> Yes. >> Um, >> I would say mostly online. >> I think like it hasn't really broken into the like TV yet, but most of the, you know, most of the volume happens on online anyways. >> So, the major streamers and the movie studios that are are using FAL, >> um, they're changing their entire workflow. Yep. using file. What are some of the models? Um like generative video is just like in the last 6 months has become production grade, >> right?

10:10 >> Um what are some of the cool things and the more like innovative things that are happening on the video side with some of these enterprises? >> Yeah. Um I mean as much as we have visibility into uh they are able to now go from like again um you know if they have merchants that are not very sophisticated, they don't have this gear, they can actually go from just like a few product shots into being able to change the lighting, change like um you know different um elements, let's say if they're doing like some localized sales, right? So they are able to actually go from and and they can generate uh assets in the thousands or tens of thousands and basically what they can do is like fully personalize um to the audience essentially. And where where this is going like with the speed that this technology is growing and how much cheaper it's getting where it's actually going is like fully personalized like onetoone ads and onetoone like shopping experience. So, anytime you're you're doing anything like commerce related online, um you know, in a few years, you're going to basically experience a world where it's very specific to you. Maybe it has to do with, you know, the things you like to do or like place extremely personalized, >> full of memory. Yeah.

11:31 >> Yeah. So, so it'll have context on like, you know, the next trip you're that's coming up, right? And uh they can show you in in certain uh you know the the the clothing you want to buy etc. Like what basically the whole experience uh you would go and have in like a real real shopping experience. Uh you can actually have that. >> I've seen some really cool some of your customers uh have some really cool virtual fashion tryon applications that they're doing and I used it. It it made me look skinnier and more fit which is kind of interesting and I didn't program that in. Um, so you guys are working some magic on the back end. But, um, one of the things that I I think is also super interesting about what's happening is that, you know, LLMs are are obviously like, you know, agents and harnesses and um, everything that's happening with chat GBT and things of the like text is very very important for instructions. It's very important for any type of copy. But um you know if you look at YouTube, Netflix, Tik Tok, uh Spotify uh you know media in terms of audio, video, voices, images is much more captivating, compelling for like mainstream human beings. Like we don't have much time I wish I could read more books >> um but we spend a lot of time um watching videos, listen to podcasts, things like that. There's a mind share >> that generative media um is expected to have. Can you talk a little bit about like what you see uh five 10 years out with generative >> media I would actually I mean this is very self- serving but I would actually make the claim that uh generative media is going to be bigger than language models in terms of especially in terms of like what we're consuming dayto-day >> I actually think that LLMs are going to be mostly a lot of LLM like executions are actually going to happen between like machine to machine right like agents orchestrating things systems talking to each other, problem solving instructions, whereas like the >> the content realm is very much like >> you know >> the the end user is humans, >> right? So and and if you look at the internet and like the consumption we have today literally like I don't know 95% of what we consume on the internet is visual. Um maybe 5% is text. Uh so just just just going by that um chances are like you know most of the compute and and also video and content generation is most more uh compute intensive.

14:00 >> So just just basically looking at all that uh chances are you know we're going to be spending most of compute on on being able to generate content. >> Let let's talk about computer. That's a good segue. There's a argument out there in financial press that we're in AI infrastructure bubble, >> right? >> Um and and we all I think we all sort of know in in Silicon Valley with a lot of the the inference platforms and the neoclouds >> that demand is outstripping supply, right? Nvidia cannot ship enough um uh high-grade GPUs um to fulfill that demand. These data centers, these Neoclouds, the clouds. What is the difference between compute for LLMs versus compute for video and images and things like that? Because it's different, >> right? Yeah, it is different. Um, you know, not getting too too deep into the technical details, but uh LLMs in general are uh memory bound. So there's memory and there's compute. So so LLMs are memory bound and video models in particular are computebound. Um, and because LLM market is bigger right now, a lot of the chip manufacturers are actually optimizing their chips to fit with, you know, they're kind of overfitting to LLM use cases, right?

15:15 Obviously like the the chip needs to be built uh somewhat generic enough that it can still do a lot of other work but you know from generation to generation uh you're seeing more improvement on uh memory memory bandwidth than actually like pure flops which is the which is compute. So with uh video models, you typically are wasting all that additional memory, right? Because the compute hasn't gotten that much faster.

15:46 So um so that's the main difference. I think the market is kind of like optimizing for LLMs right now, but what we do at F is like we actually optimize these workloads to to run much faster. So like we we're kind of squeezing the most juice out of the the GPUs. Another point here to to talk about like the compute shortage or or are we overbuilding, are we underbuilding? What we experience is there is always shortage. Like there's never been a time we we haven't gone um and you know beg for compute from comput providers.

16:23 >> I mean we were just on the phone last couple days. >> Yes. >> Haggling someone over a a big GPU deal at a data center. >> Yeah. So it's it's um >> and and like particularly I think um like we we have been mostly um buying from hyperscalers and neoclouds right that's that's that's been our main uh motion of how we procure GPUs but because of the shortage I think if you reach a certain scale like us and like many other companies you actually want to have more control over your supply so that's that's another that's another reason why like there's also so much buildout.

17:03 >> Well, control and cost. Yes. >> Right. Long-term like amortized cost on it as well. >> There there is even times even right now where like there's so much shortage that like you're you're like okay I want to get 10,000 GPUs and I would pay like you know 50% more and you just can't find it. There is literally no way you can find it right now like today. Um what you can do is you can place an order.

17:27 Maybe it'll it'll be in like 8 months, but um it is the the the supply itself is is very short and you know you you you all probably experience it too like you know going to chatt and it's slow or clog and it's slow. So it it is actually that's because of shortage. It's not actually they're having a systemwide outage. It's literally because of shortage. Well, since we um since we let Nvidia invest in the last round, we we need to call Uncle Jensen for um a little bit more uh chip expedition.

17:58 >> Yes, absolutely. >> Um so there's there's a there's a thing in compute 2 that u we've talked about. We did a little bit of an analysis on uh compute of video models versus LLMs. And one of the things that was incredible that came out of that was that if you wanted to generate a novel using an LLM, the amount of compute GPU compute to generate a 300page novel from scratch from a series of detailed text prompts was uh 150th of the compute that required to create a 3minut sorry a one minute video.

18:36 >> Yeah. >> A one minute video generated was 50x the GPU compute. >> Yeah. And >> I think I think that sounds mind-blowing. >> Yeah. Yeah. Yeah. So so like you can think of a video as like tokens. So like each pixel roughly I'm I'm oversimplifying but like those are all like words. So it is there's you know in a single video including like you know the frames uh over time there's more tokens than than a novel. possibly if you straight line out both uh prediction curves on on both LLM's AI in the LLM world and AI in the generative media world. Uh it says that generative media is going to be much more of a a consumption of GPUs and data centers and energy than uh than LMS. How do you think about like where this goes? I mean this is one part a a cost issue, an energy issue, data center compute issue, but it's also a fidelity issue. like how do you think about foul growing a thousandx from here? What what are the things that need to be true?

19:39 >> Right? I think first and foremost like the demand needs to be there, right? Which which we're experiencing like unprecedented demand. Um so and and use cases need to uh sort of scale, right? So like the personalized ads use cases that that I I was talking about, those need to scale and they are on their way too. So I I think that's that's pretty much solved in in like the next 5 years. And then I would say compute is the second biggest one. Um and for that I mean what we're doing and I think what a lot of companies are doing is like trying to secure our own um capacity right um and that basically means that uh you know you you build great partnerships with like the supply chain uh you build great relationships with like chip manufacturers and and try to kind of stay ahead of demand and you know at our scale um you know we can we can we can sort of invest we and kind of ma make this investment into into having some little bit of extra capacity. Uh I think early days like we were a lot scrappier. Um you know we were like all right you know let's let's try to optimize for uh you know the best price and whatever but now it's so much more about like controlling your own supply and and having some predictability. So th those are the types of things we're doing. Um you know we we plan to actually have like some of our own own GPUs. um possibly diversify also between different chips. Uh our team like tries to experiment with uh multiple multiple chips so so that we can have a little bit more control.

21:17 >> Yeah, there's there's a lot to unpack there with like transformer architectures versus diffusion. Um but the uh the amount of media that's already being generated today um is is incredible to see. It's this is really just in the last year, right, that um that's happened. I mean just um our platform alone and we we are like a pretty pretty big market leader here but our platform alone I think last month we've generated more than 1 billion like pieces of content I think uh this is like image video audio including um I I actually looked this up uh I was like how many ad impressions does like all of humans see and it's in the trillions a day right so That's basically uh I mean that just tells you the scale of like how much more I think it's in like couple trillions like numbers like maybe four trillion or something. So you know couple billion a month to 4 trillion. Uh this is a day.

22:20 >> So a day is is a is a big jump right. So we have like a thousandx ahead of us which is very exciting. >> It's incredible. Um, okay. We're in LA, capital of media, entertainment, fashion, sports, gaming, all of it. Um, uh, are humans getting replaced by generative media? >> You know, in in the public markets right now, there's this whole argument that like all these legacy SAS companies >> are going to get decimated by the LLM research labs and their products, not just their models, but the products that they're building. Yeah.

22:53 >> Um uh what do you think happens with generative media and all these traditional industries? >> I think the nature of the work changes. Um I think like you know the manual work, the the camera, the lighting like a lot of that you can actually you know prototype that kind of stuff a lot faster so you can iterate faster. I think um what this does is actually like it evens the playing field. So like you know a one person studio from from their bedroom can actually like will be able to build We're already seeing that, right? We are seeing that >> like amazing films.

23:26 >> Yeah. So, so the nature of the work and the economics are going to change. So, there's going to be a lot of change that's going to happen. But I think we still need the creativity. Like what when I see the the outputs like you know some of our like studio partners generate like they're just there's sort of like this AI you know video enthusiasts uh with no let's say like you know film background like the the the the quality of output they generate and like someone that's been in this industry for 30 years it's it's like miles away. You can still tell like the craft and the expertise, >> the directors, producers, good writers.

24:02 >> They're still going to produce the best stuff. >> They're still steering the generative media models and the workflow that you guys are creating for them. >> And I I prefer their output to to like, you know, someone who who's done this only for a few years. That's I mean that's this goes back to like the music industry >> their push back on the technology wave of of uh MP3s and streaming >> and before that CG and the entertainment industry.

24:27 >> I think we we've seen this so many times and you know it's always the same story. There's a bit of resistance but then there's a lot of change. >> The ones who embrace it, invest in it succeed. >> Brookai, thank you so much for doing this. Um you even dropped some numbers that are confidential so that's great. Thank you for letting me do that. That's great. >> Thank you. Thank you. >> Thanks all. [applause]