Gorkem Yurtseven is the co-founder and CTO of fal, the generative media platform powering the next wave of image, video, and audio applications. In less than two years, fal has scaled from $2M to over $100M in ARR, serving over 2 million developers and more than 300 enterprises, including Adobe, Canva, and Shopify. In this conversation, Gorkem shares the inside story of fal's pivot into explosive growth, the technical and cultural philosophies driving its success, and his predictions for the future of AI-generated media.
In today's episode, we discuss:
- How fal pivoted from data infrastructure to generative inference
- fal’s explosive year and how they scaled
- Why "generative media" is a greenfield new market
- fal's unique hiring philosophy and lean <50-person team
- Building a brand that resonates with developers
- What the world looks like in 2027 when AI-generated video becomes mainstream
- And much more…
Where to find Gorkem:
Where to find Todd:
Where to find First Round Capital:
- Website: https://firstround.com/
- First Round Review: https://review.firstround.com/
- Twitter/X: https://twitter.com/firstround
- YouTube: https://www.youtube.com/@FirstRoundCapital
- This podcast on all platforms: https://review.firstround.com/podcast
References:
- Adobe: https://www.adobe.com/
- Amazon: https://www.amazon.com/
- Anthropic: https://www.anthropic.com/
- Base10: https://base10.vc/
- Black Forest Labs: https://blackforestlabs.ai/
- Burkay Gur: https://www.linkedin.com/in/burkaygur/
- Canva: https://www.canva.com/
- Clay: https://www.clay.com/
- Coinbase: https://www.coinbase.com/
- Cursor: https://www.cursor.com/
- DALL-E: https://openai.com/dall-e-2
- Databricks: https://www.databricks.com/
- Dylan Patel: https://www.linkedin.com/in/dylanpatelsa/
- fal: https://fal.ai/
- Google DeepMind: https://deepmind.google/
- LLaMA: https://ai.meta.com/llama/
- OpenAI: https://openai.com/
- Oracle: https://www.oracle.com/
- Perplexity: https://www.perplexity.ai/
- Shopify: https://www.shopify.com/
- Snowflake: https://www.snowflake.com/
- Sora: https://openai.com/sora
- Stable Diffusion XL (SDXL): https://stability.ai/stable-diffusion
- Stability AI: https://stability.ai/
- Together AI: https://www.together.ai/
All images and videos generated using models run on fal.ai
Timestamps:
(01:43) The generative media industry
(02:29) From $2M to $100M ARR: fal's explosive year
(04:06) How Gorkem met co-founder Burkay Gur
(05:38) The hardest decision that saved the company
(09:52) Spotting the opportunity in generative media
(13:28) Turning Todd into George Clooney
(15:29) The early adopters of the first fal product
(17:54) The transition from toy to tool
(19:27) Why 2025 is the year of AI-generated video
(21:44) Staying nimble as a 45-person company
(24:42) Predicting AI-generated film in 2027
(27:24) Why generative media is a greenfield market
(30:33) fal’s greatest optimization wins
(34:42) Why fal has 500 Slack channels
(36:02) Competing in a fast-moving, fragmented market
(42:06) How to build a world-class team
(47:24) Learning sales as a technical founder
(50:55) How fal built a brand without a marketer
(53:21) The story behind "GPU Rich / GPU Poor"
(54:22) Inside fal’s rule-breaking playbook
(56:09) The hardest part of scaling fal
Gorkem Yurtseven: We even had paying customers. We did both products for a while at the same time, we tried to convince each other like the pivot is not as drastic as it actually is.
Todd Jackson: Hey everyone, it's Todd Jackson. I'm a partner at First Round for today's episode. I'm excited to sit down with Gorkem Yurtseven. He's the founder and CTO of Fal, a generative media platform for building with AI image, video and audio models. I invested in Fal's seed round in 2022, but back then Gorkem and his co-founder, Burkay, were building a completely different product for data teams.
Gorkem Yurtseven: We were doubling down on this data infrastructure. We thought at the time, compute was gonna be big.
Todd Jackson: But when Stable Diffusion emerged, they made the tough call to pivot, walking away from paying customers.
Gorkem Yurtseven: We tried to explain this to people because it was so new. No one got it. We actually had a hard time raising our Series A.
Todd Jackson:
They made these models run so fast that when they showed me a demo turning my face into George Clooney, it literally looked like a video. Since then, they've grown from 2 million in ARR to over a hundred million.
In just one year, serving 2 million developers and raising three funding rounds in 18 months.
Gorkem Yurtseven: Investors criticized us that this revenue may not be durable. So we took enterprise sales very, very seriously.
Todd Jackson: In our conversation, we dig into their pivot story, how they stay ahead in this fast moving market and what's next for AI video.
Gorkem Yurtseven: Traditional marketing doesn't work for developers. People think it's cringe. We just jumped onto the meme and created these two hats. He ran out of the GPU poor hats like way before the GPU rich ones.
Todd Jackson: Let's dive in. Gorkem, welcome to the show.
Gorkem Yurtseven: Thank you so much. Thanks for having me, Todd.
Todd Jackson: We're going to dig into a bunch of different topics today, but just to make sure everyone has enough context, could you start by explaining at a high level what Fal does?
Gorkem Yurtseven: Fal is a generative media platform for developers. We host image, video, and audio models as easy-to-use APIs using our inference engine behind the scenes and other developers working at other big companies or solo developers build products on top of these easy-to-use APIs.
Todd Jackson: I think that you guys really pioneered this category of a generative media platform and it's just been an incredible year. I know over 100 million in ARR now, over 600 models, two million developers, 300 plus enterprises, and a series A and a B and a C all in the last 18 months. It's been a crazy year. Can you share a little bit about what the business was like this time last year, one year ago versus today?
Gorkem Yurtseven: We thought we were growing fast, but we were only at two million ARR maybe in August and we had a slow summer. I think that in general, the space, the image model space had a slow summer and it was right after Stable Diffusion XL was released in April I believe, and then for the whole summer it was very, very slow and there was basically no new model releases, everything was getting cheaper and cheaper. Our usage was growing a lot, but because everything was older, people were able to run them more efficiently including us, so our revenue was hovering around two million. And then first the Flux Model family was released. That was really good for open source and for us, and then many other models right away were released. We had a really good launch with Flux models and then video models that's really big for our business. I think around October was the first video model that was commercially available in the API. Since then, we've been growing that.
Todd Jackson: The growth has just been crazy since then. Okay, well let's go rewind the clock a little bit because one of the most interesting parts to me of your path to product market fit that I think not everyone knows about is that you and Burkay started off with a completely different idea. So I'd like to go all the way back to the beginning and just sort of tease out some of the lessons from the pivot that you guys did. Can you just tell the story of how you met Burkay and how you decided to work together?
Gorkem Yurtseven: I met Burkay when I moved to San Francisco over 10 years ago. We are both from Turkey and we went to high school there and came to the U.S. to study in college and then we both moved to San Francisco right after college and we were working at big Silicon Valley companies completely independent from each other. So we met socially by just being in the city, being Turkish together through common friends and we actually never worked together until fall. He was working at Coinbase at the time, I was at Amazon. This was during COVID times. We went to Palm Springs together and everything was in a lockdown in San Francisco and we rented the house there, stayed there for a couple months just socially. We both had our own jobs and then we started discussing some ideas and that's when the first seeds of Fal was planted and then it was maybe seven, eight months after Burkay quit his job and then shortly after I followed doing the same and we kind of went around the idea maze, tried to do something in the machine learning space. We knew building for the future, building for developers was always the right track, but we had to do some exploration around what ideas that might catch on and we had to get a team of, I think at the time, five people to try some things, open source, tried to work with some of the enterprise companies that we had connections with and we were doubling down on this data infrastructure area because we thought at the time, compute was going to be big. We were following footsteps of Databricks and Snowflake. They were able to monetize compute in the cloud very well. Turns out our bet was correct. In a way we are still doing the same thing. We are doing compute in the cloud, but the workload we were targeting was specifically for data transformation in big companies where there's a lot of data so that people can transform this data to use it either for AI in the past and maybe analytics. It was first DALL·E 2 and then Stable Diffusion and then ChatGPT and then Llama 4 and all of these things released within months of each other, the whole AI world kind of came backwards in a sense that you don't need data anymore to actually train these models. The models are already trained for you. And all of a sudden like okay, you can do a lot more with a ready-made model for you. So that was the spark for us to go for the pivot. If there's a ready-made model that changes everything, this whole data preparation stage can be skipped and only the biggest of the companies are going to do that. Everyone else, they'll just use something off the shelf and that attracted us towards inference.
Todd Jackson: The thing I think is really interesting because I first round invested in 2022 on the original idea, the product for data scientists. And one of the things I think is so interesting is that that idea in my memory was kind of working. It's interesting to pivot to something else when the product, the initial product that you have is actually kind of working and you had customers. Was that a hard thing to do to walk away from something that was working to something else that you thought might be better?
Gorkem Yurtseven: We had customers, we even had paying customers. We did both products for a while at the same time. It's very hard when you are not screaming exactly what you're doing to your customers, to the potential customers, to people you work with, it's really hard to sell because they look at your website, they see something else. So it is really hard to do both of them at the same time. Probably did it for two months or three months, maybe two months we saw the revenue growth for AI inference was growing much faster than data transformation. So we decided, all right, maybe it's time to say bye to our old customers and double down on inference.
Todd Jackson: I mean is that a psychologically hard thing to do when you have users, you have customers, you have investors who invested in that original-
Gorkem Yurtseven: Yes, everyone knows us for something, right? There's the social aspect to it as well. Yes, it was very challenging.
Todd Jackson: What was the moment where you just sort of said, hey... and did you both kind of agree at the same time, like, we got to just do the new thing?
Gorkem Yurtseven: We always tried to convince each other, okay, maybe this is not a huge difference. Maybe this is not a big pivot we are doing. We're still doing compute in the cloud. It's just a completely new workload, so maybe it's not that bad. We tried to convince each other, the pivot is not as drastic as it actually is. So we lost some time actually in that phase where, okay, maybe this is not a pivot. Maybe we are just changing direction a little bit.
Todd Jackson: So if you could go back to that moment before the pivot and kind of knowing what you know now, what would you tell yourselves or what advice would you give to other founder who-
Gorkem Yurtseven: So one thing actually you told us resonated a lot with us. You said, "okay, which of the idea you think you're going to reach one million [inaudible 00:09:08] first? And then which of the ideas you think you're going to reach 10 million [inaudible 00:09:12] first?" We remember this very well.
Todd Jackson: The data science idea was faster to a million, but the generative idea was faster to 10?
Gorkem Yurtseven: Generative was faster to 10.
Todd Jackson: Yeah.
Gorkem Yurtseven: Yeah. And so it was like, okay. It's interesting. But then we actually reached one million with inference and 10 very quickly with inference as well. So our predictions were wrong, but the framework you told us actually helped us a lot in that decision-making process.
Todd Jackson: The other thing I was thinking about is when you first saw this idea, where did you see the opportunity? What was the problem that was going unsolved? Because at the time there were other inference providers. There was Together AI, there is Together AI, there's Base 10, there's a bunch of them. They were focused on language models and text. What gave you the conviction that there was a big enough other marketing media?
Gorkem Yurtseven: There is obvious value we saw in inference because all of a sudden you don't need a lot of data yourself to train a model. You can just use something off the shelf and this all of a sudden increases the number of AI product users by we thought maybe it's a 100X, 1000X, maybe it's a lot more than that. Maybe everyone becomes an AI user because it's so easy to create an AI product. That value was obvious and we thought, okay, this changes everything. Then within that idea we had to, okay, why stick with image inference and not do LLM inference? So I think that wasn't easy as well. Image came first, right? Stable Diffusion came before Llama 2. All the inference providers before some of them were doing other things as well, like Base Stand was also doing more traditional machine learning. So a lot of people saw the opportunity around the same time than us. Maybe they were able to pivot quicker than us. But either way, our first 10 customers or something like that were trying to build products on Stable Diffusion. So we had all of our customers doing that, but also within the Stable Diffusion world, we had to make some decisions as well, like are we going to just provide GPUs for people for them to just deploy any workflow they want or are we going to build easy to use APIs for them and they're just hitting an API endpoint? This was a big discussion within the company and we ended up building an API endpoint, which we called it like an inference endpoint rather than doing GPU infrastructure. So first we got over that hurdle and went with APIs and optimizing the inference process. And then we were so good at that we decided to stick with optimizing image inference because LLM inference was way different at the time. There were different technical, we thought the buyer is very different, the market is shaping up to be different. Obviously everything was happening so fast.
Todd Jackson: You thought it was a very different customer set.
Gorkem Yurtseven: Different customer.
Todd Jackson: But also a different set of optimizations that you had to do [inaudible 00:12:07].
Gorkem Yurtseven: Optimizations. Correct, correct, correct. And around that time we were actually okay, maybe it's a good time to start raising our series A. And we tried to explain this to people because it was so new no one got it. Everyone thought an inference platform is an inference platform. Doesn't matter what kind of model it is. There are other people who are more qualified or more prepared to do this than us. And we were trying to explain them how our focus on, at the time there were no video models, our focus on image models is going to actually make a difference and people thought the market was smaller as well. So all these things on top of each other, we actually had a hard time raising our series A. It wasn't that easy. And the fatigue was because all these big inference providers had raised maybe a month before or around the same time we were raising and all the investors, they were pitched the same thing over and over again and it sounds the same and that's something we underestimated how disadvantage of a situation it is to raise at the same time with seemingly all your competitors because they all say the same story. Investors hear it over and over again and there's some fatigue of hearing the same thing.
Todd Jackson: But you guys definitely had a lot of conviction that the actual stack that you needed to build, the problems that you needed to solve were very specific to what you were doing. Because I remember in 2023 when you came to our office, the first time I had seen this new idea and I think it was on Stable Diffusion STXL or something where it was on your laptop where you had this, through the webcam, this video of me turning me into George Clooney waving my hand. It was a video but it was actually like frame images-
Gorkem Yurtseven: Frame by frame.
Todd Jackson: ... images that you were generating. And it just blew my mind how fast it was. What were some of the things you were doing behind the scenes to make that possible?
Gorkem Yurtseven: That demo is when we announced this new iteration of the company to the rest of the world. Still we didn't make any money from that demo. It makes a really impressive technical demo for people to see how fast we can run inference, but we couldn't find any use case for that very fast image to image inference. Still to this day it is very impressive, but we couldn't find any commercial value with it. There was commercial value in image inference just in general, but that was very good for marketing but not so good for actually monetizing.
Todd Jackson: How did you get it to be so fast?
Gorkem Yurtseven: At the time we had a two-person inference team that Batuhan, our VP of engineering, he has a compiler's background, he loves optimizing things at a systems level and then we had another engineer who was really into writing Triton kernels. So we looked at the program, what are the parts we can optimize, how can we run the program more [inaudible 00:14:56] and Batuhan, me, the other engineer we have, Burkay, we all came together and obsessed over it over a couple of weeks and we were able to, at the time it was STXL and STXL distillations. That was a distilled model. There was one model that we were able to optimize and that was it. That's what people used to build products. So that was it.
Todd Jackson: And I remember a bunch of stuff on Twitter, some of the demos that you said kind of went very viral, but then let's get to the first actual product. When you started releasing and had developers using it, what was the first version of the product that you were giving to developers and what were the shortcomings of it? Because anytime you launch a brand new product, usually it could be very good at some things or probably not good at a bunch of other things.
Gorkem Yurtseven: Yeah.
Todd Jackson: What was the first version like?
Gorkem Yurtseven: I mentioned this before, instead of focusing on GPU orchestration and letting people deploy whatever they want, we decide to build APIs. Every single code that's deployed is owned by us and we control the whole process and that becomes okay, if the API can do what you're trying to do, yes you can do it. If not, you're restricted by what the API can do. So there were other competitors, they were allowing any workflow, any code to be deployed. So that was a big shortcoming. But we knew what people want to do. Everyone wants to do the same thing over and over again and therefore we thought there is value in actually optimizing the most common workflow and then people will realize that this is actually what they need. And it ended up being like that.
Todd Jackson: So let's talk about some of the earliest customers and their use cases. How did you get your first 10 customers and how were they using it?
Gorkem Yurtseven: I think almost all of them were horizontal design or image generation products. When a technology is so new, I think it's hard for products to be specialized. So a lot of them were general either consumer AI applications or web-based image generation apps.
Todd Jackson: And I remember them, some of the early ones being like I thought of them as a little bit like indie dev kind of hobbyist stuff. Sometimes when you see a new technology, it's a little bit like a toy. And some of your early customers were these kind of indie devs. Did you have concerns about that or did you think this is just how it's going to start?
Gorkem Yurtseven: I think one of the biggest signs of how this is going to be around is the amount of money they spent, right? Everyone was spending serious money on the platform, tens of thousands of dollars a day all of a sudden. And maybe this is temporary, but as long as it's going, it's definitely not a toy. People are spending serious money making serious products used by real people and to this day it's growing as we speak.
Todd Jackson: Did you have a sense at that time that eventually, look, enterprises are going to want this?
Gorkem Yurtseven: I think so because every piece of technology was like, it was just too magical to be ignored. The models just had to be a little more capable and we could see how every month, every three, four months, the models were getting more capable and people were moving into video. A lot of money was poured into training video models, so we knew things were going to get more serious. The timing, none of us expected it to be as quick.
Todd Jackson: I remember Flux, it was sometime in 2024, just being a very big moment for the company. And you guys had day zero support for Flux. The Flux moment in 2024 was kind of like the nano banana moment I feel like we're [inaudible 00:18:35] video, but how did you get day zero support for Flux?
Gorkem Yurtseven: The Flux team worked at Stability before, so we had a relationship with them already through their times at Stability. Of all the demos we did attracted their attention and we were able to become friends and colleague working partners during their time at Stability. So they were pretty under the radar. No one was expecting the Flux model, but we knew they were doing training and we really respected the team and we knew that they were going to do something awesome. We got in contact with them throughout during the summer and we planned a big release together and it worked great for us.
Todd Jackson: Okay, so video has been a huge part of your focus over the last 12 months, and I remember last summer you guys were telling me 2025 is going to be the year of AI generated video. It feels obvious now, but a year ago at least to me it was not that obvious. What were the signals you were seeing that were giving you that conviction?
Gorkem Yurtseven: So first of all, researchers always want to be in the cutting edge and all of a sudden, I would say a couple months after Flux, we realized we looked around, all the researchers that were good at the diffusion space, they started working on the video problem. All of a sudden, everyone left.
Todd Jackson: Pre Sora or Sora had already happened?
Gorkem Yurtseven: Flux happened in August, Sora was February before that. And I couldn't run it. Maybe some people did, but it was just a demo. After Sora, serious money was put into training video models. Yeah, exactly. And researchers left the image field and started working on video, which is interesting. I don't know if there's a term that describes the situation because not all problems were solved for image by any means. There was so much to do there, but the price for video was much shinier and there was a lot of VC money being put into it. All the researchers left doing image research and then focused on video.
Todd Jackson: So you guys saw those signals happening and you started to prepare the company, right? What were some of the new challenges that you had to solve for video?
Gorkem Yurtseven: Models were a lot bigger, which was to our advantage. The same reason why a bigger model like Flux compared to Stable Diffusion XL was to our advantage. Same way, a bigger model that requires more compute power, requires multiple of the fastest GPUs to run in the cloud. So the optimizations we make matter a lot more. If something takes one second, if you can shave off 20% of it, maybe not enough people care about it, but if something takes a minute and you can shave off a similar percentage, all of a sudden that's a lot more meaningful.
Todd Jackson: So now you guys find yourself kind of just at the center of this really interesting fast moving market. It feels like every week there's something new that happens. And so my question is how do you organize yourselves internally to be responsive to that? You guys are 40 people now?
Gorkem Yurtseven: 45 maybe.
Todd Jackson: How do you operationalize a 45 person company to be this responsive when the market is changing every week?
Gorkem Yurtseven: Among many things, positioning ourselves as a generative media company helped us hire the people who are the most interested in this, right? We have applied ML team, it's around 15 people right now, and all they do every day is either deploy these models, optimize them, play with them, and they're obsessed with it. It's as if this is their hobby and if they didn't work at Fal, they will be doing this anyways. So they're extremely on top of the market. And now with Fal's position in the market, we also get some early information from the research labs. Everyone tries to talk to us and release their models on file on day zero, so we get some additional access to information as well. These two things combined, we have some advantages over others in the market. But unexpected things happen as well and we have to react to it really fast and we are able to build that muscle very well in the team. Like something drops, immediately we drop everything and try to release that model if you don't know about it beforehand.
Todd Jackson: How do you know if a new model is worth to do that?
Gorkem Yurtseven: Exactly. The first phase is, okay, how can we evaluate this as fast as possible? Is the demo video that this model is released with, is that accurate? Can we actually compare this to other similar models? Is it actually worth it?
Todd Jackson: But if you get a sense, this thing is state of the art-
Gorkem Yurtseven: This thing's good, then we drop whatever and try. Usually on Slack, we open a huddle, eight, nine people get in there, someone shares their screen and we try to get it out as fast as possible. Some of it, we actually did it publicly too. That was interesting.
Todd Jackson: Oh, what did you do publicly?
Gorkem Yurtseven: When a model first drops, I think Batuhan did this couple times, or I'm doing a screen share, I'm speed running, deploying this model on file and people watched it. You definitely do this internally, like Batuhan would share his screen and others will watch the speed run to deploy the model to be ready.
Todd Jackson: Are there other things that sort of go along with that? Do you have huge demand spikes when a new model drops?
Gorkem Yurtseven: Yes, there is and we try to allocating GPUs to a model and doing that as elastically as possible. It's a problem that we have to think about all the time and we spend a lot of time managing our GPU capacity, how to do it the most efficient way. Sometimes we have to get GPUs really quickly. Sometimes we are able to plan it. But yeah, this is an ongoing problem that we have to solve.
Todd Jackson: One of my favorite things about working with you guys is that every time we meet, I feel like I get a crash course on the current state of the market and just a lot of opinions and insights on where the market is headed. So for example, I remember talking to you guys and you're like, "yeah, Netflix is starting to use gen AI. We're not far away from two and a half hour feature films being completely generated." Give us a picture. What does the world look like in 2027 to you?
Gorkem Yurtseven: I would say six months ago I wasn't sure if studios were going to actually use this or is it going to be so fast that all of a sudden you have independent movies on YouTube and studios completely missed the strand, right? Sometimes technology happens in light speed and you're stuck in innovators dilemma and other startups, upstarts, whatever companies leapfrog you in a way that you don't get to even compete with them. I thought maybe something like that was going to happen to studios because they're already having trouble with some of the labor situation they're having already. The film movie industry is having trouble in terms of the revenues they're making. There is very successful couple movies a year, but the long tales of movies, they've been having trouble. So I wasn't sure if studios were going to be able to actually use this technology. And it looked like they were completely avoiding it for a whole year. But something changed this summer and we are getting a ton of interest from basically all the studios in LA or elsewhere. Everyone is really interested to at least do something about it because now they understand this is good enough and they can actually save money by utilizing this technology. The creative people inside the studios I think they've spent enough time with this tool that, okay, now they see this as something that enhances their creativity rather than just replacing them. So they bought into it as well. So this summer I would say that was the biggest difference.
Todd Jackson: Are there other industries, genres, games, other things you think will be really, again, two years from now?
Gorkem Yurtseven: Yeah, gaming industry is a little more sensitive. People really care about things being hand-drawn or the creativity aspect is really important. But yeah, I think it's very suitable for gaming industry as well.
Todd Jackson: In thinking about the market, one of the things that we've talked about before is that image generation, video generation in many ways is a net new market. It's a new opportunity, it's a greenfield opportunity. Why is that an interesting idea for you?
Gorkem Yurtseven: Yeah, so this is one of the reasons why we decided to actually brand the company around generative media. We believe this is net new. Anything we are creating is not taking market share from a giant company. We are not a database company taking database revenue from Oracle or we are not in the traditional LLM market. The biggest use case seemed to be search, any LLM inference is taking market share from Google search and this is really important for Google. They would give it away for free if they had to. So they're protecting it really, really close to themselves and that's why I don't think it's a startup's game to play in that race, but for image generation video, it's completely net new. So we believe it's very suitable for a startup. And one of the other things is it was a small market in the beginning but very fast-growing. That also makes it really, really suitable for a startup as well. And even if like Google, OpenAI, even if they're active, they have their own models, but because there isn't a clear target in front of them, they're not as nimble as a startup. So it's really hard for them to go towards that target because that target doesn't exist. They have to reinvent it every single month as we do. So that puts us and the market in a very unique position for a startup like us to stay very nimble, change direction very fast, get close to maybe advertising, maybe studios, maybe more on the design side, maybe e-commerce and retail and build the whole team around this rather than just having a single target and going towards it.
Todd Jackson: In some ways I think these market conditions ended up working out so well for you in that it was an overlooked market at the beginning or people thought it was too small, which is great as a startup because it means that you enter and there isn't a lot of other competition. And then even as it's grown there's no clear giant that you are threatening and who's going to come after you. Did you predict all of it, did you know this was the case?
Gorkem Yurtseven: Maybe we got a little bit lucky. And in the whole AI market, it looks like it's very hard to differentiate with models. If you have a good model, you maybe have a three-month lead, maybe four months. People catch up for two reasons. Number one, if you had a unique research insight, that leaks and other people do it. Number two, once people see something's possible, it's way easier for them to actually go try to do it because now they know it's possible, someone else did it and other people go do the same thing or try to reach to the same idea. And number three, these models have APIs, people distill it. If you have a strong model, you can create similar models for way cheaper. So it looks like it's going to be really hard for someone to differentiate with the quality of the model. And we believe this is going to lead to even more fragmentation. So a company like Fal where you get to access many different models at the same time is going to be positioned very well in the market going forward, the better it is for us. Yeah.
Todd Jackson: I think a lot of people don't fully understand what you guys do at Fal and why it's so hard. Part of that is you've done so much work on GPU infrastructure, on model cold starts, you rarely talk about this stuff publicly I think, correctly because you put the focus on the models, you put the focus on what creative things people are doing with them. Just for a second, because I think the amount of systems work and all the creative engineering that you guys have put into the platform is wild. So you don't have to give away your secret sauce, but what are some of the biggest optimization wins that you've had that people don't know about?
Gorkem Yurtseven: Yeah. So one thing people don't appreciate, it's a different problem to host a single model and optimize it and just serve that versus hosting 600 different models at the same time, all various different architectures, various different problems and all of them have different traffic patterns. It's a much, much harder problem than hosting a single model. Any research lab either hosts their marquee model and maybe a couple other. So the number of models they have to host is less than 10. And for us it's around 600, which is complicates things a lot. And we had to build very, very serious systems to support this. And one of the things that we have to pay attention to is how we utilize these GPUs. We can't just say, all right, we are going to equally deploy these models to all the GPUs and hope that the traffic patterns are equal. Everything has to be elastically scaling up and down. And to be able to do that whenever any new request comes in, we should be able to auto-scale up and down really fast. And that is only possible with couple things. First, cold starts is important, starting a new model as fast as possible when it's a completely new start.
Todd Jackson: And we're talking about seconds or milliseconds?
Gorkem Yurtseven: Seconds when everything is fresh. But then we have to be really smart about how we cache these models. That's part of the strategy. We are running in I think 28 different data centers. So when a request comes in, we have to make sure the request goes into a node that has this model cached locally or at least close enough so it can load into memory as fast as possible. And we do cache these models in memory sometimes too, even if it's not being used at the time, it stays cached while it's serving another model. Cold starts, the caching strategy. And if the model is latency sensitive, for example audio models, it matters a lot if the model is in a data center that's close to you because you care about milliseconds, then routing the request to the closest GPU cluster is important as well. When a node serves a request, how long it's going to stay alive and when it's going to shut down, that's an important parameter as well. So we have to optimize all these things to serve these models as efficiently as possible. We need to be so good that if you were to just host this one model on your own, you need to host 600 of them but still be better as if you are doing one.
Todd Jackson: If you could wave a magic wand and solve one technical problem kind of in generative media right now, what would that be and why?
Gorkem Yurtseven: One really important problem with model optimization, you want to be able to run the models in many GPUs because let's say given that things scale linearly, instead of it taking one minute on one GPU, you want to run it on two GPUs and let it take 30 seconds. But in reality, every time you add another GPU, the gains are not linear, it's a little bit worse. You got to be able to do it as linear as possible. So if the optimizations work linearly, I think it would solve a lot of the problems you are facing. You could be able to connect many GPUs and do really, really fast inference. This is an active research and active engineering problem and we are getting close to make it linear, but the more GPUs you add to the end, it gets worse. So in the beginning it's close to linear and then it gets worse and worse.
Todd Jackson: Another thing I've noticed is that you're very obsessed with developer experience and developer support. And I remember one of the founders that I work with was having an issue recently with Fal and I connected you guys on email and your response to him was instant. It was faster than the time you replied to my emails. Is there something that you guys do to make that customer obsession with the developer [inaudible 00:34:31] scale?
Gorkem Yurtseven: From the beginning, Burkay and I, we really cared about building for developers. We believe that's the best way to multiply our input to the world. We build for smart people who then build very smart products and then the impact is multiplied. If you look at some of the newest software products like public companies, they're all one way or the other developer platforms or infrastructure platforms, which is a subset of developer platforms. So we wanted to build a company for developers. We knew that was the biggest force multiplier we can get on our work. Once you decide to do that, you got to do your marketing towards developers, you got to make sure everyone gets the right amount of support, you got to care about things that developers care about and you can only do this if you obsess over the developer experience. And I think we're doing an amazing job at Fal at that. We have, I don't know, 500 different Slack channels with all the engineers from companies we work with and the response rate of those Slack channels, you measure that daily and we obsess over that.
Todd Jackson: The market is changing rapidly. We've talked about this. You guys are growing and so naturally there's competition. And I think you both have been incredible chess players and sort of navigate these changes. Well how do you think about competition in the space and future chapters of Fal and how you navigate that competition?
Gorkem Yurtseven: One thing that helped us a lot was okay, we had a technical advantage, a technical mode, how can we turn into a business advantage and a marketing advantage? Calling ourselves generative media platform and owning that term was really helpful. As biggest of the biggest enterprises were adapting this technology because when you are screaming what you do well people believe you, right? No one else is actually calling themselves generative media inference platform. To this day, no one is defining their company as that. And I think that positioning was a big advantage. And now we can go out there, talk about generative media and talk about the industry itself, but indirectly we are just talking about ourselves. Being in that position is really, really valuable and I think that differentiates us from our competition. We are almost defining the industry while we are basically just talking about our own company. So that is a unique position. And we always say this, being the market leader, there's a premium to that and yeah, we are seeing that premium because whenever a big company wants to get into this, we want to be the first phone call that they're making.
Todd Jackson: So speaking of big companies, enterprises, you've got Adobe as a customer, Canva, Shopify, what have been the biggest lessons in building for individual developers who you still serve and enterprises at the same time?
Gorkem Yurtseven: Our head of operations jokes about this. We became a legal tech company. With all these different AI models, the enterprise companies have a lot more concerns about, all right, how are these models trained? What happens with the inputs and outputs? Things like that versus our initial set of customers, they just want to go to production as fast as possible. So we had to build that muscle and being able to accommodate the requests coming from the enterprises about the legal side of things and making sure they're comfortable with using this in production.
Todd Jackson: And so do you feel like you're able to scale and serve both of these kind of different parts of the market equally well?
Gorkem Yurtseven: 100%. Yeah. I'm not scared of any security review anything. We've been through the hardest ones so we could do anything else.
Todd Jackson: And you guys are just scaling so quickly, it's one of the fastest early revenue growth rates I think in startup history. You said two million last summer and now-
Gorkem Yurtseven: Yeah, now it's over 100.
Todd Jackson: And I know that model launches also make things a little lumpy. It seems like a hard to forecast business. Is there anything you've had to learn about managing this explosive growth and forecasting for what you need?
Gorkem Yurtseven: Not forecasting, but you mentioned model launches. It's just an incredible situation because every single model launch that happens in the platform is an opportunity for us to do more marketing, hopefully with the research lab who was launching the model and then go out there, talk about this model, talk to more customers. An opportunity for us to talk to a customer we were trying to sign or get them to use the platform. Now it's another opportunity for a touch point for us. Look, there's a new model and this is happening. I would say it used to be once a week, now three, four times a week. So many model releases are happening, it's out of control. And all of them is an opportunity for us to get another big customer or get some more eyeballs on X or on Reddit. It's just an incredible tool for us to be on top of all these model releases and be the first platform to support them.
Todd Jackson: Do you think about revenue quality and revenue durability, like the revenue from one customer is different than the revenue from another customer?
Gorkem Yurtseven: In the beginning, mostly investors criticized us that this revenue may not be durable, the quality is not high enough. So we took enterprise sales very, very seriously. We built a sales team early on, maybe earlier than some of our competitors and we tried to get as much of this revenue in form of yearly commitments rather than pay as you go. To this day I think we are doing an incredible job at that. And that protects the revenue in a bit, but also proves that these are serious companies doing serious business and willing to commit for 12 months or millions of dollars in some cases. So this was our reaction to that criticism.
Todd Jackson: What are the top handful of metrics, maybe the top two or three metrics that you watch internally the most obsessively?
Gorkem Yurtseven: Revenue is number one. Anything else we've tried to do kind of became useless. Number of requests was something that we cared about. All of a sudden when video models were part of the one video model request is a lot more valuable than image models. So that doesn't matter anymore. The number of team accounts, if you sign up as a team and you have many members in your team, that means you are probably going to start spending more. So that's one thing we pay attention closely. Other than that, revenue has been the metric to follow. I know as a PM you don't like that.
Todd Jackson: I think it's incredible that as founders and especially as technical as you guys are, that revenue has always been your north star. As an investor I think you'd like to hear that.
Gorkem Yurtseven: PMs always want to find other metrics that are not correlated to revenue to track. Maybe it's like revenue in the future, but not revenue right now. It's been hard for us to find what that is, but maybe people do that when there isn't enough revenue so they're looking for other things to track. But in our case, there is enough revenue and that's what we've been tracking.
Todd Jackson: Okay. So let's talk about the team. You guys have built an insanely talented technical team. Have you been able to tap into talent that other people overlook or how did you find some of your key hires?
Gorkem Yurtseven: I think one of the advantages of working at a big Silicon Valley company for five plus years is you get to meet a lot of other very talented engineers. You could do that at a startup too, but you don't meet enough people there. And in a big company like Amazon or Coinbase, we met a ton of talented engineers and we were able to bring the most talented ones into Fal. It was their first startup job, maybe, right? They trusted us to work at in this environment and that's how we were able to bring them into Fal. So that worked great for us. And then yes, we are both Turkish and our early engineering team had some Turkish members as well. We were able to attract incredible talent from back home. That's because people trust us more and it's just a very good company for them to work at. That's been really useful for us.
Todd Jackson: How global is the engineering team?
Gorkem Yurtseven: I would say more in person in the six months people are willing to move here. We've been trying to get both our American engineers who live in other cities to move to San Francisco, but also some of the overseas engineers to move to San Francisco as well. The center of gravity has definitely moved to San Francisco, but we still have 15 people maybe remote.
Todd Jackson: What are you looking for either in the ML engineers that you hire, the systems engineers? Is there a specific thing that you evaluate?
Gorkem Yurtseven: It's either obsession with optimization, so if they worked at database company before or if they did any low-level systems engineering, that is a big plus. Even if they haven't worked with a GPU before, we believe they can learn very fast. Or obsession with the space. In the beginning this was harder to do because the space was so new there wasn't enough time for them to get obsessed with. But now we can five minute conversation immediately understand if they love video models, image models, they know what's going on and if you can feel that obsession, it's extremely useful for us. And those are the two things we are looking for when we are hiring ML engineers.
Todd Jackson: Is there anything unique or different that you guys do through interviews or the way that you hire folks?
Gorkem Yurtseven: This is an evolving process. In the beginning we were able to get to know people, spend more time with them somehow with the amount of time we spend in the ecosystem. So we were able to know someone for a longer period of time with the work they do out in the open and we had more things to look at before deciding if they should join or not. And now we have a recruiter, we have a bigger pipeline. So some of the people, I'm meeting them for the first time at the interview. So that process has been evolving. What are the things we can do differently than Meta or Google that we can identify the unique things that these people can be valuable to a company like Fal rather than any big software organization.
Todd Jackson: How do you recruit an amazing ML engineer who's also talking to the various large labs? That's got to be hard, right? Is it just obsession with the category?
Gorkem Yurtseven: So this question used to be how do you convince someone who has also an offer at Google or Facebook? I think we are past that. Like you said, we need to look at some overlooked career paths, find people non-traditional backgrounds, maybe younger, earlier in their career that might help open AI and anthropic there. Young companies too, they can't be interviewing everyone. So there's enough talent in the world if you know where to look.
Todd Jackson: And speaking of just talent and how concentrated the talent is at Fal, you guys have kept the team lean. I think you were at 25 people when you crossed 50 million and now you're at 45 crossing 100 million. And how many folks do you have on sales now, go to market?
Gorkem Yurtseven: We have around six, maybe 10 if you include CSM and all that.
Todd Jackson: Okay. But still, how many companies do you know at 100 million revenue with 10 people on the go to market side?
Gorkem Yurtseven: I think AI kind of changed a lot of things in how software companies did sales because before the market was more stagnant, you had to actually go convince people to use your product and they were choosing between five different options and the sales process was really, really long. But with AI you have so much demand coming from the market, you have to qualify. Your problems are very different. Your problem is you have to qualify who to spend time with, you have to qualify who's going to have the most spend among these companies for you to actually go talk to them. So the go-to market motion is very fast-paced, in a way transactional. So it's very different than the traditional SaaS sales cycles. And yeah, we have to hire for that type of AE profile, which has been an interesting challenge for us as well
Todd Jackson: As technical founders has been your biggest lessons in this area of teaching yourself sales. I know you did a bunch, most of the early sales.
Gorkem Yurtseven: I think the biggest lesson was starting early to get people into commitments because immediately that's a measure to understand how serious the person in front of you is as well. If they're willing to invest in your relationship and do they trust you? Do they think you can do a good job and make them successful? Getting people to commit as early as possible and giving them that option, I was a huge skeptic. I was like, no one's going to commit to this.
Todd Jackson: You didn't do sales?
Gorkem Yurtseven: No, I didn't do sales. No. Burkay didn't either.
Todd Jackson: So how do you sell the product, especially to a large enterprise, how do you sell the product?
Gorkem Yurtseven: First we want them to try it out. We make sure that's as easy as possible. You log in, put in a credit card and you can just use the product. A lot of our contracted revenue even comes from inbound. People already tried the product and they have some sort of spend and now we are reaching out to them, all right, you're on track to spend $10,000 this month. If we give you 10% discount, 7% discount, would you be willing to do a yearly or two yearly commitments?
Todd Jackson: Your annual enterprise customer started as there's a couple engineers doing pay as you go basically.
Gorkem Yurtseven: We have some signals on Salesforce. If someone's, let's say, I think it's $300 a day or something like that, if they spend more than that, it creates an opportunity on Salesforce. One of the AEs take that opportunity. There's an email associated with it, they reach out, they try to get on the phone and convert that to a yearly contract. There are very few high profile like Neon Burkay would reach out to someone who we know they have a lot of generative media spend. There's a couple of those, but majority happened inbound first and then we converted to a yearly contract.
Todd Jackson: Outside all of the work that goes into product, are there any examples of how you guys are using AI internally?
Gorkem Yurtseven: Yeah, our product team uses Cursor or equivalent tools a lot. I see the monthly bill and it keeps increasing. Yeah, I think it's better suited for product engineering type work as opposed to some of the low level optimizations we are doing on the ML side. Yeah, I would say our product engineering team and our API team allows a Cursor and similar products. On the sales side. I believe we are using Clay to enrich some of the leads we have and figuring out who to reach out, emails, stuff like that. I think that's pretty much it.
Todd Jackson: So shifting gears a little bit, but staying on team. You guys have some awesome open roles now for marketing, but I think that you guys have done a pretty incredible job marketing the product and the company without having sort of formal marketing people. You've got this amazing swag, know the GPU rich, GPU poor hats, the website is awesome. You've sponsored the padel courts in San Francisco, nano banana hackathons, generated video awards and your first developer conference coming up. So how have you thought about brand building and having taste as founders and where does that inspiration come from?
Gorkem Yurtseven: Yeah, shout out to Adam Hall who worked on our initial branding, which to this day we use it and we had a couple other iterations to even expand on it. He did a great job and we just trusted him. We didn't have any designer in-house or a marketer, so we just asked him to go crazy with it and he came up with this amazing brand that we have today and we love it so much that we want to show it everywhere. So it's been great. In terms of developer marketing, I think that's a unique taste. Traditional marketing doesn't work for developers. People think it's cringe. People don't want everything to be very obvious on their face. They want this subtle tasteful marketing that is only towards developers. I believe indirectly we are doing it very well. We all have really active X profiles. We hired a couple people just because they had an active X and they ended up being really active members of the community as well. So that helps a lot. Yeah, we try to do hackathons, we try to go to conferences, but at some scale this needs to be more organized. Therefore, we are looking for couple, not a lot of positions to help us put more organization into this. Everything has been founder led more spontaneous. We just decided to do something one day and one of us goes and executes it, but you need to do this little more organized and we want to make sure we're not missing important conferences. You don't have a hackathon and another event on the same day. We want to make sure we are spreading this enough.
Todd Jackson: Did a lot of it just come from the way that you and Burkay like to be marketed to yourselves? A lot of it is fairly intuitive?
Gorkem Yurtseven: I think so. In a way, we didn't hire a marketer early on. This is a really important position. You're presenting the Fal brand, you're presenting the company. You wanted this person to be as good as it can be to represent us basically. So we wanted to do this on our own in the beginning to know how to do it, where we fall short, maybe fail a couple of times and exactly understand where we need help so we can hire. And I believe now is the time.
Todd Jackson: How'd you come up with the GPU rich GPU poor hats?
Gorkem Yurtseven: That's one of the earliest things we did. At the time, Dylan from Semi-Analysis blog wrote an article about how everyone is GPU poor except Google. I think that was the main point of the old article. And we had a conference coming up maybe in two weeks, three weeks. We didn't have that much time. And this became a meme on Twitter. We just jumped onto the meme and created these two hats. One of them very basic GPU poor, just white on black plain font. And then this GPU rich looks like a country club, green on white. And we just showed up to this conference with these two types of hats. We thought, oh, people would like the GPU rich all of a sudden and we made equal number of them and we ran out of the GPU poor hats way before the GPU rich ones. Everyone thought GPU poor was hilarious.
Todd Jackson: Is there anything about your culture or how you guys operate that would sort of seem, I don't know, kind of weird to outsiders, but is very essential to how Fal runs?
Gorkem Yurtseven: Yeah, we don't have engineering managers. We have around 32, 34 engineers. We do have leads, obviously, we have leaders in the team, but we don't have this engineering manager role. Everyone is always contributing writing code. There's no one whose job is just to manage people.
Todd Jackson: That reminds me of early Google actually. I think there was some engineering VPs there and each of them had 40 or 50 direct reports, something like that.
Gorkem Yurtseven: At Amazon we had engineering managers. Their job was to maybe set goals but also have one-on-ones with everyone and that was it. And we don't have that.
Todd Jackson: When's that going to break?
Gorkem Yurtseven: We'll see, I don't know. One thing I like to do is we also have very few one-on-ones. Instead of one-on-ones, we try to do smaller groups of discussions like one-on-three or whatever, one on four. And we try to bring people from within the team, but maybe someone who joined recently, someone who's been there for a while and someone who's remote, someone who's in office, different types of people together so that it's a discussion. If they want to complain about something, they can complain. But usually those meetings are a lot more constructive than the old one-on-one style where you're almost forcing someone to complain about, okay, this is your time to start complaining about it. And when you give that that time, people complain, that's what they do. So we believe this group style is a lot more constructive.
Todd Jackson: Well, some questions to wrap up. What is the hardest part about building Fal that you didn't anticipate when you started?
Gorkem Yurtseven: I would say this is the hardest part of, I would say any growing company, not specific to Fal, is hiring executives, trusting people who are very experienced to come in and take over pieces of the company that have been there for a while and trusting them to actually do a good job. When you ask the question, where did you make mistakes, this is usually what people say and we are trying to be extra careful with it not to make mistakes. And we are taking as much time as possible. Yeah, that's been the hardest. We built a sales team before we hired the head of sales. I think this is number one question like series A or series B companies ask themselves what comes first. Decided to hire, I think six AEs first, everyone reported to either me or Burkay, and then we hired the head of sales. I think that was the right thing to do, but maybe it was the wrong thing to do when we look six months from now and the whole team is fighting with each other, I don't know. Hopefully that doesn't happen, but I'm now much more confident about evaluating this new head of sales we hired and the AEs, what success looks like. I saw it. I experienced it. It was painful to get there, but now I'm much more confident in my ability to actually evaluate the situation than if I had hired the head of sales before and they hired the whole sales team.
Todd Jackson: What's been the biggest surprise about being a founder versus being a technical leader?
Gorkem Yurtseven: It's just a wide range of things you have to do any given day. You might be negotiating marketing material on a padel court, also getting on recruiting calls with ML engineers, but also giving investor updates and all of these might be happening back-to-back meetings and you're responsible for all these different things.
Todd Jackson: What are the things that you're thinking about personally right now, and maybe Burkay too, if you know, about leveling up as an incredible founder over the next year or two?
Gorkem Yurtseven: One difference from having a five-person team and 45 person team is actually the amount of good information you get from everyone in the team and the amount of market intelligence, the amount of insights that we get to discuss in our office or in our stand-ups, and now I get to go represent those ideas to other people. That is insane amount of leverage I think you don't necessarily get if you are an engineering manager in a small part of a big company or if you are founder of a smaller company. I think that's been the biggest difference and it's been huge leverage for my personal development because I get access to all this very smart engineers in my team, and I get to represent their ideas.
Todd Jackson: Gorkem, thanks for being here.
Gorkem Yurtseven: Yeah, of course.
Todd Jackson: It was great.