Inside fal’s $1.65B Series E & F Led by Sequoia & Bessemer
Topics
Video Generation and Physical World Models
Commercial Pitfalls of AI-Native Applications
How AI Reshapes Enterprise Workflows
Open Source vs Closed Source AI
Generative AI Reshaping Film Production
Pioneers
Burkay Gür
Gorkem Yurtseven
Summary
- Generative media has the spending intensity to sit alongside coding agents as a major AI market. The founders describe individuals spending $10,000–$20,000 a month generating content. Burkay recalls Veo 2 charging $5 per five-second clip against earlier models at roughly $0.30–$0.40: better quality unlocked professional buyers and what he calls “essentially infinite demand.”
- fal’s changing revenue mix challenges the assumption that video simply replaces images. Video climbed from zero to around 70% of revenue before image editing revived image demand, bringing the mix back to roughly 50/50. The founders’ premature conclusion—“image models were done”—missed editing’s usefulness both independently and in preparing frames for video generation.
- Advertising offers a compelling feedback loop, although the complete automated workflow remains unfinished. Molly raises the mismatch between effective 60-second ads and short generated clips; Burkay responds that quick cuts and imperfect, homemade-looking content can suit UGC advertising. Generating variants, measuring engagement and automatically favoring winners is the opportunity, but “no one has done it perfectly.”
- The founders’ margin thesis favors businesses that innovate in the compute path. They argue application margins are “close to zero” because competition and marketing consume the economics, reversing the conventional expectation that infrastructure commoditizes while applications retain value. fal’s role includes anticipating model launches and securing scarce inference capacity before demand arrives.
- fal is testing whether a roughly 150-person company can scale through two reporting levels and hands-on technical leaders. Its “Super ICs” combine engineering, leadership and commercial judgment; Gorkem estimates AI makes the engineering team three to four times faster. Burkay says a flat structure may hit communication bottlenecks and require more management; Gorkem thinks coding agents may let Super ICs lead five to 10 engineers with two levels of hierarchy as long as the models keep improving.
- Open models retain practical value even as fal expands its closed-model partnerships. Studios want fine-tuning, LoRAs and control over unreleased characters and styles; open releases also give trailing labs an “incredible marketing tactic.” Across both model types, customers’ confidence that inference runs on fal’s GPUs matters because confidential creative material is often the input.
- Growing Hollywood revenue has yet to translate into the labor disruption the founders hear predicted. Burkay claims “zero impact” on Hollywood labor despite the industry becoming a meaningful revenue contributor over the previous six months; legal approval remains an adoption constraint. His next-year milestone is “timeless AI content,” enjoyable after the demonstration of technical capability loses its novelty.
- World models could create another compute-intensive market, although product-market fit remains a possibility. Gorkem moved from skepticism three months earlier to enthusiasm about faster, more persistent interactive worlds. He says gaming might become a fourth category, with a GPU running continuously while someone plays.
Deep dive
1. Quality gains expand spending without making images obsolete
- Gorkem’s market map has three product-market fits: search and consumer assistants, coding agents, and generative media. The distinguishing behavior is individual spending: someone at an AI-native or conventional studio can spend $10,000–$20,000 a month generating videos, comparable to heavy coding-agent users. “This is the only other use case in AI that an individual can spend that much money”; search gets a pass because of its scale.
- His first reversal concerns revenue mix. Video climbed from zero to around 70%, prompting the assumption that “image models were done”: images seemed too simple and insufficiently compute-intensive. Image editing then revived demand and improved video workflows by creating frames for interpolation. He now puts revenue at roughly 50/50, overturning their expectation that video would completely take over.
- Burkay’s pricing example: clips started around $0.40 per five seconds and settled at maybe $0.30 before Veo 2 arrived at $5. “Who the heck is gonna pay” was their reaction. Higher quality brought in advertising, marketing and movie-production buyers whom cheaper models could not serve. Usage rose alongside prices, producing what he calls “essentially infinite demand.”
- Gorkem says the best video model currently produces maybe 10–12 seconds, with newer ones potentially reaching 30 seconds. Until generation reaches real-time, high-definition streaming, he expects continued expansion. In his explanation, duration increases inference attention and resolution increases pixel counts, making costs rise “exponentially.” Burkay says this market is “nowhere near” diminishing returns.
2. Professional demand runs ahead of finished video workflows
- Molly’s pushback comes from her AppLovin conversation: the best-performing ads are about 60 seconds, while generated clips can be around five seconds and do not fit every brand. Burkay responds that an ad need not be one continuous generation; quick cuts can work better in certain cases. UGC-style ads also tolerate imperfect lighting and disposable output: “If it doesn’t work, make a new one.”
- The larger opportunity is a feedback loop: publish variants, observe engagement, make small changes and automatically show more of what performs. Molly offers a concrete example of that testing mindset: raising her arms as a hook reportedly got her 10 times the views; she also tests thumbnails, titles and copy. “No one has done it perfectly,” Gorkem says. Burkay adds that sophisticated creators and some large companies, including fal customers, already do versions of this. The reinforcement-learning and technical work remains difficult; he thinks it becomes the norm “in a few years.”
- Sora illustrates demand alongside operating constraints. Burkay says fal saw huge growth during the last three months before it was taken down; Molly calls it too expensive, and Burkay agrees. Gorkem says they saw substantial volume, while the explanation that there wasn’t enough compute for other products was “the rumor”; he also says Sora was expensive to run. He thinks it had a future: “Maybe it’ll come back sometime.”
3. fal kept its compute thesis and changed its market
- Burkay dates fal’s start to 2021 and a thesis drawn from Snowflake and Databricks: valuable data infrastructure businesses owned the compute. Their initial bet was that Python would follow SQL into elastic cloud workloads because AI was next. They served data-science and machine-learning teams before ChatGPT, Stable Diffusion and DALL-E forced a strategic choice.
- With five or six people and limited money, competing in open language models meant joining a price race they could not finance: “We can’t really go and sell tokens at a negative margin.” Instead, they accelerated image models on existing infrastructure. Burkay acknowledges they had no fully formed thesis about where those models would go.
- The commercial signal settled it: the earlier business ran for a year without making a dollar, while image-model revenue started growing. Roughly three to six months later, they recognized different use cases and buyers, especially ads and marketing, and put “the generative media company” on the website. Within about 18 months, people were discussing feature films, an application they had not initially imagined.
- The name FAL comes from “Features and Labels,” old machine-learning terms; “fal” also refers to having a fortune told from Turkish coffee grounds. The founders describe fragmentation across image generation, editing, video, 3D, text-to-speech and speech-to-text, with perhaps five to 10 companies per subcategory and frequent leapfrogging. Their explanation for continued progress includes abundant video data: nobody claims to have trained on all of it, and selecting interesting footage is the hard part. Modeling physics and the wider world also leaves multiple directions for improvement.
4. A flat company needs engineers with commercial agency
- Burkay’s recruiting model comes from joining Coinbase in 2016, when it had almost 100 people; fal now has about 150. Coinbase’s Reddit crypto quiz screened for knowledge before a phone call. fal similarly wants people consumed by the subject: “If you didn’t hire them, they would be doing this anyway.” Early engineering hires also came heavily from trusted former colleagues.
- The organization has at most two reporting levels and builds teams around “Super ICs”: leaders who remain excellent hands-on contributors. Burkay says this approach has a limit: communication bottlenecks may force more management layers, and they are watching for when the structure breaks. Gorkem thinks coding agents may let them maintain two levels, with Super ICs leading five to 10 engineers while agents do much of the work, as long as coding models keep improving at this pace.
- Gorkem says agency includes commercial responsibility. Because customers switch models so quickly, fal wants engineers to work with customers and make business decisions—a broader remit than the software-only roles he describes at large companies. He looks for independent action, responsibility and ownership of decisions, but says these are hard to assess without having worked with someone.
- Burkay pairs hunger with humility and uses “reverse selling” to explain the workload: crunch periods can mean 60 or 80 hours. He rejects toxic work hours: “We go home during weekends,” and not doing so is “nothing to brag about.” He also wants technically exceptional people who are low-ego, friendly and easy for customers to work with.
5. Compute planning and customer trust anchor the platform
- Burkay describes fal as infrastructure partner to model providers and API partner to applications and studios. Customers use multiple models and often synchronize product launches with model releases. A successful lab may not procure enough inference capacity itself, so fal forecasts demand and invests ahead. Otherwise, scarce compute can go to coding agents or whoever commits more capital; coordinating both sides has become an ecosystem responsibility.
- On open versus closed models, Gorkem expects a swinging pendulum: trailing labs gain attention by open-sourcing, while studios have technical reasons to want it. He calls open source “the only way” to fine-tune, apply LoRAs, change lighting and use unreleased characters or styles. fal initially feared closed models because it mainly ran open ones; partnerships serving closed-model inference now give it exposure to both.
- Molly frames the broader issue as protecting IP across media, source code and customer lists, and notes that local execution can help. Burkay says customers trust inference more when models run on fal’s GPUs, open or closed, because unreleased content is often an input. Open models provide a self-hosting option; customers choose fal when they believe it can run them better.
6. Hollywood adoption is early, with lasting content and games ahead
- Burkay’s Hollywood claim is sharply bounded: business from the industry reached a meaningful share of revenue over the last six months, yet AI has had “zero impact” on its labor and barely produced worthwhile Hollywood content. He calls four years of displacement discussion fearmongering while expecting change ahead. Adoption requires company lawyers to approve the workflow as well as models becoming good enough.
- His next-12-month milestone is the first “timeless AI content”: work people enjoy equally when revisiting it later. Existing clips mostly demonstrate capabilities. His benchmark is Toy Story, which he dates to “1996, something like that”: computer-generated work that remains enjoyable. He thinks the next year will bring a widely accepted work, maybe a short movie or animation, whose appeal survives its technical novelty.
- Gorkem’s change of mind concerns world models: “I was very skeptical even three months ago,” but recent progress made him bullish. These are interactive worlds navigated with keyboard actions. Better speed, quality and architectures allow more persistence: previously, you could see a flower, look away and return to find it gone. Keeping it there is his concrete example of improvement.
- He thinks prompting a playable game could become another product-market fit: “There’s a big chance that may happen,” and “it might be this fourth category, maybe.” The constraint is continuous compute: playing requires a GPU running all the time. Molly responds, “Very expensive.” He identifies compute as the big challenge even as he expects game quality to improve.
7. The margin thesis follows the compute path
- Gorkem answers AI-capex skepticism with internal ROI: he estimates fal’s engineers move three to four times faster and would have required roughly four times the people before AI to match today’s pace. The company pays less than that extra headcount would cost. He urges measuring actual users rather than relying on reports or nontechnical commentary, describing himself as “long compute.”
- Burkay’s framing is that “every big company is gonna be a compute company,” with different storefronts: coding agents, data platforms or fal’s media offering. Behind those interfaces, he says, revenue follows who runs the most compute. He also says headline figures such as $200 billion in CapEx include capacity not yet built; construction is slow and current supply remains insufficient.
- A founder’s margin hot take reverses the old infrastructure-versus-applications consensus. The argument is that application margins are “close to zero” because competition and marketing absorb the economics: “You can only create margin if you’re in the compute path and can innovate in that path.” Molly cites Tony Kim’s observation that chips have better margins, and both founders agree. The speaker’s instruction to investors: “Wake up, VCs.”
8. The closing bets span consumer consolidation and programmable media
- Burkay revisits a year-old OpenAI–Meta merger idea: Anthropic has established an enterprise position, Google has made its consumer move, and he no longer sees OpenAI clearly winning either segment. Meta also needs something; combining forces could strengthen its consumer position. He stresses that he has no information and is “completely speculating.”
- The other speaker says a merger would be bad for the whole industry, but does not elaborate on the mechanism.
- A newer closing bet is that coding is also part of generative media. Developers on fal increasingly make richer applications and programmable video, including websites that feel like videos, so the company may have to invest because customers already want it. The shared closing line is “Everything’s a computer.”