BREAKING: fal Raises $1.65B… | Sourcery Newsletter Summary

1. Key Themes

Theme: Generative media is a distinct, durable AI product-market fit with unusually high spend per user

Generative media as one of only three AI product-market fits

fal's founders frame the AI landscape as having just three clear PMFs, with generative media sitting alongside search and coding agents.

"AI really has 3 product market fits. One of them is search. The other really clear one that's growing really fast now is coding agents. And the third one is, we believe, generative media."

Individual spend of $10K–$20K/month is the signal

Gür points to per-user spend as evidence of real value. Only coding and generative media have users spending at that level, with the generative media user being a full-time professional generating video.

"This is the only other use case in AI that an individual can spend that much money."

Price increases expanded demand rather than shrinking it

Early video models cost about 40 cents per 5-second clip. Google's Veo 2 launched at $5 per clip, and quality gains pulled usage into professional ad, marketing, and film work.

"Veo 2 was very surprisingly $5 per 5-second video. So it jumped more than 10X. And we were like, 'Who the heck is gonna pay all this money?' Turns out there is essentially infinite demand."

Cost keeps climbing with resolution and duration

Gür said cost rises exponentially with resolution and duration, and he expects spend to keep growing until models can stream real-time, high-definition video. Current best models produce 10–12 seconds per clip, with about 30 seconds expected from newer ones.

Theme: Compute is where the value and the bottleneck sit

Margin accrues to compute providers, not applications

Gür says the early consensus that margin would flow to the application layer has reversed.

"Margins are close to zero in the application layer. There's so much competition. People need to spend so much money for marketing. You can only create margin if you are in the compute path."

Every big AI company becomes a compute company with a different storefront

"Every big company is gonna be a compute company. Just the storefront is gonna be different. Our storefront is unique and interesting. No one else has a storefront similar to us."

Compute shortages now shape model launches

Labs must secure inference capacity before launching, and many partner with fal to do so. fal forecasts customer demand and procures compute ahead of it so capacity does not all flow to coding agents.

"If you are coming up with a new model, you really have to plan ahead of time. You have to make sure you have enough compute to run inference."

Theme: The category is multi-model, fast-leapfrogging, and not winner-take-all

More labs and more leapfrogging than LLMs

Yurtseven says the generative media space has more labs and leapfrogging than LLMs across image, editing, video, 3D, and speech. Gür attributes part of this to abundant, unclaimed training data.

"There is so much video data out there and so much new video data is being produced every day that no one claims to train all of that data."

Image editing models revived image and now feed video

Video grew from 0 to roughly 70% of fal's revenue, and the split is now 50/50 after image editing models revived image usage. Keyframes made with image editing also feed video models.

"Turns out image editing models brought life back to image models, and then it's also useful for video models as well because you can create different frames, and video models does such better job interpolating between those frames."

Open source persists as a distribution strategy

Gür expects labs that fall behind to keep open sourcing models. Studios want open models they can fine-tune with LoRAs for lighting, unreleased characters, and house styles, and enterprises trust the process more when inference runs on fal's GPUs.

Theme: Next frontiers: ads, world models, physical AI, and Hollywood

User-generated-style ads are the near-term killer use case

Lower polish requirements, short clips stitched into longer ads, and engagement data enabling automated iteration make the format a fit.

"The content is usually throwaway. If it doesn't work, make a new one."

World models could become a fourth PMF

Yurtseven says world models went from unusable to compelling in about three months, with newer architectures keeping objects persistent. He sees AI gaming as constrained mainly by needing a GPU per player.

"You'll be able to actually build a game with a prompt, which I didn't believe in literally 3 months ago, but I think that space is just gonna blow up."

Physical AI and robotics

Black Forest Labs worked with robotics company Mimic to apply the backbone of its latest Flux model to robotics, using the model's learned physics to generate synthetic training data.

2. Contrarian Perspectives

The application layer has no margin; compute does

Against the early consensus that value would accrue to applications, Gür argues competition and marketing spend crush app-layer economics. The supporting reasoning is that compute spend is what drives revenue regardless of storefront.

"Margins are close to zero in the application layer. There's so much competition. People need to spend so much money for marketing. You can only create margin if you are in the compute path."

Hollywood is buying, but labor impact is near zero

Despite the narrative of AI displacing film workers, Gür says Hollywood revenue to fal has grown meaningfully over six months while labor impact has been negligible. Legal review and model quality, not just resistance, are slowing adoption.

"So far I would say AI had zero impact on anything in Hollywood. The business we get from Hollywood has been increasing in the last 6 months. But I would say it had zero impact on the labor."

Higher prices created demand instead of choking it

The conventional assumption is that AI prices must fall to drive adoption. fal's experience was the opposite: a 10X price jump with Veo 2 moved usage into professional work, with the founders surprised by the response.

"Turns out there is essentially infinite demand."

3. Companies Identified

fal

Description: Generative media inference platform for image, video, audio, and 3D models, serving developers, studios, enterprises, and model labs; founded 2021 with about 150 employees.
Why mentioned: The subject of the article, which raised $1.65B in a combined Series E and F (Sequoia led E, Bessemer led F) after a $140M Series D at a $4.5B valuation in December 2025.
Quotes: > "fal runs inference for image, video, audio, & 3D models, serving developers, studios, enterprises, and the model labs that build the models."

Sequoia Capital

Description: Venture firm.
Why mentioned: Led fal's Series E and its prior $140M Series D.
Quotes: > "Sequoia led the Series E and Bessemer Venture Partners led the Series F."

Bessemer Venture Partners

Description: Venture firm.
Why mentioned: Led the Series F and has been an investor since the $49M Series B in February 2025.
Quotes: > "Bessemer has been an investor since the $49M Series B in February 2025."

Meritech

Description: Venture firm.
Why mentioned: Led fal's $125M Series C in July 2025.
Quotes: > "...followed a $125M Series C led by Meritech in July 2025."

Google (Veo 2)

Description: Maker of the Veo video models.
Why mentioned: Veo 2's $5 per 5-second clip pricing marked the 10X price jump that expanded professional demand.
Quotes: > "Google's Veo 2 launched at $5 per 5-second clip, and more expensive models followed."

Kling and MiniMax

Description: Early video model providers.
Why mentioned: Their early models set the low-price baseline of roughly 40 cents (later about 30 cents) per 5-second clip.
Quotes: > "Early video models from Kling and MiniMax launched at roughly 40 cents per 5-second clip and settled near 30 cents."

OpenAI (Sora)

Description: Maker of the Sora video model.
Why mentioned: fal hosted Sora and saw volume grow sharply before it was taken down due to expensive compute demand.
Quotes: > "fal also hosted OpenAI's Sora and saw volume grow sharply in the 3 months before it was taken down because of it's expensive compute demand."

Black Forest Labs

Description: Image model lab behind Flux.
Why mentioned: Case study in physical AI, applying Flux's backbone to robotics with Mimic.
Quotes: > "Black Forest Labs worked with robotics company Mimic to apply the backbone of its latest Flux model to robotics, using the model's learned physics to generate synthetic training data."

Mimic

Description: Robotics company.
Why mentioned: Partner on the Flux-for-robotics synthetic training data work.
Quotes: See Black Forest Labs quote above.

Coinbase

Description: Crypto exchange.
Why mentioned: Where Gür joined in 2016 at about 100 employees and drew inspiration for hiring via its crypto knowledge quiz; also where the founders worked on ML infrastructure.
Quotes: > "If you didn't hire them, they would be doing this anyways."

Amazon

Description: Technology company.
Why mentioned: Where the founders previously built machine learning infrastructure.
Quotes: > "...founded the company in 2021 after working on machine learning infrastructure at Coinbase and Amazon."

Snowflake and Databricks

Description: Cloud data platform companies.
Why mentioned: Inspiration for fal's original 2021 thesis of owning compute for Python data science and ML workloads.
Quotes: > "The founders saw Snowflake and Databricks as compute owners running SQL in the cloud and set out to do the same for Python data science and machine learning workloads."

Dalida

Description: Eastern Mediterranean restaurant in San Francisco run by chefs Sayat and Laura Ozyilmaz.
Why mentioned: Interview location with personal significance to the founders.
Quotes: > "Yurtseven got married at Dalida with Gür officiating."

4. People Identified

Burkay Gür

Description: Co-founder and CEO of fal.
Why mentioned: Primary source of the PMF framework, margin thesis, hiring philosophy, and Hollywood observations.
Quotes: > "As you do more rounds, I think it's irrelevant. It's the business that matters more."

Görkem Yurtseven

Description: Co-founder and CTO of fal.
Why mentioned: Source of insights on pricing/demand, super ICs, world models, and ad testing loops.
Quotes: > "Basically you want your management layer to be super ICs. So people who are just so epic at their job, they can actually do hands-on work. And then they can also lead teams."

Molly O'Shea

Description: Author of Sourcery and interviewer.
Why mentioned: Conducted the interview and broke the funding news.
Quotes: > "I sat down w/ Co-Founders: CEO Burkay Gür... & CTO Görkem Yurtseven..."

Michelle Wu

Description: fal's Storyteller.
Why mentioned: Hired into a dedicated narrative role as part of fal's brand strategy.
Quotes: > "fal hired Michelle Wu as its Storyteller..."

Tina Sang

Description: fal Marketing Lead.
Why mentioned: Hosts long-form conversations on creativity in AI through fal Studios.
Quotes: > "...Marketing Lead Tina Sang hosts long-form conversations on creativity in AI through fal Studios."

Hyun Jung Lee

Description: Designer at fal design lab.
Why mentioned: Co-designed the "Your creativity layer" out-of-home campaign.
Quotes: > "Designed at fal design lab / with Hyun Jung Lee" (from Adam Ho's post)

Adam Ho

Description: fal team member who posted about the campaign.
Why mentioned: Shared the billboard campaign.
Quotes: > "Out-of-home campaign for @fal : Your creativity layer. Really proud of the whole team for getting this out!"

Sayat and Laura Ozyilmaz

Description: Chefs running Dalida.
Why mentioned: Hosts of the interview venue.
Quotes: > "...run by chefs Sayat and Laura Ozyilmaz."

5. Operating Insights

Flatten the org around "super ICs"

fal runs at most two reporting levels with no pure managers, with senior individual contributors both building and leading. Gür expects this to hold as long as coding agents improve, with each technical leader supervising 5–10 engineers. Yurtseven estimates coding agents make the engineering team 3–4X faster, equivalent to roughly 4X the headcount.

"Basically you want your management layer to be super ICs. So people who are just so epic at their job, they can actually do hands-on work. And then they can also lead teams."

Hire for obsession, and set intensity expectations upfront

Screening targets agency, hunger, and humility, with engineers expected to work directly with customers and make business decisions. Candidates are told to expect 60–80 hour crunch weeks, though the team generally does not work weekends.

"If you didn't hire them, they would be doing this anyways."

Pivot toward revenue signal, and secure compute ahead of demand

fal abandoned its original Python data infrastructure business after a year with no revenue and moved to image models where revenue accelerated. In the current shortage, it forecasts customer launch timing and procures capacity in advance.

"With this other business, we ran it for a year, we didn't make a single dollar. With this one, our revenue's starting to go like this. There's clearly something here."

6. Overlooked Insights

Enterprise trust is a hidden advantage of hosting inference

Studios often feed unreleased content into models as inputs. Enterprise customers place more trust in the process when they know inference runs on fal's GPUs, a data-handling moat that could matter as adoption moves from consumer experimentation to IP-sensitive production work.

Coding is merging into generative media

Yurtseven noted that developers on fal are building programmable video and richer applications, suggesting the two PMFs are converging rather than staying separate.

"...coding is becoming part of generative media as developers on fal build programmable video and richer applications."