Inside fal’s $1.65B Series E & F Led by Sequoia & Bessemer

fal sells the compute beneath generative media's boom

fal sells the infrastructure that runs generative media. It began in 2021 as a compute company, pivoted to serving image models faster than anyone else in the cloud, and now runs inference for open and closed model labs while studios call its API. The bet is that generative media is a third product-market fit alongside search and coding agents, and that a small team of senior engineers can serve every lab at once.

Takeaways

• Generative media is one of AI's top fits. Yurtseven puts generative media alongside search and coding agents as AI's product-market fits. A single person generating video can spend $10,000 to $20,000 a month, which is why the category draws the spending.
• Higher prices revealed effectively infinite demand. Veo 2 launched at $5 per five-second clip when Kling and MiniMax charged 30 to 40 cents. Gür expected pushback and found none. Cost rises exponentially with resolution and duration, so better models keep finding buyers.
• Video peaked, then images came back. Video reached roughly 70% of fal's revenue before image editing models revived images and returned the mix to about 50/50. Editing also helps video models, which interpolate between frames.
• Two levels and no peer managers. fal keeps two levels of reporting and no peer managers, with senior individuals leading five to ten engineers. The founders say coding agents make that sustainable and compare it with AppLovin's roughly 400 employees and $250bn peak market cap.
• They hire for agency and low ego. Yurtseven wants engineers who make business decisions and work directly with customers. Gür filters for humility and warns candidates about 60 to 80 hour weeks. Gür joined Coinbase in 2016, when it was nearly 100 people.
• Hollywood has not been disrupted yet. Yurtseven says AI has had zero impact on Hollywood labour and has barely produced worthwhile content there, though fal's revenue from studios is growing. AWS and AMD already sell to media companies, and generative media is the next wave.

From compute company to generative media

1. Started in 2021 as a compute company, betting that data infrastructure winners such as Snowflake and Databricks own compute, and that Python would be the next SQL because AI was next.
2. ChatGPT, Stable Diffusion and DALL-E showed what that infrastructure could do.
3. After Llama, they chose not to compete on open-source token pricing.
4. Pivoted to running image models much faster in the cloud on infrastructure they had already built.
5. Coined generative media for the category, and now run inference for closed-source models as well.