How Did Fal Move From Data Infrastructure to Image Models?

fal co-founders Burkay Gür and Gorkem Yurtseven say fal started in 2021 as a compute company, before the current AI wave. The thesis was that the next data-infrastructure company had to own compute, the way Snowflake and Databricks let SQL scale in the cloud. fal bet that Python would be the next layer and worked with teams to move machine-learning jobs into the cloud.

ChatGPT, Stable Diffusion, and the early image models then forced a choice between staying in data infrastructure and entering AI. Open-source language models drew a price race toward the bottom. With five or six people and little money in the bank, selling tokens at a negative margin was off the table. Image models were where the trend still held. Video was not even being discussed. The same infrastructure could run image models much faster.

A separate effort had run for a year without making a single dollar, while image revenue was already climbing, so the focus moved there. Roughly three to six months later the trend looked much bigger, useful for ads and marketing, with buyers different from language-model buyers. The founders coined the phrase generative media, put it on the website, and told people to follow the generative media company. Three years earlier almost nobody understood the phrase. About 18 months after the first image experiments, people started thinking the tools could make feature films.