A startup with paying customers is not usually a thing its founders volunteer to abandon. Early on, Burkay Gur and Gorkem Yurtseven had one: a tool for data teams working with Python and large sets of company data. Their early investor, Todd Jackson, could see it was working. The two founders could see something else. New image models were arriving, and the heavy lift was shifting from preparing data to making finished models run quickly enough for people to use. For a while, they tried to run both businesses. Their website described one future while their experiments pointed to another. Eventually they said goodbye to the customers they had and followed the faster signal.
The name on the door survived. The meaning of the work changed. Fal had begun as “features and labels,” terms from machine learning. It also happened to be Turkish for fortune telling, an inside joke appreciated by a company with many Turkish employees. A little foresight was now required: Gur and Yurtseven had to back a market that most people still treated as a toy.
The conversation before the company
Their partnership began as a friendship. Both grew up in Turkey, studied in the United States and ended up in San Francisco, where they met through mutual friends in the early 2010s. Gur worked at Oracle and later Coinbase; Yurtseven worked at Amazon. During the COVID lockdowns, they spent a couple of months sharing a house in Palm Springs. They talked through ideas while still holding their jobs. Gur left Coinbase first; Yurtseven followed several months later. Fal was founded in 2021.
There is something pleasantly ordinary in that beginning. No laboratory revelation, no staged light-bulb moment. Two friends had time to talk, and they had each spent years close to the sorts of technical problems that turn a clever demo into an actual service. At Coinbase, Gur had dealt with the infrastructure needed to use machine learning in a business that was continuously fighting fraud. The original fal idea followed from that experience: help companies build the pipelines that train models. The founders were looking for a developer product, even before they knew precisely which developer problem would matter most.
Gur had studied electrical engineering and computer science at MIT, including work at the Computer Science and Artificial Intelligence Laboratory. Those credentials help explain his technical fluency, though a degree alone would not have predicted the turn fal took. It was a change in the market, paired with an appetite to revise a plan, that made the difference.
When the demo stopped being a toy
Stable Diffusion's arrival sharpened the founders' thinking. Rather than asking every developer to collect data, train a model and manage its hardware, fal could give them quick access to models that already existed. This is inference: running a trained model to produce an answer, an image, a piece of audio or a video. A developer can then build an app around that capability. Fal made those models available through APIs and worked on the less glamorous details of speed, reliability and cost.
The first direction had some revenue, which made the choice awkward. Yurtseven later recalled that they tried to persuade themselves it was only a small adjustment: still compute in the cloud, just a different workload. It was more than that. Customers of a data product and developers making new media applications wanted different things. The founders kept both offerings alive briefly, then committed to the new one. Their early backers had funded the previous plan, and raising the next round was difficult. Conviction does not make a pivot tidy.
“We saw the signal and chased it.”Burkay Gur, recalling fal's pivot
A demo for Jackson made the shift tangible. The team turned his face into George Clooney at a speed that made still-image generation look almost like video. It was funny, and the joke did useful work. A person could feel the technical improvement without reading a benchmark. The GPU is seldom a charming storyteller; a familiar face becoming a movie star is.
The business underneath the trick was specific. Fal did not need every model in the world to be its own invention. It needed developers to reach many models, quickly and predictably, with fewer hours spent arranging servers. In an August 2025 conversation with Andreessen Horowitz, Gur described the early company as working around scarce GPU capacity. Constraints pushed the engineers to tune the machinery. Speed became both an operating habit and the product's selling point.
A platform for other people's ideas
Gur and Yurtseven had an early principle that held through the change: build for developers. The bet was that one useful infrastructure product could multiply through the applications other people made. A design app, a video editor and an advertising tool might all need fast access to different models. Fal could be the common plumbing without deciding which creative work should win.
This explains a second choice: remain flexible about models. Video, image, audio and editing systems improve at different rates. One model can be excellent on Monday and less compelling after Thursday's release. Fal responded by adding many models, including open and closed ones, and by trying to make fresh releases available quickly. Gur told an investor gathering that being ready on launch day eventually brought model makers to fal in advance. A product decision had become a way to earn attention.
The approach also puts the company in a curious position. It benefits from model makers competing, and it must keep up with them all. Gur has spoken about the speed of that competition in weeks. In 2026, fal's model catalog passed a thousand; at the Upfront Summit in February, Gur said the company was adding about ten models a week. The numbers describe a demanding editorial job as much as an engineering one: which model is worth putting in front of a developer today?
The hiring restraint behind the speed
The early team remained small. For roughly a year and a half, six people formed its core. Later, the founders hired against particular bottlenecks: first the work of getting more models onto the platform, then the work of staying close to fast-growing enterprise customers. In a May 2026 investor account, Gur admitted the sales team had lagged behind demand. This is a less photogenic startup metric than a valuation, but a more revealing one. The company had to learn when an engineer's quick answer to a customer would no longer scale.
The founders also cultivated a developer culture with a sense of humor. At one conference, the team handed out caps marked “GPU Rich” and “GPU Poor,” borrowing a joke from the AI hardware conversation. The supposedly less desirable “GPU Poor” caps disappeared first. Engineers can usually be trusted to find the joke in a shortage, especially if they have spent months working around one.
By December 2025, fal announced a $140 million Series D, with Sequoia, Kleiner Perkins and NVIDIA joining its backers. Gur said the funding would support a global scale-up and a new fund for companies building in generative media. The team, he wrote, had reached 70 people. These are company milestones, not a personal scorecard. They do show how far the founders' once awkward decision had traveled.
The pace continued into 2026. Fal introduced H3 Max in August, a video model its own research team had post-trained and its inference team had optimized. In September, Gur welcomed the founders of Lucent after fal acquired their creative AI company. That purchase hinted at a widening question for the platform: when model quality grows more similar, how much value sits in helping people direct several models toward one finished piece of work? Gur had once chosen to stay below the application layer. The addition of a team that worked on creative direction suggests fal is studying the layer above its APIs too. The decision is still young, but it belongs to the same pattern: watch the work customers are trying to finish, then adjust the machinery.
A human on the other side of the API
At the 2026 Upfront Summit, Gur argued that generative media could become larger than language models over the next five to ten years. His reasoning was pointed: a great deal of language-model work may pass between machines, but media still reaches a person. He also pushed back on the idea that software alone erases craft. Better tools can widen access; experience still shapes the result. He said he would rather see the output of a veteran using the tools than the output of someone newly arrived with the same tools.
It is a useful ending for a story that began with a technology decision. The infrastructure is there to shorten the distance from an idea to a usable image, scene or sound. Gur's job is to make that distance smaller while the models, the customers and the definition of “good enough” keep moving. Fal's first product taught its founders that a workable plan can have an expiry date. Their second asks them to notice the next one early.