What You Could Have Withheld

Fal doubled its valuation in the months it had to turn customers away. A price pays you for what you could refuse, which is not what you gave.

Share

"We hit a point in summer where we were not able to serve our customers and had to rate limit them." That is Burkay Gur, chief executive of fal, describing to Bloomberg the season his company ran out of computing power. On Tuesday fal announced it had raised $1.3 billion at a $16 billion valuation, in a round led by Bessemer Venture Partners with Nvidia taking part. In April it had been valued at $8 billion. Last December, $4.5 billion. It has 185 employees, and according to a person familiar with its finances it passed $800 million in annualized revenue earlier this year.

What fal sells is speed. Studios, ad agencies and game developers send it a prompt and get back an image, a clip or a sound, generated by a model running on fal's machines. Most of those models were built by someone else: Black Forest Labs' Flux, Google's Veo, Kuaishou's Kling, Alibaba's Wan, whose earlier versions Alibaba released under an open license for anyone to use. Fal tunes some of them to run faster, and Gur says it has no near-term plan to pretrain a foundation model of its own.

So who made the eight billion dollars that appeared between April and November? The usual way to answer is to list contributors and argue about the weights. The founders saw generative media coming in 2021. The engineers made it fast. The model labs made the things people actually want. Nvidia made the chips. The customers brought the demand, and the moment brought the customers. Then someone says "luck," someone else says "you didn't build that," and everyone goes home with the opinion they arrived with.

I think the list is the wrong instrument, because a price was never trying to measure contribution. It measures something narrower: what the world would have to do without if you refused. Fal's value doubled during the very months it was saying no to its own customers, and that is not a paradox. The queue was the proof. Every customer waiting in it was evidence that fal held something others could not easily get elsewhere. A price pays you for what you could have withheld, not for what you gave.

Seen that way, the open-weight model makers are not underpaid by some mistake in the market. They chose a form of giving that cannot be withheld, and a price has no way to see what cannot be withheld. Their share of fal's eight billion is zero by design. The same logic reaches further down than anyone tends to notice. The researcher who published the technique, the engineer who maintains the library everyone imports, the early user whose bug report saved a month: all of them contributed something real and none of them can stand in front of the door.

The trouble begins when the price's answer leaks into the private feeling of what is "mine." Ask people what share of the value they create feels like theirs, and the honest answers tend to follow their leverage rather than their effort. The founder with equity feels the whole curve is her doing. The salaried engineer whose kernel optimization made the summer possible feels, at most, that she did her job. Neither feeling tracks what either of them gave. Both track who could have said no.

The best reply to all this comes from the investor that wrote the check. Fal, Bessemer's announcement says, "has spent years on this exact problem, all the way down to the kernel." Being hard to replace is not a lottery ticket. It is built, slowly, by people who chose an unglamorous layer of the stack before anyone cared about it. A company of 185 people that becomes the place where everyone's models run fastest has done something most companies of 185 people never do. To call its reward "mere leverage" sneers at the most honest form of work a market offers, which is making yourself necessary by being good. And the open-weight labs were not tricked. They gave in the open because they wanted reach, and reach is what they got.

Most of that is right. Making yourself hard to replace is real work, and the people who did it at fal earned something. But it narrows what they earned to a particular slice of time. The kernels were written before the summer. The decision to bet on images and video before the boom was made in 2021. What doubled between April and November was not the work; it was how badly the world wanted the work at once, more than there were chips to serve it. Bessemer itself credits the company for moving fast "to meet the demands of this moment," and the phrase gives the game away. The moment is not something anyone built.

So I would split the claim. The part of value that is yours is what you made before you knew it would be scarce. The scarcity itself belongs to no one. Market prices fold the two together into one number, and most of us fold them together in our own heads too, because it is far more pleasant to read a surge in demand as a verdict on your character. That is the habit I distrust in myself. When something I made suddenly becomes worth more, the question worth asking is whether I made more, or whether more people suddenly could not do without me. The second is a fact about them.

None of this tells fal what to do with its money, and it shouldn't. The valuation is a forecast, and forecasts are allowed to be fortunate. What it can tell anyone is how to read their own good year. Somewhere this summer a small studio sat in fal's queue, waiting for a video model someone else had trained, to run on a chip someone else had designed. The wait was the most valuable thing in the building, and fal had only built the building.

— The Andes