The inference layer keeps getting repriced upward. Fal, a five-year-old startup that runs hosted image, video and audio generation models for creators and developers, has raised a $1.3 billion round led by Bessemer Venture Partners at a $16 billion valuation — double the $8 billion it commanded in April, according to a TechCrunch report citing regulatory filings and a matching Bloomberg account of the round.
An unusual set of investors joined the round: Nvidia itself, Ashton Kutcher's Sound Ventures, HOF Capital and T. Rowe Price. Fal disclosed that its Series E and F rounds together total $1.65 billion; the earlier E round was a $350 million raise led by Sequoia in April. Nvidia's participation is a data point in itself — the chipmaker now routinely anchors the financing stack of the inference platforms that buy its GPUs by the rack.
Fal's business is the plumbing of generative media: it hosts and scales open-source and proprietary models for image, video and audio generation so that developers do not have to manage GPU fleets themselves. Demand for that layer has exploded as enterprises move from training experiments to serving production traffic, with customers leaning heavily on open-weight models.
The company's revenue trajectory is the strongest signal of that shift. A person familiar with the matter told Bloomberg that annualized revenue crossed $800 million earlier this year — a company-linked figure that has not been independently audited, but which would put Fal among the fastest-scaling infrastructure startups of this cycle.
Fal is hardly alone. Baseten is reportedly seeking $2 billion at a valuation of up to $30 billion, while Fireworks AI and Modal have both raised large rounds at sharply higher marks in recent months. The caveat investors keep returning to: buying or leasing compute for hosted inference is capital-hungry, and margins across the category remain thin relative to the software businesses that sit above it.
Fal was founded in 2021 by Burkay Gur, who previously led machine learning at Coinbase and worked as an Oracle engineer, and Gorkem Yurtseven, a former Amazon developer. If the April-to-November doubling holds as a pattern, the inference layer is on track to be one of the most expensive pieces of the AI stack to buy into by the time the next round of term sheets lands.
Comments (0)
Log in to join the discussion
Log InNo comments yet