AI media host fal raises $1.3B, doubling its valuation to $16B

fal, the San Francisco company that hosts and runs image, video and audio AI models, has raised a $1.3 billion Series F led by Bessemer Venture Partners, with Nvidia, Sound Ventures and T. Rowe Price taking part. Bloomberg reports that the deal values fal at $16 billion, double the $8 billion valuation it reached in April.

fal announced the Series F on November 3 together with an earlier Series E, and says the two rounds total $1.65 billion. Sequoia Capital led the Series E, which fal closed earlier this year and only disclosed now. Bessemer confirms in its own investment note that it is leading the Series F.

The deal: $1.3 billion now, $1.65 billion across two rounds

The headline numbers come from two places. fal states the combined $1.65 billion total itself. The $1.3 billion Series F size and the $16 billion valuation are reported by Bloomberg. fal did not break out the Series E amount, but ShiftDelete.Net calculates that the combined total implies about $350 million for that round.

Nvidia joins the Series F as a strategic investor, according to Bloomberg. A strategic investor is a company that invests partly for business reasons, such as a supplier or partner relationship, rather than only for financial return.

Item	Detail
Series F size	$1.3 billion
Lead investor	Bessemer Venture Partners
Other participants	Nvidia, Sound Ventures, T. Rowe Price
Valuation	$16 billion, up from $8 billion in April (Bloomberg)
Series E lead	Sequoia Capital
Series E + F combined	$1.65 billion, per fal
Previous round	$140 million Series D, December 2025
Headcount	More than 150, up from about 70 at the Series D

The jump in staff tracks the funding. fal had about 70 employees when it raised its $140 million Series D in December 2025 and now has more than 150.

Revenue: one unnamed source

The revenue figure is the weakest-sourced number in the story. A person familiar with the matter told Bloomberg that fal reached $800 million in annualized revenue this year. TechCrunch, citing the same Bloomberg report, says fal passed that mark earlier this year. Annualized revenue usually means a recent period’s revenue extended over a full year, not revenue actually booked over twelve months. fal has not published revenue figures itself. On that basis, ShiftDelete.Net calculates that the $16 billion valuation is about 20 times the revenue figure Bloomberg cited.

What fal sells: inference for image, video and audio models

fal’s business is inference: running a model that is already trained each time someone asks it for an output, such as an image or a video clip. Training builds the model once; inference is the recurring compute cost of every request, and it is what API customers pay for.

fal works on two sides. Developers get access to thousands of models through a single API, and AI labs use fal for distribution and serverless infrastructure. Serverless here means the customer calls the model without renting or managing specific machines. Bessemer names Adobe, Canva, Shopify, World Labs and Amazon among the large companies that have moved production volume to fal.

According to TechCrunch, the round comes amid strong demand for inference services, especially from customers relying on open source models. ShiftDelete.Net adds that demand keeps rising as studios and ad agencies adopt high-resolution video models such as BytePlus Seedance 2.5.

Bessemer’s case: media is a different serving problem than text

Bessemer has backed fal since its Series B two years ago and has invested in every round since. Its investment note, signed by partners including Talia Goldberg and Sameer Dholakia, centres on a technical argument:

Media is a different infrastructure problem than text. LLMs produce one token at a time. Image and video models refine a whole frame over many passes, so they need memory, compute, and latency handled differently. The same serving stacks built for LLMs don’t carry over.

A token is the small chunk of text a language model outputs at each step. An image or video model instead works on a whole frame repeatedly, which, in Bessemer’s description, changes how memory, compute and response time need to be managed. Bessemer says fal has worked on this problem “all the way down to the kernel”, meaning the low-level code that runs on the GPU, and describes the result as infrastructure built for real-time generation at scale with low latency, low cost and high throughput. These are the lead investor’s characterisations; the sources include no independent performance measurements.

Bessemer also lists the media uses it sees growing fast: real-time ad creatives, interactive video, product imagery, game assets, music and voice, avatars and video editing. It says most of these barely existed when it first invested in fal.

From hosting models to post-training them

The more notable change for builders is that fal is no longer only a host. Bessemer describes fal’s growth over two years from an inference engine to a fuller stack: workflow tools for creators, the API, custom infrastructure on which enterprises run their own model weights, and now post-trained model development alongside the labs. Post-training means further training an existing model to improve it for specific uses.

• H3 Max (August): fal Research released a post-trained version of MiniMax H3, with the serving path designed together with the model. Bessemer says it “immediately topped all major arenas and leaderboards”. The note does not name the leaderboards or give scores, and none of the sources provides independent results.
• Flux 3 Max (last week): fal Research partnered with Black Forest Labs on the model.
• Exclusive custom models: enterprises can commission post-trained versions of models that run only on fal.

One inconsistency is worth noting. Bessemer’s note says developers reach “thousands of models” through fal’s API in one passage and describes “the API that reaches hundreds of models” in another. The sources do not resolve the difference.

Where the money goes: compute first

fal says the new capital pays for two things in 2027: more inference capacity to run models at scale, and a larger team for research and new products. Compute is the pressing need. In an interview released alongside the announcement, the founders described a compute shortage that is shaping how and when new models launch.

fal was founded in 2021 by CEO Burkay Gur and CTO Gorkem Yurtseven. TechCrunch describes Gur as a former Coinbase machine learning leader and Oracle engineer, and Yurtseven as a former developer at Amazon.

What it means for builders

For developers who call image, video or audio models through an API, the round has several practical implications, along with clear limits.

• Capacity is the bottleneck. fal’s own founders say a compute shortage affects how and when models launch. If a well-funded host is constrained, teams that depend on new video models at launch should plan for availability to vary.
• Thin margins across the sector. Inference providers including Baseten, Fireworks AI and Modal are growing fast and raising large rounds, but their gross margins remain thin because of the high cost of buying or leasing compute, according to several people familiar with their financials who spoke to TechCrunch. Pricing for media generation depends on that cost structure. ShiftDelete.Net notes that whether fal’s valuation holds will depend on securing enough GPUs while prices for video generation stay high.
• Exclusive models change the choice of provider. When enterprises can commission post-trained models that run only on fal, choosing a model can also mean choosing a host. Teams that value portability should check whether a given model variant is exclusive.
• Hosts now tune models too. fal’s work on H3 Max and with Black Forest Labs on Flux 3 Max means the line between model maker and model host is getting less clear. Performance claims for such variants currently come from fal’s backer, not from independent tests.

The money race in AI infrastructure is broad. ShiftDelete.Net reports that rival inference platform Baseten is also raising about $2 billion at a reported $30 billion valuation. For other recent large rounds and filings, see Unicorn API’s coverage of FuriosaAI’s $620 million raise and Anthropic’s IPO filing.

What to watch next

• Revenue disclosure. The $800 million annualized figure rests on one unnamed source. Any figure fal publishes itself would put the reported 20x revenue multiple on firmer ground.
• Compute deals. With Nvidia now an investor and more inference capacity planned for 2027, watch for announcements on GPU supply and how fast new models arrive on the platform.
• More post-trained models. Bessemer calls H3 Max “just the beginning”. Further fal Research releases, and independent evaluations of them, will show how far the company moves into model development.
• Rival rounds and margins. Baseten’s reported raise and the thin gross margins TechCrunch describes across inference providers will shape how the sector prices image and video generation.