Fal announces Series E and F to expand capacity and custom AI models

That requires spending on both serving workloads and research. The announcement specifies work across 2027, but gives no capacity target, deployment schedule or division of spending between the two.

Disclosed individual funding rounds in 2025, as reported by fal and Bessemer Venture Partners. These historical amounts do not establish the size of Series E or F. Source: Bessemer Venture Partners — Series B investment.

A compute business that found its media market

Fal’s specialization followed an earlier, broader ambition. In its July 2025 account of its development, the company said it began in 2021 intending to scale computing for Python, then found its opportunity in speeding up image-generation models such as Stable Diffusion.

That announcement disclosed a $125 million Series C and described plans for a larger graphics-processing-unit fleet. It also introduced fal Serverless, infrastructure for teams deploying custom generative-media models. Hosting customers’ own models thus predates the newly announced rounds.

A February 12, 2025 investment statement described a $49 million Series B. Bessemer said then that fal handled more than 100 million inference requests daily and served more than one million developers. These are historical investor-provided measures, not current usage figures; the statement does not define how many developers were active or paying.

Bessemer now describes a platform combining creator workflow tools, a common application programming interface for accessing models, custom infrastructure and post-trained model development. It names Adobe, Canva and Shopify among companies that have moved production volume to fal, without specifying spending or the proportion of their workloads involved.

Custom models add a stake beyond distribution

Bessemer says fal Research released H3 Max in August, a post-trained version of MiniMax’s H3 model, with the serving path designed alongside the model. It also says fal partnered with Black Forest Labs, the developer of the Flux image-model family, on Flux 3 Max last week. Enterprises can commission post-trained versions of models exclusively on fal, the investor says.

Post-training adapts an existing model. These projects expand fal’s contribution beyond running a lab’s output, while still relying on the underlying model developers. The announcements do not spell out ownership, portability or commercial terms for commissioned versions.

Access to media models alone is an established offering elsewhere. Competing model-hosting service Replicate offered Black Forest Labs’ FLUX1.1 [pro] in October 2024, alongside the existing FLUX.1 [pro] model accessible through its API. That earlier service is not a comparison with fal’s new customized models, but it shows why distribution by itself cannot establish exclusivity or a technical advantage.

Bessemer argues that image and video generation require different memory, compute and latency handling from text generation, and credits fal’s specialized infrastructure with low costs and high throughput. Its investment statement supplies no like-for-like measurements to test those advantages. Costs for different models, output settings or serving configurations would not settle that comparison.

Replicate’s company-reported October 2024 FLUX1.1 [pro] demonstration: prompt upsampling above, the prompt “bird” below. This is not a comparison with fal’s new models. Image: Replicate. Source: Replicate — FLUX1.1 announcement.

What the 2027 spending must demonstrate

Fal’s next task is to turn the funded expansion into sustained production use. More capacity and model research could reinforce each other if improved models attract workloads that justify the resources needed to serve them; the announcements do not establish those economics.

The evidence needed is specific: current usage, revenue and margins, comparable serving costs, and the rights customers receive when commissioning an exclusive model. Those disclosures would help customers and investors judge whether fal’s broader role offers a durable advantage over choosing another model host or deploying their own models.