U.S. portal to Chinese video models
Jan-Erik Asplund
TL;DR: As Chinese video models took off in the West, fal and Higgsfield became their leading gateways for developers & creatives, respectively, and are now moving into each other's markets as they race to power AI agents that generate video ads. Sacra estimates fal hit $800M in annualized revenue in September 2026, up from $285M at year-end, and Higgsfield hit $1.35B in October, up from $200M. For more, check out our updated reports on Fal.ai (dataset) and Higgsfield (dataset).


We first covered Fal.ai in August 2025 at $95M/year as the inference API for image, video and audio models.
Our Higgsfield coverage started with our January 2026 interview with CEO Alex Mashrabov on orchestrating AI video models, and we followed up December 2025 as the "Pinterest for AI video" crossed $100M/year, then again at $230M ARR in January 2026.
Key points via Sacra AI:
- fal & Higgsfield have become the key U.S. gateways to Chinese video models for developers & creatives respectively, accounting for roughly 50% of all Western inference to Seedance (ByteDance), Kling (Kuaishou) & H3 (MiniMax), with Sacra estimating that fal grew from $285M in annualized revenue at the end of 2025 to $800M in September 2026 and Higgsfield from $200M to $1.35B in October 2026.
- With Replicate ($350M peak valuation) sold to Cloudflare for $57.4M (December 2025) as the cautionary tale of pure model aggregation, fal & Higgsfield have both moved into custom models, with Higgsfield post-training models like Soul 2.0 (February 2026) for style and video-editing model Genjutsu (August 2026) for 15-25x cheaper product swaps & character consistency, generating 80%+ gross margins from these models vs. 20-30% on closed third-party models.
- Where Higgsfield post-trains models, fal licenses base models from companies like MiniMax ($800M annualized revenue in August 2026) & Black Forest Labs ($96M annualized revenue in August 2025) and re-engineers how they run on their GPUs to build fal-exclusive versions like H3 Max (August 2026) with ~35x the throughput speed of MiniMax’s own endpoint, splitting revenue with those labs, alongside its core business of reselling third-party models through its API & renting out GPUs.
- While fal has served as an inference provider to Higgsfield alongside GPU clouds like Nebius (NASDAQ: NBIS), Higgsfield has now launched its own pay-per-second API (September 2026) serving 50+ models including its own to go after the same developers fal sells to, while both plug into AI assistants to enable users to generate video directly from ChatGPT & Claude, creating a frenemy dynamic not unlike Anthropic renting compute from SpaceX's xAI while the two compete on models.
- As AI video moves from one-off prompts to agents that iterate & continuously generate new variations, fal & Higgsfield are converging on the orchestration layer for those agents, starting with the first killer use case of programmatic video ads, with fal acquiring AI ad campaign builder Lucent (30K+ users, September 2026) & launching fal Agent (September 2026) as it looks to catch up to Higgsfield, which launched its autonomous ads agent Supercomputer (May 2026), grew agent users 42x in three months, and now turns a product URL into up to 100 ad variants (Marketing Studio, October 2026).



For more, check out this other research from our platform:
- Fal.ai (dataset)
- Higgsfield (dataset)
- Fal.ai at $95M/year growing 4,650% YoY
- Alex Mashrabov, CEO of Higgsfield, on orchestrating AI video models
- MiniMax (dataset)
- Kling (dataset)
- Black Forest Labs (dataset)
- OpenArt (dataset)
- Runway (dataset)
- Replicate
- Modal Labs (dataset)
- $100M/year Pinterest for AI video
- Higgsfield at $230M ARR
- OpenArt vs. Higgsfield
- $150M/year free speech Higgsfield
- Video scarcity to video abundance
