fal's 35x speedup: engineering expanding the market, not a toll
By Bram de Vries · Reporting from Amsterdam ·
fal is bypassing regulatory caution by using engineering to remove friction for Hollywood creators.
European firms often treat regulation as a fixed cost of doing business, a toll paid to keep the peace. Gorkem Yurtseven and Batuhan Taskaya of fal, speaking on the a16z podcast with Jennifer Li, challenged this view. They showed how their team optimized MiniMax's H3 Max video model by rewriting kernels and refining post-training. These changes made generation ten times faster and nearly tripled the amount of the chip's power they actually used. This is not the work of a committee; it is the work of engineers making a machine faster to meet market demand.
Why Tokens Matter
Yurtseven said the AI video industry has found its footing because professionals now pay thousands of dollars a month for continuous generation. While critics call for state-mandated caution, the real bottleneck is compute. By cutting the diffusion pipeline by more than half and mastering chips like Hopper and Blackwell, fal changed the physics of the render.
| Performance Metric | Baseline | fal Optimized | | :--- | :--- | :--- | | Diffusion Steps | 50 | 20 | | Render Time (5s video) | ~7 seconds | < 2 seconds | | Overall Speedup | 1x | 35x |
A five-second clip now renders in the time it takes to blink twice. When a firm achieves a 35x speedup without losing quality, they are not waiting for a regulatory green light; they are expanding the market through engineering.
How did Hollywood adapt?
Studios adopted these tools almost overnight. Jennifer Li noted that Hollywood usage jumped from zero to fal's fastest-growing segment in a single year. Production houses now plug these models into Blender and use custom camera controls to bypass traditional bottlenecks.
The common fear in Europe is that automation erases culture. In practice, creators are simply desperate for tools that remove friction. A director does not want a manual on algorithmic ethics; they want a JSON-described camera angle at ninety degrees that executes reliably every time. To get this, studios build custom LoRAs and host their endpoints in the US to avoid lag and oversight.
Speed is the only arbiter
Europe can continue to wrap every new enterprise in compliance tape until the capital moves to more hospitable shores, or it can accept that efficiency defines modern power. The engineers at fal did not ask permission to optimize their kernels, and the creators using their platform do not care about bureaucratic hesitation. Wealth goes to those who move the cargo, process the tokens, and optimize every cycle of the chip before the competition finishes reading the rulebook.