Enterprise spend on @fal more than 4x’d over the last six months. AI-native products are still the majority of usage (and growing). But there’s a new set of companies with no prior stake in generation building the kind of layered workflows you'd expect from a specialist studio: chaining models, switching modalities inside a single job, training adapters for narrow tasks. Generative media has moved from something companies experimented with to something they build on. Compute has been the binding constraint across generative media since April, by fal's own account, and their answer was post-training MiniMax's H3 to roughly 35x the throughput of the base model at matching or better quality. Tons more interesting insights into where the gen media industry is headed in the State of Generative Media Report Vol. 2. Congrats @burkaygur, @gorkem @isidentical & team.