AI film startups open their own Hollywood studios
The AI video boom is going physical: startups that used to make clips on demand are now planting flags in Hollywood itself. Meanwhile, a Peking University research push is quietly attacking the one bottleneck that makes chatty models feel sluggish — the wait for the first word.
AI film startups are setting up studios in Hollywood, using both US and Chinese models to cut production costs and route around traditional financing. The Guardian reports a wave of AI-native film companies moving onto actual studio lots, betting the same models behind the viral-video wave — from OpenAI's video tools to Chinese labs like Qwen — can let small teams produce movies without a studio's distribution machinery or a traditional financier's checks. Filmmakers behind the move argue the economics let them take creative risks the majors won't, even as industry unions keep warning about the jobs at stake.
This is the production-side sequel to the funding boom we tracked this morning — Higgsfield raises $400M at $5.4B for AI marketing video. When capital and now real estate both flow toward AI film, the "AI studio" stops being a thought experiment and becomes a business category. The interesting tension is that these startups are deliberately source-agnostic, mixing US and Chinese models wherever they're cheapest or best — a reminder that in creative AI, the geopolitical split between model blocs means less to an operator than raw capability and price.
A Peking University–StepFun research project targets the model's first-word lag, cutting time-to-first-token by up to 93.2 percent. The paper, called TensorCast, reframes tensors — the multidimensional arrays a model shuffles around during inference — as a managed layer that sits on top of mainstream serving stacks like vLLM and SGLang. Its pitch is that today's systems treat tensor movement as an afterthought scattered across components, and a unified, programmable view of it lets optimizations flow across the whole pipeline instead of inside one part.
That's meaningfully faster time-to-first-token, the seconds-long hesitation before a model answers — the single most noticeable defect in interactive AI today. The up-to-93.2 percent drop comes from a cross-component policy the team implemented and measured. The bigger idea is architectural: treat tensor management as its own layer, the way memory management was carved out of early operating systems. It's a strong sign that the local-inference arms race is shifting from raw model quality toward serving efficiency — the part users actually feel.
What to watch: whether AI film studios can hold real production schedules, and whether TensorCast-style optimizations spread beyond vLLM and SGLang into the closed-model serving stacks most people actually use.
Do you think AI-native studios will reshape Hollywood, or is it a funding bubble wearing a director's chair? Tell us in the comments.
