What We Know About Jeffrey Katzenberg’s Unnamed AI Video Startup

Reports describe an early-stage, still-unnamed AI video venture led by DreamWorks co-founder Jeffrey Katzenberg, former OpenAI Sora leader Bill Peebles, and former Dropbox CFO Sujay Jaswa. Its reported objective is unusually specific: develop proprietary video models for professional filmmakers. 45

That combination gives the prospective company a potentially valuable mix of technical video-AI experience, financing experience, and entertainment-industry access. But it also enters a market where infrastructure costs, creative control, copyright, and performer consent are central business questions.

What has been reported—and what has not

According to reports citing people familiar with the plans, Katzenberg, Peebles, and Jaswa intend to train their own AI video models and aim them at filmmakers. The company reportedly has spoken with prospective investors, including Andreessen Horowitz (a16z). 45

Those reports should not be read as a completed deal. The venture has not publicly announced:

• a company name;
• a product or release date;
• a valuation or financing size; or
• a confirmed investment from a16z.

The distinction matters. Speaking with a venture firm is not evidence of an investment commitment.

Why this team could matter

Peebles’ experience leading Sora gives the group direct familiarity with a prominent generative-video effort. Reporting also identifies him as a co-author of the Diffusion Transformer architecture. 4

Katzenberg brings deep studio and animation credentials as a DreamWorks co-founder, while Jaswa provides operating and financial experience from Dropbox. Katzenberg and Jaswa also founded WndrCo, their technology investment firm. 5

That background could help the startup understand professional production workflows and build relationships with studios, producers, and creators. It does not, however, eliminate the hard work of building a research, infrastructure, safety, and product organization capable of competing in video generation.

Sora’s departure creates an opening—and a warning

OpenAI discontinued Sora as a product on April 26, 2026, and says its Sora API will be discontinued on September 24, 2026. Reuters reported that operating the video app consumed significant computational resources and reduced capacity available to other teams. 2

For a filmmaker-oriented newcomer, Sora’s exit may create demand from teams looking for video-generation tools or a replacement for workflows built around the API. Yet the shutdown is also a reminder that video generation can be exceptionally resource-intensive. A new model builder will need enough capital and infrastructure to train models, serve repeat generations, and offer pricing that works for professional users.

Hollywood will judge more than output quality

A tool designed for film production must address questions that consumer video apps can sometimes defer: What data was used to train the model? Can a studio control characters, styles, shots, and access? How are outputs tracked? How are copyrighted works and performers’ likenesses protected?

The recent dispute around ByteDance’s Seedance illustrates the stakes. Disney accused ByteDance of providing Seedance with a “pirated library” of copyrighted characters, while studios also raised concerns about unauthorized content featuring recognizable characters and people. In August, ByteDance and the Motion Picture Association reached an agreement to strengthen copyright safeguards for Seedance and Seedream. 17

That agreement does not establish a universal solution for AI video. It does show that stronger intellectual-property protections are becoming a practical requirement for vendors seeking acceptance from major entertainment companies.

Katzenberg’s record offers both credibility and caution

Katzenberg’s career gives the project unusual entertainment-industry credibility. His history with DreamWorks and animation makes him a distinctive partner for a company that wants to sell creative tools into filmmaking.

His technology track record is also a reason to avoid assuming that access and capital guarantee adoption. Quibi, the short-form streaming venture he founded, shut down in 2020, roughly six months after launch despite raising about $1.75 billion. The comparison is not a forecast for this startup; it is a useful reminder that a well-connected team still needs a product that fits how people actually work.

What would make the startup differentiated?

The reported plan to train proprietary models is ambitious. A filmmaker-first company would need to make a case that it can offer more than general-purpose prompt-to-video generation.

The strongest proposition would likely combine high-quality generation with production-grade controls and credible safeguards: reliable iteration, controllable shots and characters, permissions management, clear provenance, and protections against unauthorized use of copyrighted material or likenesses. The company’s technical and Hollywood credentials may help it pursue that position. Its success will depend on whether it can deliver those capabilities at sustainable cost—and win the trust of the professionals it hopes to serve.