OpenAI is reportedly exploring a path to resurrect its video generation model, Sora, months after shelving the project. According to a report from The Information, the artificial intelligence company is in talks with a stealth startup that aims to adapt the model for the filmmaking industry. The unverified negotiations would grant the new venture access to tweak Sora and build commercial software or services around it.
The startup at the center of the talks brings together a notable mix of entertainment and technology figures. Co-founded in recent months, the unnamed company is reportedly led by Hollywood executive Jeffrey Katzenberg, former Dropbox chief financial officer Sujay Jaswa, and Bill Peebles, who previously served as the research lead for Sora at OpenAI. The potential partnership emerges after OpenAI previously halted development on both the core model and a related consumer application, citing the need to cull server-intensive side projects.
The compute calculus of video generation
The initial decision to pause Sora highlighted the severe infrastructure constraints facing generative artificial intelligence developers. Video generation requires exponentially more compute power than text or static image processing, forcing companies to prioritize their most lucrative enterprise and consumer text models. OpenAI, the Microsoft-backed developer behind ChatGPT, has increasingly focused its internal resources on its core foundational models. By potentially licensing Sora to an external entity, OpenAI could offload the immense capital expenditure required to refine and serve video models while still maintaining a foothold in the emerging medium.
For the stealth startup, securing access to Sora’s underlying architecture would provide a significant technological head start. Rather than training a foundational video model from scratch—a process requiring thousands of specialized graphics processing units—the company could focus its resources on fine-tuning the system for specific industry applications. The involvement of Katzenberg and Peebles suggests a targeted approach aimed directly at professional filmmakers, a demographic that requires high-fidelity outputs and precise control mechanisms that general-purpose consumer models often lack.
Navigating a crowded visual frontier
The reported talks occur against a backdrop of intensifying competition in the visual AI sector. While OpenAI paused its video efforts, competitors have aggressively pushed forward. Google has continued to develop its own video generation capabilities, and xAI remains active in the space. Concurrently, Meta, the parent company of Facebook and Instagram, recently unveiled new AI models designed to challenge both OpenAI and Anthropic across various modalities. This persistent competitive pressure means OpenAI cannot afford to abandon video entirely, making a strategic licensing agreement a pragmatic compromise.
Structuring the revival of Sora through a specialized third party also insulates OpenAI from the complex copyright and labor tensions currently roiling the entertainment industry. Hollywood has been particularly defensive regarding generative AI, with unions and studios clashing over the use of synthetic media. A standalone startup, particularly one guided by an established industry veteran like Katzenberg, may be better positioned to navigate these delicate commercial and legal negotiations than a broad foundational model provider.
Whether the talks materialize into a formal agreement remains unconfirmed, but the negotiations illustrate a potential shift in how foundational AI companies manage their peripheral research. As the cost of training and operating multimodal systems grows, the industry may see more developers spinning out or licensing specialized models to external ventures, allowing them to monetize breakthroughs without bearing the full operational burden.
With reporting from The Information.
Source · The Information
