Video Generation

Today's Stories

1. 1

   Wistron chief: sovereign AI expands beyond US, boosting infrastructure demand

   Wistron chairman Simon Lin said AI demand is expanding beyond the US as sovereign AI projects gain traction in more countries, marking the industry entering a new phase rather than a bubble. The shift toward sovereign AI (government-backed AI systems in individual countries) suggests wider adoption, more paid services, and a longer runway for AI infrastructure spending globally—extending the investment cycle beyond concentrated US demand.

   The transition from US-centric to distributed sovereign AI deployment may reshape which regions and companies benefit from the infrastructure buildout over the coming years.

2. 2

   Apple researchers tackle text-to-video generation with audio sync

   Apple researchers published a study addressing two core challenges in text-to-sounding-video generation—the process of creating video with synchronized audio from text descriptions. The work proposes methods to improve text conditioning and cross-modal feature interaction, tackling problems where shared captions between video and audio create interference and where gaps exist between training captions and user prompts at inference time. Text-to-sounding-video generation is a developing capability that could eventually improve creative and media workflows. By identifying and addressing fundamental technical bottlenecks—particularly how text instructions are processed and how audio and video features interact—this research may help make such systems more reliable and responsive to user intent, benefiting developers and creative professionals who rely on automated media generation.

   The study was published on Apple's Machine Learning Research site, making the findings publicly accessible. The work represents Apple's contribution to open research in multimodal AI, an area where video and audio generation remain complex technical problems requiring solutions beyond current industry approaches.

3. 3

   Soracom, Reborn as an After AI Organization in One Year, Becomes a Safe Vessel for "Token Capital

   Soracom, Reborn as an After AI Organization in One Year, Becomes a Safe Vessel for "Token Capital

4. 4

   The ROI Runway Could Be Long Outside the Tech Sector

   The ROI Runway Could Be Long Outside the Tech Sector

5. 5

   OpenAI shuts Sora video app, Disney deal scrapped

   OpenAI announced Tuesday it is shutting down the Sora app, the video generation tool the company launched in December 2024. The $1 billion（約1600億円） partnership deal with Disney, announced in December 2025 to let users create videos with Disney characters, is also being terminated. OpenAI did not confirm reports that Sora would move into ChatGPT, but promised to share timelines for the app and API soon. Sora faced steep competition from other AI video tools like Google's Veo and Luma Ray and never captured the early momentum from its launch. The sudden exit and Disney deal cancellation, just months after CEO Bob Iger was promoting the partnership in February 2026, signal internal turmoil at OpenAI—a company also dealing with wrongful death lawsuits tied to teen mental health concerns.

   OpenAI has promised to share more details soon, including timelines for preserving user work. The deal was originally set to last three years; Disney said it "respects" OpenAI's decision to exit video generation and shift priorities elsewhere.

6. 6

   Apple releases VideoFlexTok video tokenizer for AI models

   Apple published research on VideoFlexTok, a new approach to video tokenization—the process of converting raw video into a compressed representation that AI models can work with. Unlike traditional methods that treat all videos uniformly, VideoFlexTok adjusts the level of detail based on what information matters most in each part of the video. Video tokenization directly shapes what information an AI model can see and how it organizes that information. Traditional approaches force downstream models like text-to-video generators to predict all low-level pixel details regardless of the video's actual complexity, which wastes computational resources. VideoFlexTok's adaptive approach could make video AI models more efficient and capable.

   The research is published open-source on Apple's machine learning research site, making the methodology available to the broader AI research community.