Data Marketplaces & Intermediaries

For video and audio, licensing isn't the safer path to training data — it's the only one

Source: Troveo · Sep 22, 2026

For text, licensing arrived as the responsible alternative to scraping. For video and audio, there was never a scraping option to begin with. Real-world video, professional-grade audio, gameplay, robotics telemetry and genuine business data barely exist on the open web — which means for these modalities, licensing isn't the safer route to training data, it's the only route.

The market is already forming around that reality. Music labels turned the Suno and Udio copyright fights into licensing arrangements in late 2025, converting litigation into a paid supply channel. Marketplaces like Troveo now sit between rights holders and AI labs, and Troveo reports having paid more than $20 million to over 7,000 rights holders for licensed footage and audio.

This is the core of the multimodal thesis. The most valuable data for the next generation of models is precisely the material that cannot be scraped — real people, in real settings, captured on video and audio. That data only exists if someone collects it deliberately, with the consent of the people in it and clear rights to use it. Long-form interview conversations sit squarely in that category: high-value, multimodal, and impossible to source responsibly except through licensing and consent. As labs run out of open text, the scarce asset becomes exactly the kind of rights-cleared, de-identified conversational data that has to be built rather than found.

Key Points

• Real-world video, professional audio, gameplay, robotics and business data barely exist on the open web, so scraping was never an option for these modalities
• Music labels converted the Suno/Udio copyright suits into licensing arrangements in late 2025
• Troveo reports paying $20M+ to 7,000+ rights holders for licensed footage and audio
• Directly reinforces the multimodal-interview-corpus thesis: high-value modalities must be licensed and consented, not harvested