Sora 2: Licensing is all you need (jperla.com)

OpenAI this week unveiled Sora, but the bigger story isn’t a new model — it’s a reckoning: licensing and revenue-sharing, not neural architectures, are the hard problems. The piece argues neural networks are largely commoditized (GPT faces rivals from Anthropic, Google, Meta’s Llama, Mistral, and Chinese models), so the competitive moat can’t be the model itself. Instead, the battleground is content rights: training on vast amounts of creative work without permission creates ethical backlash and legal exposure, and admitting the need to license going forward implicitly raises questions about past training data for GPT‑4, DALL·E and others. The practical implications are severe. OpenAI’s SaaS-style margin assumptions (80–90%) clash with content businesses’ economics (Spotify ~25% gross margins, Netflix ~40%), so revenue-sharing deals could blow out unit economics and turn AI firms into low-margin media companies while exposing them to “tens of billions” in liability. The author likens the missing piece to Spotify’s painstaking licensing deals and calls for a new generation of “Spotify for AI” startups that bake reciprocity and contracts into their models. Bottom line: this is primarily a business-model and legal problem — not a technical one — and the industry must learn to negotiate rights and share value with creators or face legal, financial, and ethical blowback.