Barry Diller recently made waves by arguing that physical assets and deep back-catalogs—like those owned by MGM—are suddenly worth a premium because AI companies need them for training data. It is a comforting narrative for aging media moguls. It suggests that a century of celluloid is a golden shield against digital disruption.
It is also completely wrong.
The belief that legacy film libraries are an unassailable goldmine for artificial intelligence models is the current lazy consensus of Hollywood. Wall Street is buying into it, executives are banking on it, and boards are pricing their assets based on a fundamental misunderstanding of how generative technology actually scales.
Having spent two decades analyzing media acquisitions and tech integrations, I can tell you exactly where this thesis falls apart. Hollywood is preparing to sell water to a tech industry that is already learning how to build desalination plants. The value of physical libraries is not skyrocketing; it is about to hit a hard ceiling.
The Myth of the Infinite Training Premium
The core argument for the high valuation of studios like MGM rests on a simple premise: AI companies have run out of public internet data, so they must pay ransom fees to Hollywood for high-quality video and scripts.
This assumes that AI development requires a continuous, linear consumption of copyrighted human media to improve. It treats large language and video models like traditional distribution channels—as if OpenAI or Runway are just newer, hungrier versions of Netflix.
They are not. Look at how data utilization is actually evolving.
First, synthetic data generation is rapidly displacing the need for raw, historical human inputs. Models are increasingly trained on data generated by other, highly optimized models, filtered through rigorous reward systems.
Second, the structural utility of a 1950s cinematic masterpiece to a video generation model is vastly overrated. A model does not need to ingest all 20-plus James Bond movies to understand how a sleek car moves through a European street or how explosions behave under physical laws. It needs high-fidelity spatial data, multi-angle telemetry, and clean, high-resolution physics simulations.
When tech companies license studio data, they are not buying the "art" or the "intellectual property." They are buying structured text-to-video alignment. Once a model understands the relationship between a complex prompt and a photorealistic 3D environment, the utility of buying another 5,000 hours of B-movies drops to near zero.
Imagine a scenario where a tech giant pays $500 million for a multi-year studio license. They scrape the structural patterns, map the cinematic lighting configurations, and train their foundational nodes. Year three arrives. Do they renew that license for another $500 million? No. The model has already mastered those weights. The data is depreciated.
Dismantling the Copyright Trap
The second pillar of the Diller thesis is legal protection. The assumption is that strict copyright enforcement will force tech companies to keep their checkbooks open forever.
This misjudges the history of technology platforms and fair use doctrine. While early lawsuits from authors, visual artists, and music labels look menacing on paper, the legal system historically favors transformative utility over total protectionism.
Even if courts rule that tech companies must pay for training data, the financial windfalls will not distribute evenly. The legal precedents being set right now point toward one-off licensing settlements, not perpetual royalty streams.
Furthermore, Hollywood's libraries are messy. A studio may own the distribution rights to a film, but the underlying rights—music cues, talent likenesses, literary source material—are a tangled web of union contracts and legacy estate agreements. If a tech company trains a model on an MGM film and that model later outputs a character that vaguely resembles a specific actor or uses a specific musical cadence, the legal liability ripples outward.
Tech companies are realizing that licensing a messy, legally complicated 80-year-old catalog is a massive headache compared to hiring thousands of creators to build clean, fully cleared, bespoke training environments from scratch.
The Real Power Shift: Distribution vs. Compute
Hollywood has always survived by controlling distribution bottlenecks. First it was theaters, then cable packages, then proprietary streaming apps. The studio executive's job was to sit at the gate and toll the creators and the viewers.
AI completely removes the bottleneck of production capability, which means the bottleneck shifts entirely to compute infrastructure and user attention interfaces.
If anyone can generate a feature-length, photorealistic film tailored specifically to their personal taste in real time, the value of a static library of old movies plummets. Why watch a generic 1990s action movie from a studio vault when you can prompt a system to create an entirely original, highly engaging narrative that reacts to your choices, featuring characters optimized for your psychological profile?
The current valuation models for studios do not account for this shift in consumer behavior. They assume that audiences will always want to consume linear, pre-baked content.
Let's address the flawed premise behind the standard industry questions.
People Also Ask: "How much is the MGM library worth to an AI company?"
The Real Answer: Far less than it was worth twelve months ago. Its value is front-loaded. Once the foundational models extract the structural physics and linguistic patterns from high-quality film, the residual value of that data for training purposes approaches zero. It is a non-renewable resource for the tech sector.
💡 You might also like: The Bloodline and the Bottom Line Inside the Mango Homicide Scandal
The Vulnerability of the Contrarian Take
To be fair, there is a distinct risk to dismissing the value of these physical assets entirely. If the legal landscape shifts into absolute protectionism—where even the outputted style of a studio's catalog is deemed a copyright violation—then tech platforms will be forced into permanent revenue-share models with legacy media.
If the Supreme Court rules that any model that has ever ingested a copyrighted image or video must be completely destroyed, then Hollywood wins a temporary lease on life.
But betting on the legal system to freeze technological progress is a strategy that failed the music industry in the Napster era, failed print media in the Google era, and will fail the film industry now. Progress moves faster than precedent.
Stop Hoarding Libraries, Build Foundations
If media companies want to survive, they need to stop acting like museums hoarding artifacts for tech companies to study. They need to change their operational model completely.
Instead of licensing content out to be devoured and forgotten, studios should be building proprietary, specialized micro-models trained exclusively on their unique IP.
- Own the Persona: Don't sell the rights to train on a character; build a sovereign model of that character that can interact with audiences across games, films, and virtual environments.
- Monetize the Style: Turn a director’s unique cinematic style into a licensed, secure AI plugin that creators can rent, rather than letting a third-party engine replicate it for free.
- Clean the Data Immediately: Spend capital fixing the underlying rights of the back-catalog so it can be deployed instantly in interactive environments, rather than waiting for a lawsuit to stall production.
The belief that physical assets are more valuable in the AI age is a comforting illusion designed to prop up legacy stock prices. Tech companies are not buying Hollywood's past to preserve it; they are buying it to learn how to replace it. Once the lesson is over, the classroom is useless.
