Two years on, Lionsgate and Runway have shipped nothing

Lionsgate and Runway announced an exclusive AI model trained on Lionsgate's catalogue of more than 20,000 titles. It was the most-publicised AI deal in entertainment. Two years later nothing has come out of it. No film, no series, no short, no tool built on top of it that anyone can point to.

The reporting on why is thin but specific. According to TheWrap, development ran into two walls: copyright, and data scale. A source put the second one in a single sentence. "The Lionsgate catalog is too small to create a model." That is a strange thing to say about twenty thousand films and television shows, and it is the most useful sentence published about AI in entertainment all year.

The date matters, so here it is plainly. Futurism's summary of TheWrap's reporting, alongside a Vulture interview with Lionsgate vice chairman Michael Burns, was published on 27 September 2025. We went looking for anything newer. We could not find any 2026 reporting that the model has shipped, and Runway's own Wikipedia entry carries no 2026 Lionsgate update. Silence is not proof a project is dead. It is the only evidence available, and after two years it counts for something.

Catalogue size is not a moat

The logic of the deal was clean enough to fit in a headline. A studio owns decades of finished film. A model needs film to learn from. Put the two together and the studio gets a model nobody else can have, trained on material it already paid for, producing work in a house style it already owns. Every studio with a library read that and started counting its titles.

Two things broke. The first is arithmetic. We have no published figure for the volume of material a frontier video model is trained on and will not invent one, so take the claim as the source made it: in the context of training a model from scratch, twenty thousand titles is small. A film library is also narrow in a way that a count of titles hides. One corporate taste, a handful of genres, a repeating set of camera moves, lighting conventions held steady on purpose across decades. That is an asset for a brand. It is thin diet for a model that needs variety more than it needs quality.

The second is rights, and here the reporting says only that copyright problems arose. It does not say which. Anyone who has cleared a library knows the shape of the problem. A finished film is not one asset with one owner. Performer likeness, music, third-party footage, licensed design, co-production terms drafted by people who had no reason to think about model training. A catalogue can be fully owned for the purpose of distribution and not owned at all for the purpose of training. Those are different permissions and the paperwork is decades old.

Runway is fine. The deal is what stalled.

In February 2026 Runway raised $315M at a reported $5.3bn valuation. The company is healthy. That is the part marketers tend to read as reassurance, and it is the part that proves least. A funding round is evidence that investors believe in a model company. It is not evidence that any specific thing announced on top of that company exists. Keep the two separate when a vendor's deck leads with its cap table.

The gap, repeated

Lionsgate is the loudest example, not the only one. Set aside the question of whether AI production works and simply read the announcement calendar against the release calendar.

“A catalogue can be fully owned for the purpose of distribution and not owned at all for the purpose of training.”

• WPP. Restructured into four units in February 2026, creating WPP Production with roughly 10,000 production staff. On 22 September 2026 it opened Devon's Point, a 34,000 sq ft production hub in east London housing the creative-tech studio HEX. Adweek's on-site piece names no client work produced there and gives no output figures. The restructure coverage is similarly short on output.
• Framestore. Announced an AI joint venture with iwot studios on 11 February 2026 to build a cross-media production platform, first IP The Wheel of Time. Nothing released. No launch date given.
• Asteria. Launched Continuum Suite on 11 March 2026 with Pressman Film as an initial client. It is explicitly a production operating system, not a video generator, which is a more honest product than most. Asteria has shipped: The Odd Birds Show, animated shorts released to social platforms in April 2025. Its feature Uncanny Valley, Natasha Lyonne's directorial debut, remains announced.
• Utopai Studios. Has shipped products. PAI went to public release on 5 March 2026, Utopai X on 30 September 2026. No completed film or series. Variety reports three films and two series premiering in 2027.
• Fable Studio and Showrunner. No completed series documented. Everything Is Fine was planned for 2026 and has not shipped.

Read that list as a group and the pattern is not failure. It is a category error about what the hard part is. Every one of those announcements solved for capacity: a model, a platform, a building, a headcount. None of them was an announcement about a finished piece of work with a client's name on it.

What shipped instead

Two deliveries stand up to checking. Promise, operating as The Generation Company, delivered more than 500 AI-generated shots for Harlan Coben's Final Twist Season 2, which premiered on CBS on 5 October 2026. Netflix's Q2 2026 shareholder letter states that generative workflows were used in roughly 300 of its titles in 2026, concentrated in post-production. Neither of those came from a model trained on a proprietary library. Both came from hybrid pipelines run by people who already knew how to finish things.

Four questions worth asking a vendor

This is the part of the story a brand marketer can use. The announcement economy is not going to slow down, and the only defence is boring.

• What has this specific capability produced that aired, and where can I watch it?
• Who paid for that piece, and will they say so?
• Which parts were generated and which were shot, built or composited?
• If the answer is a tool rather than a film, who finished a film with it?

Two years and twenty thousand titles have produced no releasable minutes. On 5 October, 500 AI shots went to air on CBS. The difference between those two outcomes was never the size of the library.