Netflix bought a cost, not a product

Netflix paid $587 million for a sixteen-person AI company and then refused to sell the technology to anyone. The refusal is the interesting part, and it explains which kind of AI is shipping in film and which kind is still stuck in beta.

01The purchase

In March 2026, Netflix acquired InterPositive, a company Ben Affleck founded in 2022 and ran in stealth for four years. A filing in July put the price at $587 million. The team was sixteen people. That is roughly $37 million a head for a company with no public product, no disclosed revenue, and no customer other than the buyer.

The technology is post-production. It reads the footage from a given production, learns how that production looks, and helps with the work that comes after the shoot: lighting corrections, color, background replacement, shots that were never captured. Affleck has been careful to describe it by what it rules out. The system builds a model from the production's own material. No text prompting, nothing generated from nothing.

Then Netflix added the detail worth stopping on. It has no plans to sell the technology. Access stays inside Netflix productions and the creative partners working on them. It will not be licensed to rival studios.

02The missing business model

Nearly every large AI acquisition of the past three years has been a product acquisition. The buyer wants a thing to sell, or a team to build a thing to sell, and the price is justified by a market of strangers who might one day pay for it. That is the standard shape, and it is the shape investors know how to underwrite.

This one has no such market attached. A tool that measurably lowers post-production cost has an obvious second life. Post houses would buy it. Independent producers would buy it. Competing studios, carrying the same lighting problems and the same missing shots, would pay a great deal for it. Netflix declined that revenue before anyone got the chance to offer it.

Refusing to sell is not free. It means paying for the engineering, the support, the roadmap, and the integration work without any external revenue to offset it. Netflix looked at a version of this purchase that pays for itself and a version that does not, and took the second one. There is only one reading of that choice that holds up.

03A cost is not a product

What Netflix bought is not a product it declined to sell. It is a factor of production, and factors of production are valued differently. A product is worth the revenue it might earn from other people. A cost is worth the spend it removes from you. Which means the right comparison for InterPositive was never a software multiple. It sits next to the invoice from a visual effects vendor, the day rate on a reshoot, the week you lose when a shot does not exist.

Set $587 million against a content budget measured in the tens of billions a year and the arithmetic gets simple. Netflix has said generative workflows have already been used across roughly three hundred of its titles. A tool that shaves a real percentage off post-production across a slate that size pays for itself on the spend, not on the sale.

Licensing it would trade that advantage for a modest software business. It would also hand competitors the same cost curve, which is the entire asset. At almost any licensing price that trade is bad, and Netflix appears to have run it.

04What was actually bought

Read the description of the technology again as a description of an asset rather than a feature. Your own material. The input is footage the studio already owns and already paid to create. The system does not invent a look, it learns one that exists, from plates and dailies with a clean chain of title.

That makes the tool much less portable than software usually is. Sold to a studio with a thin catalog and four productions a year, it produces less, because the thing it feeds on is scarce there. The value scales with owned material, and Netflix owns an enormous amount of it. The acquisition is better understood as a way of converting a library into a production cost advantage than as a piece of software that happens to live in-house.

05The other stack

There is a second kind of AI in this market, built on the opposite premise. Deep Agency and companies like it sell synthetic models: photographic humans generated from nothing, hired the way a brand once hired a person and a studio for a shoot. No production, no plates, no source footage. The model is the origin of the image rather than a compression of something already filmed.

The two get discussed as one story. They differ on the question that actually governs whether a frame can ship.

• Provenance. One trains on assets the buyer owns outright. The other produces output whose only ancestry is the model.
• Clearance. A background rebuilt from the production's own plates raises no consent question. A synthetic human raises every one at once.
• Insurability. Studio releases carry errors-and-omissions cover. Underwriters price what can be traced, and discount what cannot.

The 2026 SAG-AFTRA agreement makes the asymmetry concrete. Digital replicas of real performers require explicit signed consent and detailed compensation. Synthetic performers carry penalty terms and have to justify themselves as adding value rather than saving money. Studios owe notice when performer data is licensed onward for training. None of that attaches to relighting a shot you filmed yourself.

This is why one stack is running across three hundred titles while the other sits in closed beta. The gap has nothing to do with model quality. What holds synthetic humans back is not whether the image convinces anyone, it is whether anyone will insure a release built on it. Expect the premium tier of this market to be sold on custody and clearance rather than capability. And expect the underwriter, not the union, to set the actual pace.

06The week before

One week before the InterPositive announcement, Netflix walked away from Warner Bros. Paramount took it at $31 a share, with the cable assets and CNN attached. Netflix had the option to raise and did not. Seven days later it spent $587 million on post-production tooling.

Those look like opposite decisions and they are the same one. Netflix declined to buy supply and paid to lower the cost of producing supply. For twenty years the streaming thesis was accumulation: buy the library, rent the library, win on catalog. That thesis assumed content was scarce. It is not scarce now. The scarce thing is an hour of watchable content produced cheaply enough to survive its own viewing figures, and that is a cost problem, not an acquisition problem.

Whether Paramount overpaid depends entirely on which of those two views turns out to be right.

07If you are not a studio

The generalizable part has nothing to do with film. Netflix took material it already owned, which no vendor could sell to anyone else, and built a narrow system on top of it that lowered the cost of the work it does every day. Then it kept the system. That sequence is available to a lot of companies that do not think of themselves as having data.

Most operating businesses are sitting on material nobody has ever modeled: years of call recordings, ticket histories, inspection photographs, quotes that were won and lost, closed files with the reasoning still in them. A system built on that material is hard to copy, because the input is not for sale. A system built on a generic model with no proprietary input is a subscription your competitor can also buy this afternoon.

The second lesson is the one most people will skip. Having built the thing, Netflix did not productize it. Not every internal advantage should become a business line, and the instinct to sell what works often destroys the reason it worked.

If you want an honest read on what your own material is worth, tell us what you have and what it costs you today. We will tell you whether there is a system in it.