The Paramount Problem

In 1948 the Supreme Court told the movie studios they could not own the theaters. At the time the five majors, Paramount, MGM, Warner, Fox, and RKO, made the films, controlled how they were distributed, and owned the cinemas that showed them. They used that integrated grip to run a racket called block booking: a theater that wanted the one film audiences were actually lining up for had to agree, sight unseen, to rent a year's worth of the studio's other dreck along with it. The Court looked at the arrangement, called it what it was, and forced the studios to sell the theaters. That settlement, the Paramount Decrees, governed Hollywood for seventy-two years.

Then in 2020 the Justice Department asked a judge to throw the whole thing out. The argument was that the decrees were a relic, that the market had moved on, that streaming had made the old theater monopoly irrelevant. The judge agreed, with a two-year sunset. And here is the part that should stop you cold. The instant the rule came off, the structure it had banned reassembled itself, just with fiber instead of film reels. Netflix makes the content, owns the distribution pipe, and owns the direct relationship with the customer. So does Amazon. So does Apple, so does Disney. The exact vertical integration the Supreme Court spent a generation dismantling is now the default business model of everyone who streams video, and we call it innovation.

I bring this up because the AI industry is running the same experiment right now, at roughly ten times the speed, and oddly no one is bringing up the Paramount Decrees. In late June OpenAI revealed its first co-designed chip, stood up a consulting subsidiary to deploy its own models, and kept building its own data centers. The model vendor increasingly also owns the silicon, the cloud, the agent framework, and the people who walk into your office to install it. Every cloud provider is racing to own more of that same stack. So, while the pieces are different, the shape is looking a whole lot like the old thing.

The thing the court disagreed with was not bigness for its own sake; it was the anti-monopoly provisions set in stone a half a century earlier: tying. Using control of the one input you cannot do without to force you into taking the things you would never choose on their own merits. If there is something that destabilizes the AI stack, that's going to be it. When the company that owns the model you depend on also owns the chip it runs on, the cloud it runs in, the framework that orchestrates it, the consulting arm that deploys it, and the eval harness that grades whether it's working, the bundle stops being a convenience so significant you're willing to overlook all the downsides. If you're in a situation where you wanted the model, but now you are now renting the year's worth of dreck along with it, sight unseen, people are going to get mad.

None of this means integration is evil, and the Columbia Law Review crowd that treats every merger as a crime misses the same point the deregulators do. Vertical integration genuinely lowers transaction costs, can make the product better, and, sometimes (as with electricity right now), it is the only way to secure a critical input at all. The studios made some of the best films in the history of the medium inside the integrated system, and nobody who saw them in a studio-owned theater felt robbed in the moment. The point of the decree was never that integration is always bad. It was narrower and more durable than that: an integrated incumbent can quietly convert a great product into a captive market, and by the time enough people notice, the only available fix is a court order and thirty years of waiting.

If you build systems rather than antitrust briefs, the reason to care is that the architecture decision and the market-structure decision turn out to be the same decision wearing two hats. A modular stack, open weights you can run yourself, a data plane you control, compute that moves to where your data already lives, clean seams between layers you can actually pull apart, is not only the better engineering pattern. It is the thing that keeps you from waking up one morning inside someone else's block-booking arrangement with no exit that doesn't cost you a rewrite. Every layer you let collapse into a single vendor's bundle is a layer you have agreed, in advance, never to have an opinion about again.

Hollywood needed the Supreme Court to unbundle it because the studios were never going to do it themselves. Nobody standing inside an integrated monopoly wakes up wanting to break it apart. The companies assembling the AI stack this year won't either, and we should stop expecting them to. The only thing that keeps the layers separable is whether the people writing the checks insist on the seams while the seams still exist. Right now, mostly, they are buying the bundle and calling it the future. We have seen this movie. We even know the runtime: about seventy-two years to break the thing up, and roughly eighteen months to put it back together.

Wondering whether your AI stack has any seams left to pull apart? Check out Expanso. Or don't. Who am I to tell you what to do.

NOTE: I'm currently writing a book based on what I have seen about the real-world challenges of data preparation for machine learning, focusing on operational, compliance, and cost. I'd love to hear your thoughts!