Home Screens Got Free and IMAX Grew 40%. There Is a Lesson in That for AI Products

The conventional story about cinema is that streaming killed it. Free-at-the-margin content, an enormous library, a large screen in your own house, no ticket price and no travel.

IMAX grossed $1.28 billion worldwide in 2025 — up 40% on 2024 and 13% above its 2019 record, and spent 2026 signing laser expansions across Australia, New Zealand, the southeastern United States, Vietnam and Paris. It ended March 2026 with 1,865 systems in 91 countries.

The premium format grew sharply while the commodity alternative became nearly free. That is not a paradox. It is the standard outcome, and it is about to matter a great deal to anyone building AI products.

What actually happens when a baseline goes free#

The intuitive model is that free competition destroys paid alternatives. The observed pattern is more specific: free commoditises the middle and concentrates value at both ends.

What died in cinema was not the premium experience. It was the undifferentiated one — the ordinary screen in the ordinary multiplex showing the ordinary film, which offered nothing the living room did not offer more conveniently. That proposition had no answer to “why leave the house”, so people stopped.

What survived and grew was the experience that is physically impossible to replicate at home. You cannot install a screen that size. You cannot reproduce the sound field. The differentiator is not marketing; it is a set of properties the substitute cannot have at any price.

The same restructuring happened to music (streaming free at the margin, live events at record prices), to photography, to news, and to software distribution. Each time, the mid-market was hollowed out and the ends thickened.

The AI parallel is immediate#

This month made model capability dramatically cheaper. GPT-5.6 shipped in three variants, Grok 4.5 and Muse Spark 1.1 landed, DeepSeek V4 went stable, and Kimi K3 arrived at 2.8 trillion parameters with open weights and cache-hit pricing around $0.30 per million input tokens. Capability that was frontier in January is a commodity API call in July.

Which means a very large number of AI products just had their core proposition commoditised.

If your product is a thin wrapper that takes text, sends it to a model, and returns the response, your differentiator was access to the model. Your user can now do that themselves, from a chat interface they already pay for, at a price that keeps falling. You are the ordinary screen in the ordinary multiplex.

The question every AI product team should be answering this quarter is the cinema question: what do we offer that the free substitute cannot?

The three answers that survive#

From what we see working across client portfolios, there are three durable answers, and only three.

1. Proprietary data the model cannot reach.

A general model knows the world. It does not know your organisation’s twelve years of maintenance records, your specific claim denial patterns, or which of your students historically drop out and why. That data is not on the internet and never will be.

This is the strongest position available, and it compounds — the longer you operate, the wider the gap. It is also the one that requires the least glamorous work: joining, resolving, and cleaning the operational data you already have. A model over clean proprietary data beats a better model over public data on any question specific to your business.

2. Integration into a workflow the substitute cannot enter.

A chat interface can tell a clinician what the guidance says. It cannot write into the patient record, check the formulary, route to the pharmacy, and appear in the ward round view at the right moment. That last mile is unglamorous integration work, and it is exactly what makes an AI feature part of the job rather than a separate errand.

In a Hospital Management System or a School ERP, this is nearly the whole value proposition. The insight is commodity. The insight arriving in the register screen at 08:40, already attributed to the right student, is not.

3. A guarantee the substitute will not make.

Accuracy under audit, data residency, retention control, an SLA, liability. Free general tools explicitly disclaim all of this. Regulated buyers cannot use a system that disclaims all of this. The gap between what a general model offers and what a hospital’s information governance committee requires is a market, and it is one that gets larger as models get cheaper, not smaller.

What does not survive#

Worth being blunt, because a lot of roadmaps are still built on these:

• Prompt engineering as a product. Whatever clever prompting you have wrapped up is reproducible by any user in an afternoon, and each model generation makes it less necessary.
• Being first. Six months of head start against a commodity that halves in price each quarter is not a moat.
• Model choice as differentiation. Your competitor can call the same endpoint. If your pitch is which model you use, you have told the buyer the thing they should stop paying you for.
• A better chat interface. The incumbents ship a good one for free and iterate faster than you can.

The strategic read#

IMAX’s growth is not nostalgia and it is not marketing. It is the result of holding a position that the free alternative structurally cannot occupy — and then investing hard into that position exactly when the commodity got cheaper, which is what the 2026 laser rollout is.

That timing is the actual lesson. The correct response to your inputs becoming cheap is not to cut prices and compete on the commoditised layer. It is to spend the savings deepening the thing that cannot be commoditised.

For most organisations we work with, that means the falling inference bill should be funding data engineering, integration, and evaluation infrastructure — not more model calls. Cheaper tokens make it affordable to build the differentiated version. Most teams instead use the savings to send more tokens, which buys them nothing durable.

The screen in your living room is very good now, and it is free. People are still paying a premium for the one that is 80 feet tall. Work out which one you are building.

---

If your product’s core is now a cheap API call, the differentiator has to be your data, your workflow, or your guarantee. That is where we build. Ask us which one you have.