It is now possible, with the toolchain available to a single person on a normal laptop, to produce work that ten years ago would have required a team. The work is not toy work. Films are coming out — short-form, mostly, but real films — assembled by individuals using language models for scripts, image generators for keyframes, video models for motion, audio models for sound, and a small amount of glue code to push the artifacts through Blender and a non-linear editor. Visual systems, generative installations, dome films, interactive worlds — all the things that used to require a production house — are appearing in the wild with one name on the credits.
The capability is real. The capability is also rented.
I want to take that observation seriously rather than treat it as another instance of the platform-versus-protocol argument. Every layer of the new toolchain sits, for almost every solo practitioner, on top of an inference API whose terms can be revised on a Tuesday. The model weights belong to a company. The usage policy decides what queries get good answers and which ones get filtered into uselessness. The credits bill arrives monthly. None of this is hidden. All of it is in the documentation, and most of it shipped with the dialog box that asked for the credit card. The interesting fact is not that the studio is rented. The interesting fact is what kind of landlord runs a rental at this layer of the production stack.
The new asymmetry is in the wrong place
The capture story most engineers know is about distribution. Search engines decide which pages get found. Social feeds decide which posts get shown. App stores decide which apps get installed. Streaming platforms decide which films get watched. Twenty years of platform critique have been about that layer: the channel between work and audience.
The new capture is below that layer. The platforms that matter now are not the ones that distribute the finished work. They are the ones that decide what the work can be made of. Whoever owns the model weights, the inference endpoints, and the usage policies owns, in effect, the language of production — what queries get good answers, what topics survive content moderation, what styles cost more credits, and what kinds of work are practical at all in a given month.
A working visual artist who builds dome films and generative installations describes the situation as: "AI does not replace the artist. It removes the production friction around the artist." That is the marketing version, and it is true as far as it goes. The harder version is that the production friction has been moved, not removed. It used to live in studio politics, equipment rental, payroll, distribution deals. It now lives in the API contract, the rate limit, the model deprecation schedule, and the policy update.
What history says about freedom-from-the-old-power
The historical pattern here is not Big Tech of the 2010s. It is older, and worse.
Venice was a free commune. By the height of the Renaissance, it was a maritime empire whose monopoly on the eastern Mediterranean was enforced with a navy. Amsterdam organised its commerce as a refusal of Habsburg overreach. Within a generation it produced the Dutch East India Company, founded in 1602 by the States General of the Netherlands and granted in its founding charter the right to wage war, make treaties with foreign rulers, coin its own money, and establish colonies — a corporation written, on paper, as a state. The Hanseatic League was a defensive merchant alliance among Northern European trading cities; it became a closed cartel that kept other merchants out for two centuries before declining. The City of London was originally a chartered liberty inside the medieval English kingdom; it is now one of the world's largest financial centres, governed under arrangements that predate Parliament.
This is not a moral argument. It is a description of the gravity that pulls on alliances of any sort. A structure organised as the answer to the previous power becomes, given enough time, the next power. The Hansa worked while it was useful as resistance. After that, it worked as itself. The pattern is consistent across enough centuries — roughly nine hundred years between William the Conqueror's 1067 charter to the City of London and the same City's current arrangements — that "communes become empires" is closer to a law than to a warning. The relevant question for any new freedom-from-the-old-power structure is not whether it will follow the same trajectory but how long the window is before it does.
What open source actually fixes, and what it doesn't
The first reflex, when somebody points out that the studio is rented, is to point at the free alternatives. They exist. Stable Diffusion is open-weights. ComfyUI is an open-source node-based interface for diffusion models, now widely used by exactly the practitioners under discussion here. llama.cpp lets a competent person run a sizable language model on their own hardware. Whisper runs locally for speech. The garage with no rent on it does exist.
The garage is not the network. ComfyUI is a workbench. It does not describe how a workflow assembled in it travels to another workbench, what license attaches to the intermediate frames, or who in a multi-tool pipeline counts as the author of the result. Hugging Face is the closest thing the field has to a shared hub for models and datasets, and is a remarkable piece of community infrastructure, and is also a single central platform owned by a single company. A kind feudal lord is still a feudal lord.
The missing layer is not better tools. The missing layer is a shared standard for routes — the path a piece of work takes through several models and several human interventions, with provenance, authorship, license, and (where applicable) value-flow attached at each step. Without that standard, every solo practitioner builds private plumbing and discovers, the first time they want to collaborate or distribute a workflow, that the plumbing does not connect to anyone else's. A network of workshops is not a network. It is a list of workshops.
What a protocol actually is, and what it isn't
The internet generation worked out, by accident more than design, what a protocol has to be in order to outlive the company that wrote it. HTTP, SMTP, RSS, BitTorrent, and more recently ActivityPub all share the same basic profile: an open specification, multiple independent implementations, no single mandatory node, the option of local execution, and the right to fork without losing compatibility. None of them are perfect. All of them are still alive after their original sponsors lost interest, which is the test that matters.
A platform pretending to be a protocol fails this test on every line. There is one implementation, run by one company, with one mandatory endpoint, no local-execution mode, and a fork would lose compatibility because compatibility is defined by what the company does this week. Sometimes the marketing copy uses the word "protocol" anyway. The test is not the marketing copy. The test is what happens when the company quits.
The current creative-AI landscape has exactly this problem at the routing layer. Almost every multi-model pipeline of any sophistication is built on top of one or two cloud APIs whose particular shape — input format, output format, parameter names, rate-limit semantics — is not specified anywhere except the vendor's documentation, and is subject to change. There is no neutral description of what a route is. There is only what a route looks like inside the systems that currently dominate the routing.
What the early internet got wrong, what AI tooling could still get right
The original internet protocols were architecturally silent on authorship, attribution, and microeconomics. This was not negligence. The use cases for those features had not surfaced yet, and the engineers building TCP/IP were not in the business of imagining what payment for a file would look like in 2016. The consequence was that the consumer internet has spent the last twenty-five years bolting authorship and payment on top of protocols that do not have a place for them: copyright enforcement at the platform layer, DRM at the application layer, ad networks at the social-graph layer, all bolted on, all leaky, all the source of a great deal of daily friction.
AI tooling has a brief opportunity to not do that. If the standard format for a creative-pipeline route — what the input was, which model touched it next, what license attaches to the result, who counts as the author of which step — is defined now, with attribution and optional value-flow as first-class fields, those features survive forever. If the standard is defined later, they are again a platform's problem to solve, and the solution again ends up looking like a platform.
The bar is not high. A YAML schema with provenance and licensing fields, a CLI validator, a reference converter from existing workflow formats, and a manifest format that travels with the output file would do more for the routing layer than any new "creator marketplace" launched in 2026 will do for solo creators. None of those four pieces require a new company. All four require somebody to write them down before the existing routing layer is ratified by sheer mass.
Why the lawsuits matter, and what they do not guarantee
Right now, in the spring of 2026, the count of major copyright lawsuits filed against AI services is past seventy, with the New York Times v. OpenAI/Microsoft case — filed December 2023 — the most prominent of the publishing-industry actions, and the Disney/NBCUniversal/DreamWorks case against Midjourney, filed June 2025 and since consolidated with Warner Bros.'s parallel September 2025 suit, the most prominent of the entertainment-industry actions. The legal phase is already shaping which kinds of training data are economically viable and which are not.
It is tempting to read this as a force pushing reliably toward open alternatives. The pressure is real, but the direction is not automatic. Lawsuits can equally produce a regulatory settlement that locks in three or four large incumbents and makes life harder for everyone else, exactly the way the YouTube content-ID architecture did for music in the 2000s. The window the legal turbulence opens is the fact of the turbulence, not the outcome of any individual case. While nobody knows which way the law will land, there is room to standardise the parts of the production layer that do not depend on the law landing one way or the other. Routing, provenance, attribution, licensing — none of these need a court ruling to specify. They need somebody to specify them now, before the post-settlement landscape decides for everyone.
The studio, again
A solo practitioner who today produces a film using six models from three vendors and a half-dozen API keys has produced a real film, on a real budget, on real hardware, with real distribution. The work is not less real because the production stack is rented. It is also not the same kind of independent work that a studio with its own equipment would have produced ten years ago. The freedom is structural in the same way the freedom of a sharecropper is structural: real on the surface, conditional on a stack of contracts underneath. Sharecropping is not slavery. It is also not yeoman ownership. The middle category is the one most solo creators have entered without noticing.
The way out of that middle category, if there is a way out, is not another platform built by people who sincerely intend to be kinder than the existing platforms. The road from kind landlord to landlord, in the historical record, is consistently shorter than the marketing copy of any new entrant suggests. The way out is a layer below the platforms — a routing protocol, a manifest format, a provenance standard — that nobody owns and nobody can quit. Whether that layer gets written before the field consolidates is the open question of the next twenty-four months.
What the studio metaphor was always describing
AI does not empower individual creators. AI offers individual creators the capabilities of a studio. Whether those capabilities arrive with the autonomy of a studio or the autonomy of a tenant depends on a layer that does not yet exist. The capabilities are not in question. The autonomy is.
The question is not whether AI will change creative work. It will. The question is who owns the ground that the change is standing on, and whether that ground is being designed now to be owned by somebody or to be owned by nobody. The historical default has occasionally been beaten by people who got the architecture right early. The window for that kind of work is open. It is not very wide. The decisions that close it are being made now, mostly by people who do not realise they are making them.
Top comments (0)