Closed AI products are starting to eat each other on price.
GPT-4 launched in 2023 at $30 per million input tokens. By the time you read this, the equivalent capability sits below $0.60 per million tokens — a hundred-fold collapse in under three years. The same trajectory is visible at Anthropic, Google, and inside every Chinese frontier lab. None of the labs is happy about it. None of them can stop it. The capability has commodified faster than anyone with a private investor deck was prepared for.
There is a temptation, watching this, to assume the AI age is therefore deflating. It is not. What is collapsing is the layer at which value used to be captured. The layer the labs spent ten years building. The layer where each company tried to be the place you went to talk to an AI.
When a capability commodifies, value moves up-stack. That has happened with every great commodification before this one. The interesting question is what the up-stack layer looks like for AI. The interesting answer is: it cannot be a closed product. It has to be a protocol.
This essay is about why.
What history teaches us about commodity inflection points
In 1995, the question on the table was which browser would win. Netscape and Internet Explorer were fighting a brutal war over which company would own the front door to the World Wide Web. Both companies are now archaeological footnotes. The thing that won was neither. It was HTTP — the protocol underneath both of them. HTTP did not compete with Netscape; HTTP made Netscape possible, and outlived it because it was neutral. Today HTTP runs trillions of dollars of commerce, none of it captured by any single browser vendor.
The same pattern played out with SQL. In the 2000s every database vendor wanted to own a customer's data. Oracle, IBM, Microsoft, Sybase. Each of them is still alive, but the layer that won the decade — the layer where developers spent their time and where applications became portable — was the standardized query language. ANSI SQL is owned by nobody and run by everybody.
It happened again with mobile. The browser wars repeated, this time as iOS versus Android. Closed and open coexist on the device. But the web — the protocol layer — thrives above both of them, and most of the value users get out of their phones flows through it.
This is not a coincidence. It is the structural shape of commodification. When a capability becomes cheap to provide, the value of providing it shrinks. The value moves to whoever defines how the capability is coordinated, composed, and made interoperable. Whoever owns the coordination layer wins, and the only coordination layer that can actually unify a market with multiple credible vendors is one that no single vendor owns.
That is what a protocol is.
The structural problem with closed AI products
Now apply this back to AI.
Closed AI products today share a set of properties that look like product decisions but are actually structural consequences of the business model. Try, as a thought experiment, to do any of the following with current closed AI offerings:
- Take an agent you built on platform A and run it on platform B without rewriting.
- Have your agent on platform A call a capability published by someone on platform B, and pay them for it.
- Compose three skills from three different vendors into a single workflow your agent executes autonomously.
- Resell — actually resell, with revenue accruing to you — the work your agent does for someone else's agent.
- Run the entire stack on your own machine, with your own data, when the network is down.
You cannot do any of these. Not because the engineers at the closed-AI companies are bad at their jobs. They are excellent at their jobs. You cannot do these things because the business model that funds those companies requires that you cannot. A platform whose revenue depends on you returning to its servers cannot, structurally, let you run elsewhere. A platform whose moat is the lock-in of your data cannot, structurally, let your data live somewhere else. A platform that prices its API per token cannot, structurally, let three of its competitors' tokens flow through one of your workflows and earn anyone but it.
This is not malice. It is gravity. Closed products in a commodifying market converge on the same set of constraints because those constraints are the only way to defend margins when the underlying capability is becoming free.
The user experience of this convergence is the experience of feeling that the AI ecosystem is narrowing even as the models get smarter. Better models, fewer choices about how to use them. Better autocomplete, worse interoperability.
There is only one way out of that gravity well, and it is structural.
What an open agent protocol must provide
If protocols are how value gets coordinated in a commodified market, then the question for the AI age is what an agent-native protocol has to provide that older protocols did not.
Six properties, all of which the closed products cannot deliver because their business model forbids them:
Composable units. Capability has to come in pieces that can be combined by anyone, the way a function call composes. This is why prompts and skills feel insufficient — they don't compose, they get embedded. The composable unit for agents has to be a first-class artifact, identifiable, versioned, transferable. Not a configuration string buried inside a chat app's settings panel.
Verifiable identity and lineage. When capability is composable, "where did this come from" becomes a question with safety, attribution, and economic consequences. The protocol must answer it without asking a centralized vendor. This is exactly the role public-key cryptography played for the web's identity layer, and it has to play it again for agents.
A local-first execution path. Not local-only — there will always be reasons to call cloud frontiers for the hardest problems — but local-first: the default place an agent runs is on hardware its owner controls. Sovereignty is not a feature; it is the difference between a tool you use and a tool that uses you. Closed AI products cannot give you this because their entire pricing model is built on you not having it.
Federation without lock-in. Capability built in one place must reach users in another without forced migration. Email proved this works at scale. The web proved this works at scale. AI has not yet been allowed to.
Honest fitness measurement. When capabilities compete, somebody has to measure which ones are good. The protocol cannot delegate this to any single vendor — that turns the protocol back into a closed market with extra steps. The measurement has to be observable from outside, reproducible, and resistant to the gaming that all marketing claims invite. Call this what it is: a public adjudication layer, owned by nobody.
A creator-owned economy. When the people who build capabilities can earn from them — directly, in proportion to use, without a platform deciding their cut — you get an ecosystem. When they can't, you get an extraction stack with creators serving the platform. Substack rediscovered this for writers. App stores half-rediscovered it for developers. The agent layer has not yet had its turn.
Each of these maps to a precedent: HTTP for transport, public-key cryptography for identity, RSS for federation, app stores (the good parts) for economy. None of these precedents was invented from scratch when its moment came. Each was a recombination of older primitives applied to a new layer of the stack.
The agent layer is the new layer of the stack. The recombination is what an open protocol for agents looks like.
Why Rotifer chose this path
Rotifer is not building this protocol because we have a clever take on agent architecture. We have a clever take on agent architecture, but several other teams do too, and that is not the part of the work that matters. We are building it because the structural problem is now visible, the window during which it can be solved is short, and nobody whose business model could survive solving it is currently positioned to do so.
The shape of our bet is straightforward and worth being explicit about:
The protocol is open. The reference implementation at rotifer.ai exists to prove the protocol can be implemented, used, and grown. The reference implementation is not the protocol. It is replaceable. It has to be replaceable, or the protocol is just a closed product with extra documentation.
Other bindings are part of the design, not announcements. The protocol leaves clean slots for a Web3-native binding (where agent lineage lives on a chain), for market-localized bindings (handling regulatory and language realities specific to particular jurisdictions), and for on-device, mobile, embedded, and TEE-backed bindings. Each binding implements the same protocol against a different operational reality, and any of them can be built by anyone. If, three years from now, the most-used Rotifer binding is one we did not build — that is the goal, not a failure mode.
This is not how startups usually think. Startups try to be the binding nobody can replace. We are trying to be the protocol that makes the bindings replaceable, on the theory that the layer that survives commodification is the layer that does not need to win.
The deeper version of the bet is that, in the AI age, sovereignty over your own agents is going to matter to people the way sovereignty over your data started to matter in the late 2010s — quietly at first, then suddenly. When that flip happens, the only places to run will be places that were designed from the start to let you leave.
The two- to three-year window
Windows like this do not stay open.
The closed AI labs cannot pursue this strategy because their revenue model forbids it. Their margins depend on tokens flowing through their inference, on agents living in their app, on developers being unable to move. They will not pivot. They cannot pivot. Watching this is one of the more interesting strategic exercises available in technology right now — observing a class of companies be structurally unable to do the thing the market needs done.
There is exactly one organization with the cultural DNA to compete on open protocols and the creator economy at once: Hugging Face. They are open by default, friendly to creators, and have a model hub the rest of the industry actively uses. They have not yet built an agent protocol layer. If they do — and on a long enough timeline they probably will — Rotifer's job becomes harder. That is the contest we have to win.
We estimate the window at two to three years. That is the time available to define what an open agent protocol means in practice before either Hugging Face moves or before some chunk of the closed-AI labs collectively decide they would rather have a slice of an open pie than starve on a closed one. Either of those outcomes is fine for the world. Both of them require the protocol to be sufficiently real, by then, that it can absorb the entrants.
So we are moving now. The first reference binding is live. The genome-level composition primitives are documented. The local-first execution path is being implemented this year. The creator economy mechanics are scheduled for v1.0. The papers that explain why the underlying decisions look the way they do are public.
If you have read this far, you are probably the kind of person who notices when a layer of the stack is about to be redefined. This is one of those moments. The window is short.
An invitation
The protocol is public. So is the reference implementation. So are the papers.
If you build agents, the simplest thing you can do today is compose one against an open protocol and see whether it behaves the way the closed-product versions do not. The CLI takes a minute to install. The cloud surface at rotifer.ai takes about that long to sign into.
If you build infrastructure, the more interesting move is to build the next binding. The architecture deliberately separates protocol from implementation, and the binding slots are all open: Web3-native (on-chain agent lineage), market-localized (jurisdiction-specific regulatory and language realities), mobile, on-device, embedded, TEE-backed. Pick one.
If you write, build, research, or do anything that produces capability the world wants to compose — the agent economy is starting now, and the layer it runs on can either be one a few companies own, or one all of us do. The choice gets made in the next few years. By people who showed up.
We would rather it be open. That is why we are building it that way.
— Rotifer Protocol