Own your data. Rent the model.

Prices that don't reflect cost

Flat-rate AI subscriptions ($20 to $200 a month) sit on top of inference economics that remain genuinely contested. Critics argue frontier labs spend more delivering compute than they collect: reporting cited by Ed Zitron puts Anthropic's 2025 gross margins around 40%, roughly 10 points below projection, attributed to inference costs running 23% higher than anticipated, and OpenAI spending half its revenue on inference compute alone^[1][2]. Counter-analyses, notably Martin Alderson's, argue per-user inference is likely profitable and that retail API prices overstate true compute cost^[3][4].

The point of this paper does not depend on resolving that debate. What matters is the observable behavior: providers keep repricing, and always in the direction of less for the same money.

The observable corrections

5 events across 4 providers in 12 months, all in the same direction. This is what the end of a subsidy looks like: not a single dramatic price hike, but a steady conversion of flat-rate access into metered access.

Why this lands hard on small operators

The people most exposed are not enterprises with negotiated contracts. They are individuals and small businesses who built "vibe-coded" systems, intricate AI-assembled automations, document pipelines, and agent workflows, on the assumption that $20 to $200 a month buys effectively unlimited capability. For a small business, an AI plan repricing becomes 1 more input cost, and input costs only have 2 obvious exits: raise what the business charges, or cut something else.

There is a 3rd exit, and it is usually discovered under duress: abandon or rebuild the workflow. That is where the real asymmetry lives. Its cost is determined almost entirely by where your data and operating knowledge live. If they live inside the vendor's platform, in chat histories and proprietary project formats, the rebuild cost is enormous. If they live in organized local files, the rebuild is a re-pointing exercise.

Vibe coding builds real infrastructure

"Vibe coding" started as a pejorative for casually AI-generated software. In practice, by 2026 it describes a real and growing class of production infrastructure: small operators with no engineering staff running deal pipelines, document processors, scheduled agents, and client-facing automations that an AI built and an AI maintains. These systems work. That is exactly the problem. They work well enough to become load-bearing.

The double dependency

A vibe-coded system carries 2 distinct dependencies that are easy to conflate:

1. Capability dependency. You need a model to run and maintain the workflow. This is healthy and unavoidable; it is what you are paying for.

2. Substrate dependency. Your data, context, prompts, business logic, and institutional knowledge live inside a vendor's product. This is unhealthy and avoidable.

Capability is a commodity that is getting cheaper and more interchangeable over time; multiple frontier labs, open-weight models, and aggregators compete on it. Substrate is the opposite: the longer it accumulates inside 1 vendor, the more expensive it becomes to leave, precisely as that vendor converts flat-rate access into metered access.

Local-first, restated for the AI era

In 2019 Kleppmann, Wiggins, van Hardenberg, McGranaghan, and the Ink & Switch lab articulated local-first software: the primary copy of your data lives on your device; servers hold secondary copies; ownership means "user agency, autonomy, and control over data: you should be able to copy and modify data in any way, and no company should restrict what you are allowed to do"^[10].

That essay predates the AI subscription era, but it anticipated it. The AI-era corollary is:

Concretely, owning your data in the AI era means: every knowledge base, prompt, skill definition, and workflow specification exists as plain files (markdown, code, structured text) on storage you control; everything is versioned in repositories you can clone anywhere; and any AI tool you use reads from and writes to that substrate rather than accumulating it internally.

Case study: this lab's working drive

This lab runs the discipline on a single encrypted portable SSD. As of the June 2026 internal survey, the drive carries 11 git repositories, knowledge bases, agent workspaces, and a skill monorepo, all plain markdown and code, all pushed to remotes we control^[11]. The architecture has 3 rules worth generalizing:

1 canonical copy per knowledge base. Each domain knowledge base (a "nexus") exists exactly once at the drive root as the single source of truth. Agent workspaces that need it mount lightweight references, never second full clones. When the survey found 2 nested duplicate clones, removing them freed ~108 MB and, more importantly, eliminated 2 competing versions of the truth that AI agents could read^[11].

Workspaces are disposable; nexuses are not. Operator rooms (per-client agent workspaces) can be rebuilt from the canonical knowledge bases at any time. The deploy path is 1-directional: edit canonically, review, merge, then fast-forward pull into the live machine. If any AI vendor disappeared tomorrow, the workspaces are re-creatable; nothing irreplaceable lives in them.

Skills are files, not platform features. Every AI capability we rely on, document readers, financial classifiers, processing-desk checklists, is a plain-text skill definition in a versioned monorepo. They are portable across any agent runtime that can read instructions from disk. The vendor supplies the model; we supply the expertise.

What this buys you when the price changes

Return to the repricing table in §01. For an operator on the owned-substrate model, each of those events is a procurement decision, not a crisis: compare the new price against alternatives, point the same files at a different runtime if warranted, continue. For an operator whose context lives in vendor chat history and platform-native project formats, each event is a forced negotiation in which the vendor holds the asset.

For an individual or small business building on AI subscriptions today, in rough order of leverage:

This is a first draft and the argument has honest edges. The inference-economics debate is unresolved; if Alderson's analysis is right and flat-rate plans are sustainably profitable, the urgency (though not the prudence) of the hedge decreases^[3][4]. Local-first discipline has real overhead, governance, deduplication, deploy scripts, that a solo operator may reasonably defer. And some vendor lock-in is capability lock-in in disguise: if only 1 model can run your workflow well, file portability alone doesn't free you. These deserve fuller treatment in a later revision.

The subsidized AI subscription is the cheapest infrastructure most small operators will ever rent, and it is being repriced in real time, 1 rate-limit announcement at a time. The wrong responses are denial (assuming today's pricing persists) and abstinence (forgoing the capability). The right response is architectural: own your data, in an organized and localized way, so that the only thing you ever rent is the model. Buildings appreciate; tenants change.