The AI tax: you're renting the model you already pay for

A quarter of Y Combinator's Winter 2025 batch shipped codebases that were roughly 95% AI-generated. Collins Dictionary made vibe coding its 2025 Word of the Year. The model is good enough to write the software now. So why do so many AI products charge you every single time you use them?

Because the charge is the product. Most of what gets sold as "an AI tool" is a thin wrapper around a model you could call directly, with a meter bolted to the front. You pay the vendor, the vendor pays the model provider, and the spread is the business. Run the same task tomorrow and you pay again. That is the AI tax: a per-token toll on access to a capability that is not scarce and not theirs.

What the AI tax actually is

Strip the UI off a typical AI SaaS and you find an API key, a prompt template, and a billing integration. The model is OpenAI's or Anthropic's or Google's. The prompt is a few hundred lines someone tuned over a weekend. The defensible part, in the vendor's eyes, is the meter.

You are renting access to a model you could pay for at cost. The wrapper adds a markup and a dependency. Every run goes through their servers, on their key, at their price, forever. Stop paying and the thing stops working, because there was never an artifact. There was only a subscription to someone else's API call.

Why the meter persists

The meter survives because it is the cleanest revenue story a founder can tell an investor. Usage-based pricing scales with adoption, looks like growth, and never asks the vendor to deliver something you keep. A tool that runs without them is a tool they cannot bill twice.

So the incentive runs against you. The vendor profits when you run more, not when you need them less. A tool that genuinely solved your problem would reduce its own usage, which is the last thing a per-token business wants. The pricing model quietly shapes the product: keep the AI in the loop, keep it necessary, keep the meter spinning. You are paying for the vendor's recurring revenue, and you are paying for it in compute you could have bought wholesale.

The costs you don't see on the invoice

The per-run price is the visible cost. The expensive ones hide underneath.

Variance. A model in the live path means the same input can produce different output on Tuesday than it did on Monday. Tests that pass once and fail later, summaries that drift, classifications that wobble. You are buying nondeterminism and calling it intelligence.

Vendor dependency. Your workflow now depends on their uptime, their rate limits, their model deprecations, and their next price change. When the provider sunsets a model or triples a price, your tool inherits it. You did not sign up to track someone else's roadmap.

Your data. Every run ships your inputs through their infrastructure. For code, customer records, or internal docs, that is a standing exposure with no expiry.

Your bill, indexed to your success. The more useful the tool, the more you use it, the more you pay. Cost scales with the exact thing you wanted to grow. You succeed and the tax goes up.

The alternative: author once, own the artifact

There is a different shape, and the AI does not have to leave the picture. It does the expensive part once. Let the model explore, reason, and figure out the hard thing. Then capture the result as a deterministic artifact: code, a config, a test, a script. Something that runs on its own, the same way every time, with no model in the loop.

This is the line we keep drawing at Hyperyond: AI authors it, you own it. Hover is the worked example. An agent drives your real browser to explore a flow, then crystallizes the verified session into a plain Playwright spec. The spec is standard @playwright/test code. It runs in your CI with no agent, no API call, no token meter. The AI did the authoring. The artifact does the running, forever, for free.

When you do need a model in the path, run it on your own keys, against your own endpoint, including local models. You pay the provider at cost. No spread, no middleman, no dependency on a startup outliving its runway.

Which tools to trust

Ask one question of any AI tool: what do I own when I stop paying?

If the answer is nothing, it is a meter. If the answer is a file, a model, a deterministic thing that runs on your hardware and your keys, the vendor has aligned with you. They got paid for the work and then got out of your way. That only happens when the business is built to hand you something durable instead of billing your every move.

Open and local-first is not a feature list. It is what alignment looks like in code. The artifact is yours, the keys are yours, the run costs you nothing extra, and the vendor cannot reach into your usage to charge you again. That is the only version of AI tooling worth depending on for the next decade.

The model is a commodity now. The next thing worth building is software that uses it once and then lets go.