AI has a Subsidization Problem

AI has a subsidization problem.

Or at the very least, the largest models have a large subsidization problem.

When you pay the 200$ a month for Claude Code, you get around 5k worth of inference. You don’t need to be a rocket scientist to know that this is not sustainable long term.

I see a lot of developers use agentic AI to handle even the most basic of tasks, like editing the width of an element. Then they go and complain that the agent is taking too long to return a result. They end up writing a very long blogpost about how to optimize their prompt to get the agent to return a little bit faster, 98% of which is unintelligible buzzwords that make neither the writer or the reader better at their craft.

Imagine how much faster it would be if they just opened the file and changed the css ruleset themselves, without having to endure the roundtrip-time to OpenAI or Anthropic HQ.

It’s not even all that complicated.

The companies are in customer-acquisition mode, so they are willing to pay a premium to retain users. The economics don’t make sense, so sooner or later revenue-extraction will begin, and the products will get worse and worse.

The alternative is to run your own models, but that is technically challenging, expensive, time-consuming, and beyond the perview of the average user.