So, what did I tell you all? That when the rug is pulled, you're all going to have your backsides hanging in the wind?

Recall that Cursor burns $5000 for your $200 plan, and weep for the wild times. I've heard similar numbers from industry. Now you're hooked. You're dependent on that API, along with a million others who didn't want to miss out. Now they're coming for the other $4800, and that is just to break even. Cursor didn't even train their own model. They finetuned existing models. Cursor ran out of runway, got acquihired by x.ai, and now they're going to pull the rug.

It costs over a billion dollars to train a frontier model like the type OpenAI or Anthropic have. Each frontier model has a life of six months. In that six months, that model needs to rake in just under $170m, for training, not including inference. It needs 833,334 people to pay for a $200 plan, again before inference costs, R&D, the mountain of debt these companies have for upfront costs, and assuming nobody uses the free product.

Anthropic would need to charge $2000 per user per month to break even, assuming they even want to do flat pricing. They won't want to. Someone who uses Claude to create React components isn't going to use as many tokens as someone using Claude to refactor 20k+ lines of legacy Visual Basic code. There are a load more React Claudes than Visual Basic Claudes. Claude Code is expensive because the context window is absolutely massive for programming. Anthropic doesn't want React people. It wants legacy Visual Basic people. Legacy Visual Basic people work for insurance companies who get billed $5000/day by legacy contractors. Insurance companies pay the bills more reliably than the few million slop developers out there. Expect the rugpull.

There's a market for maybe five computers.

-- IBM, 1943

Now the next question: does the median developer need Claude? I don't think so. My experience is that open-source models, while not as good as Claude, are still pretty good given decent prompts. But what you give up in model quality you make up in other ways.

• No token worry - you own the hardware.

• Experimentability - you can finetune if you want.

• Privacy - you can tell it everything and unless something's gone wrong it won't sell your personal thoughts to data brokers.

You can start by buying a used ThinkPad P1 or T15 or something like that with a beefy graphics card. Install Ubuntu on it, keep it plugged into your router, put it behind Tailscale, then install Ollama and Open WebUI, and voila: local AI compute. Then real experimentation can occur.