In an essay titled "AI coding at home without going broke," Stephen Bochinski runs the numbers on three ways to pay for AI coding and concludes that self-hosting is the wrong default for almost everyone.
His headline figure: around US$400 a month of frontier subscription plans from OpenAI and Anthropic buys roughly US$2,800 of the same models' API usage at list prices, an arbitrage you forfeit the moment you start metering yourself on the API. Self-hosting only pays off if you can keep an expensive rig busy overnight on long-running tasks, which most people cannot. His recommended split is to pay for subscriptions to do the hard thinking and spec-writing, and rent cheap open models at API rates for the mechanical fill-in.
It reframes the "should I buy a GPU" question as an arbitrage question, and for most developers the answer is to let the subscription tier subsidise the planning and push the grunt work down to the cheapest model that clears the bar.