I compared a $500 GPU to a $2/month Claude API — here's what actually makes sense for most developers

The debate that's been heating up HN this week You've seen the posts: someone runs a benchmark, a local GPU beats Claude Sonnet on some coding task, and the comments explode. "Just buy a GPU. You'll save money in the long run." I wanted to actually do the math. Here's what I found. The upfront cost nobody talks about A decent GPU for local inference: RTX 4090: ~$1,600 RTX 3090 (used): ~$500-700 MacBook Pro M3 Max (if you're using it for AI): ~$3,500 At $2/month for API access, you'd need 250-1,750 months (20-145 years) to break even on hardware costs alone. Yes, that's ignoring electricity.