#opinion

Two weeks ago I said the DGX Spark's bandwidth wall was physics, not a driver update — fast or big, pick one. Then I ran a 122B model on the same box at six times the speed of a 70B. The law was right. The number I kept feeding it wasn't.

I bought Andromeda, a DGX Spark, to fine-tune models — but first I owed my last post a rematch. I re-ran the same local/free/paid LLM inference benchmark on 120 GB of Grace-Blackwell instead of a CPU NUC. It fixes the one thing the NUC got fatally wrong, hits a bandwidth ceiling I didn't see coming, and makes clear why the real reason I bought it isn't inference at all.

I bought a NUC to run local LLMs and escape API costs. Then I tested local, free cloud, and paid cloud against the same prompts. The result wasn't 'pay for inference' or 'run it yourself' — it was that 'fast and free' isn't a quadrant that exists right now, and that has implications for anyone planning a personal AI setup.

I've spent months building Android apps — todo, drawing, live wallpaper, launcher, Kotlin Multiplatform, full-stack with Ktor — entirely with Claude Code, without writing a line myself. This is an honest account of what worked, what didn't, what it cost, and what I think it means for how software gets built next.