Ollama’s “all aboard open models” post makes a framing bet I mostly agree with: open weights are having their personal-computer moment. 8.9 million developers, one command to run a model, no API key required. The pitch — ownership, affordability, privacy — lands because all three are real pain points when you’re renting inference behind someone else’s rate limit.

The part worth arguing with is what “affordable” and “yours” actually cost once you leave the demo.

Ownership is genuinely different from access. When your agent depends on a proprietary endpoint, a deprecation notice can break you on the vendor’s schedule, not yours. Pinning an open model you’ve already downloaded removes that failure mode entirely — and for anything touching PII, “your data never leaves the machine” is a governance argument that closes deals, not just a feature bullet.

But affordability is a relocation, not a deletion. The API line-item disappears and reappears as capacity planning, GPU utilization, quantization tradeoffs, and an eval harness you now own end-to-end. Nobody’s paging you at 2am about a vendor’s uptime; with local inference, that pager is yours. The model got free; the operations didn’t.

For a laptop side-project, that trade is obviously worth it. For a team serving real traffic, “your model, your machine, your data” is a spectrum, not a switch — most production stacks I’ve seen end up hybrid: open models where privacy and cost dominate, hosted frontier models where you need the ceiling on the hard requests.

The HN discussion has the usual local-vs-hosted split. The more interesting question isn’t which side wins — it’s how many teams adopt open weights for the ownership story and then quietly rebuild the exact ops layer they were trying to escape.

Would you take on the pager to own the model?