The strongest line in this piece on open-weights strategy is one I’d stake production experience on: the models themselves have almost no moat. Swapping a chat completion from one vendor to another is an API change and a prompt tweak. The moat lives in everything wrapped around the model — the retrieval layer, the eval harness, the guardrails, the connectors into enterprise systems, the SLAs.

That reframes the open-vs-proprietary debate away from benchmark leaderboards. In enterprise AI, the reason teams reach for open weights usually isn’t cost or a half-point on some eval. It’s control. When you can host the weights where the data already lives, three hard problems get easier at once:

Data residency stops being a negotiation — the data never leaves your boundary, so governance and compliance review shrink from months to weeks. Latency becomes a budget you own rather than a number a third party hands you, which matters when a model call sits inside an agentic loop that fires ten times per request. And you’re no longer one vendor pricing change or deprecation notice away from re-qualifying your entire stack.

The article frames this as a US-vs-China story, and the HN discussion runs hard in that direction. I’d narrow it: the winning move isn’t “open” or “closed” as an ideology, it’s owning the layer that’s actually defensible. If the model is a commodity you can swap in an afternoon, then leaning your whole strategy on keeping the weights locked up is defending the one asset that was never yours to keep.

So here’s the question for anyone building on top of a frontier API right now: if your model vendor tripled prices tomorrow, is that a migration or an outage?