If a capability you spent years building can be downloaded on a Sunday afternoon, the question is not how to stop that. It is what you own that survives it. Ben Thompson’s recent Stratechery piece works through the economics of the new Chinese open-weight models and lands somewhere useful for anyone thinking about IP. Open weights are not free. They shift spending out of R&D, which is fixed, and into cost of goods sold, which is not. Serving a model costs real money every time, so the winner in a commoditised market is not the one who charges more. It is the one with the better cost structure. Thompson also makes the point that tokens are not fungible but intelligence is, which is a precise way of saying the capability commoditises while the efficiency of delivering it does not. That is the same problem in a new setting: when anyone can build it in an afternoon, what is left to own is a question about defensibility, not about speed. The model itself was never the asset. As I have said before about the AI stack, the returns accrue to whoever owns the foundations, and the foundations here are everything that makes your version cheaper, stickier and better informed than the one anyone can download.

That reframing changes what is worth protecting and how. Note that the mechanisms Thompson identifies as durable – inference and memory efficiency, batching and caching, token efficiency, the harness the customer actually works in, and the data flywheel that comes from running inference at scale – are precisely the things that patents, trade secrets and well-drafted contracts can hold, in a way that published weights never could. Note too that the fight over distillation is a licensing fight, not a patent fight. It is being run through terms of service, which means it depends on contract drafting, on jurisdiction and on whether you can realistically enforce against the counterparty. This is the decision I have described elsewhere as choosing, in advance, which right to assert, where, and to what end.

Four questions follow, and none of them are about AI specifically:

– Ask what remains proprietary if a competitor obtains your core capability tomorrow. If the answer is nothing, the moat was rented.

– Check whether your operational efficiencies – the serving, caching and pipeline work nobody outside the team can see – have been captured as filings or as controlled trade secrets, or whether they are simply undocumented practice.

– Read the licence, not the label. Open-weight is not open source. Field-of-use limits, acceptable-use terms and downstream obligations travel with the weights into your product, and they are a freedom-to-operate question for your business, not a procurement footnote.

– If a supply contract is doing your protective work, confirm it is enforceable where the counterparty actually sits. A term of service that cannot be enforced is a preference, not a right.

The wider pattern is one that shows up well outside AI. Whenever a capability becomes broadly available, value moves to cost position, to proprietary data and to whoever owns the customer relationship. Those are IP decisions, taken early, or they are not taken at all.

Read the article: Who’s Afraid of Chinese Models?, Ben Thompson, Stratechery, 20 July 2026.