The price of frontier AI capability is falling towards zero. The consequence for an IP strategy is not that AI has become cheaper. It is that anything you thought you were buying at the model layer, you are not buying exclusively.

In July a Chinese lab released a model, free to use, that appeared to rival the paid frontier models from the large US developers. When the weights were published a week later, they came under the developer’s own bespoke licence rather than a standard open-source one. American companies had one less reason to pay for a model, and Washington had a problem it has not resolved. Part of the response has been a pre-release security review for frontier models, which one former official described as a licensing regime for frontier AI in all but name, on top of earlier measures aimed at stopping Chinese developers training on the outputs of US models.

Strip out the politics and a control-point question is left, and it is the same one that arrives whenever a capability becomes broadly available. If everyone can run the model, what is left to own?

Four IP strategy questions for anyone buying AI

– Are you paying for exclusivity at the model layer? You are not getting it. Price and term your AI contracts on the assumption the underlying capability keeps getting cheaper. Short terms, benchmarked pricing, and a right to take your prompts, evaluations and fine-tuned weights with you.

– Have you read the licence, rather than the label? Open weights are not open source, as I set out in When Everyone Can Run the Model. Bespoke licences carry attribution triggers, revenue thresholds, field-of-use limits and acceptable-use terms, and all of them travel with the weights into your product. That is a freedom-to-operate question for your business, not a procurement footnote.

– What happens if the model becomes unavailable? Export controls, release reviews and pressure not to use particular vendors make the right to deploy a regulatory variable rather than a commercial one. Any architecture only one model can serve is a dependency, and it belongs in the risk register with a substitution plan beside it.

– What have you captured at the layers that do not commoditise? Your proprietary data, the workflow your customer actually works in, your evaluations and your domain-specific tuning. In most companies these exist as undocumented practice rather than as filings, controlled trade secrets or contract terms.

None of that is new, and none of it is specific to AI. Value moves to whatever stays scarce when the headline capability does not, which is the argument in AI Isn’t Your Advantage—Your IP Strategy Is. So the question for your next strategy review is not which model is best this quarter. It is what your business still owns on the morning the model you rely on becomes free to everyone, or unavailable to you.

Read the article: China’s AI models have Trump’s AI world at war with itself, James O’Donnell, MIT Technology Review.