Will foundation-model companies eat my AI app?

What gets eaten is the patch: a feature that exists because today's model cannot do something goes to zero with the next release. Model companies are indeed moving up into applications. But the model layer cannot hold value itself: large models have no network effects and no switching costs. The top US and Chinese models sit about 2.7% apart (LM Arena, March 2026), and the two leading labs swapped ARR rankings within a year. A moat can only sit one layer up, and the test is whether the cost a customer pays again on leaving belongs to the app company. The four moats are taste, extreme efficiency in one workflow, combining several models, and process as product.

Frontier labs also shut products that cannot hold revenue: Sora's consumer app reportedly closed on April 26, 2026, at a daily cost of about $1 million against about $2.1 million in cumulative in-app purchases.

The more practical risk is an upstream cut-off. On August 28, 2026 OpenAI said it would stop supplying models to Cursor from November 12, because after SpaceX's acquisition it could not confirm Cursor would use the technology under its terms of service. Three outcomes lead to three conclusions: if users follow the model, workflow position is no moat; if users stay and accept a substitute model, the position holds and the model is replaceable; if Cursor switches to in-house models, it completes vertical integration. Retention after November 12 is the verdict. For founders, any product that relies heavily on one model supplier should stress-test a cut-off alongside price swings.

What a workflow position is worth: Cursor's annualized revenue passed $4 billion in early June 2026, about 32 months after $1 million in October 2023. SpaceX's $60 billion acquisition of its parent Anysphere closed on August 14. The independent route is also scaling: per company disclosure, Cognition said on September 25 that its annualized revenue run rate had passed $1 billion.