1% of OpenAI and Anthropic's customers generate 80% of their enterprise revenue, and the compute promises stacked on top of that revenue start coming due in 2027.
The Cursor story is the clearest illustration of the fragility you describe. There is a competitive layer underneath the customer concentration that makes the 2027 math even harder. Anthropic's own threat report (published September 10) documents that 7 China-based labs ran distillation attacks totaling roughly 200 million exchanges — Alibaba's operation alone accounted for 151 million, used to train the Qwen 3.5 through 3.7 series. If the top-1% of spenders keeping that $1.3 trillion commitment chain solvent start routing even a share of their spend to models trained that way, at a fraction of frontier pricing, the concentration risk at the foundation becomes harder to price.
I use AI every day, so I don’t need convincing there’s real demand for it. Yet I'm not convinced that real demand automatically makes the AI money currently moving around sustainable.
Concentration shows up differently at the bottom of the stack, where the buyer has no leverage over the contract. EU law moves on that in January: Article 29 of the Data Act, Regulation (EU) 2023/2854, bars providers of data processing services from imposing any switching charges from 12 January 2027, after a transitional period of reduced charges. The rule covers cloud and edge services rather than AI models as such, so how far it reaches a model API bundled into a platform is unsettled. For a firm buying one vendor's assistant, the practical question is whether its data leaves in a usable format, not whether the vendor survives.
The Cursor story is the clearest illustration of the fragility you describe. There is a competitive layer underneath the customer concentration that makes the 2027 math even harder. Anthropic's own threat report (published September 10) documents that 7 China-based labs ran distillation attacks totaling roughly 200 million exchanges — Alibaba's operation alone accounted for 151 million, used to train the Qwen 3.5 through 3.7 series. If the top-1% of spenders keeping that $1.3 trillion commitment chain solvent start routing even a share of their spend to models trained that way, at a fraction of frontier pricing, the concentration risk at the foundation becomes harder to price.
I use AI every day, so I don’t need convincing there’s real demand for it. Yet I'm not convinced that real demand automatically makes the AI money currently moving around sustainable.
Concentration shows up differently at the bottom of the stack, where the buyer has no leverage over the contract. EU law moves on that in January: Article 29 of the Data Act, Regulation (EU) 2023/2854, bars providers of data processing services from imposing any switching charges from 12 January 2027, after a transitional period of reduced charges. The rule covers cloud and edge services rather than AI models as such, so how far it reaches a model API bundled into a platform is unsettled. For a firm buying one vendor's assistant, the practical question is whether its data leaves in a usable format, not whether the vendor survives.