Customers can spend more with an AI provider while depending on it less.

A customer can use one provider more heavily for coding while moving other work to a rival. On Vercel’s AI Gateway, Anthropic’s share of estimated spending fell from about 64% to 44% between August and September 2026. That’s one gateway’s traffic, and it shows how quickly the mix can change.

We usually apply our AI defensibility diagnostic to software companies facing disruption. This time, we turned it on the frontier model providers themselves.

Switching costs can protect someone else

Replacing a model can mean testing outputs again, changing integrations, and rebuilding parts of an application. But whose position does that work protect?

It depends on what the customer would have to leave behind. That might be something specific to the model, the application that holds the customer’s context, or a cloud contract that shapes where the customer spends. A model can be important to a workflow while another company owns the relationship and can choose which model runs it.

What would make a lead last?

The cost of building a frontier model limits who can compete. Among those that can, a promising source of advantage is the ability to produce competitive models repeatedly, deliver them efficiently, and keep valuable customer work. The economic question is whether the returns justify doing that over and over.

For software owners, there’s a practical place to start. If a model your product relies on were retired tomorrow, how long would it take you to switch? And what would a competitor still have to build if it could buy access to the same model?

The report applies the eight-question diagnostic from our AI-Era Valuation Framework to the frontier model industry. It also sets out five questions software owners can put to their teams about model dependence and what their business owns.

If you’re interested, the full report is available below.