Cursor and Databricks have launched their own model selection layers while Ramp is building comparable functionality, CEO Alex Atallah told 20VC, pulling the routing decision into the application stack.
OpenRouter routes roughly a million AI queries a day to whichever model is cheapest, fastest, or best at the task: a kind of stock exchange for large language models, where the bid and ask are latency and price per token. Its CEO, Alex Atallah, told 20VC this week that the very companies relying on that routing are now building their own.
Cursor, the AI coding editor, has launched its own routing layer. Ramp, the corporate-card company last valued at $44 billion, is building comparable functionality. Databricks has shipped similar features. The shift, Atallah said, is that the routing decision is moving into the application layer, the software a customer already pays for, rather than living as a standalone service.
The pressure shows up in OpenRouter's customer pipeline. The risk is not falling token prices; Atallah's bet is that elasticity stays above 1, so falling prices expand volume. The risk is that the routers themselves stop being a business. A marketplace needs a margin between what it pays providers and what it charges customers. When the customer can run the auction inside its own product, the marketplace loses that margin.
The same interview surfaced a separate, contested claim: that OpenRouter is in acquisition talks with Stripe at roughly $10 billion, against OpenRouter's last-disclosed $1.3 billion valuation in May. Neither company has confirmed the figure publicly.