OpenRouter Hits $1.3B as Model Routing Becomes Its Own Market

OpenRouter, the aggregator that lets developers swap between dozens of large language models through a single API, has closed a $113 million Series B at a $1.3 billion post-money valuation. CapitalG, Alphabet's growth fund, led the round, with Nvidia's NVentures, a16z and Menlo Ventures alongside the corporate arms of ServiceNow, MongoDB, Snowflake and Databricks all participating (TechCrunch, May 26).

Weekly token volume through OpenRouter's API now sits at about 25 trillion, up roughly fivefold in six months according to the company. The round more than doubles its valuation from last year's Series A and places it firmly in the picks-and-shovels camp: a company that does not train a frontier model of its own, but that takes a small cut every time a customer calls one through its router (BusinessWire has the full investor list).

What enterprises are actually buying

The bull case for OpenRouter is straightforward. Enterprises that built AI features against OpenAI or Anthropic in 2024 spent 2025 discovering they wanted to swap models without rewriting their code, either to chase a cheaper price-per-token or to hedge against a single vendor going down. A router solves that. It also gives operations teams a single bill and a single audit log across providers, which matters more than it sounds to anyone who has tried to reconcile usage across five APIs.

That convenience comes at a price. OpenRouter takes a margin on every call, and some voices in the open-source community argue that the company is reinventing a layer that should be commoditised. The counter-argument is the boring one: it ships, the latency overhead is low, and 25 trillion tokens a week is hard to wave away.

What the deal signals

CapitalG and Menlo have both spent the past year warning portfolio companies about vendor lock-in. Funding the router that breaks the lock-in is consistent with that view. As Menlo's own writeup puts it, weekly traffic has crossed a quadrillion tokens a year, a useful reminder of how quickly the inference market has scaled past the headlines about training-cluster economics.

The wider thesis behind the round is that enterprises are quietly giving up on the "pick one frontier lab and build everything on it" plan. The infrastructure that lets them keep their options open is now a billion-dollar category on its own, sitting between the model providers and the application layer. For other startups eyeing that middle, the obvious question is how much of the margin actually stays at the router once the underlying labs decide they want it back.