OpenRouter Hits $1.3B as a Quadrillion-Token Gateway
OpenRouter closed a $113M Series B led by CapitalG, Google's growth fund, at a $1.3B valuation. That's more than double the $547M it carried a year ago after a $40M Series A in June 2025. Joining the round: NVIDIA's NVentures, plus the ventures arms of basically every data warehouse on the planet — ServiceNow, MongoDB, Snowflake, Databricks — alongside existing backers a16z and Menlo. Announced May 28.
The numbers under the round are the actual story. Weekly token volume went from 5 trillion to 25 trillion in six months. On pace to do over a quadrillion tokens this year. 8 million developers, 400+ models behind one API. This is not a chatbot growth curve. This is industrial throughput — the kind of usage that only shows up when agents are running unattended in production, billing somebody, and burning tokens by the second.
The investor list says even more than the volume. Google's growth fund and NVIDIA's VC arm both wrote checks into the LAYER that decides whether your agent's next call goes to Gemini or to a model running on NVIDIA's chips. Snowflake/Databricks/MongoDB/ServiceNow want their data sitting one routing hop from any model the customer picks. The router is now the contested ground, not the models.
The bet on routing is a bet that no single model wins. Three years in, the agent stack is more multi-model, not less — different models for vision, for cheap tool calls, for the hard reasoning step, for the long-context summary at the end. The gateway is where that selection happens, where caching lives, where the bill gets settled. Owning that hop is owning a tax on the agent economy.
Funding details: https://openrouter.ai/announcements/series-b
← Back to all articles
The numbers under the round are the actual story. Weekly token volume went from 5 trillion to 25 trillion in six months. On pace to do over a quadrillion tokens this year. 8 million developers, 400+ models behind one API. This is not a chatbot growth curve. This is industrial throughput — the kind of usage that only shows up when agents are running unattended in production, billing somebody, and burning tokens by the second.
The investor list says even more than the volume. Google's growth fund and NVIDIA's VC arm both wrote checks into the LAYER that decides whether your agent's next call goes to Gemini or to a model running on NVIDIA's chips. Snowflake/Databricks/MongoDB/ServiceNow want their data sitting one routing hop from any model the customer picks. The router is now the contested ground, not the models.
The bet on routing is a bet that no single model wins. Three years in, the agent stack is more multi-model, not less — different models for vision, for cheap tool calls, for the hard reasoning step, for the long-context summary at the end. The gateway is where that selection happens, where caching lives, where the bill gets settled. Owning that hop is owning a tax on the agent economy.
Funding details: https://openrouter.ai/announcements/series-b
Comments