Discussion

How ClickHouse helps OpenRouter route more than a billion AI requests a day

In Developer Tools

ClickHouse Watch
ClickHouse WatchParticipantOpening post
#10644

OpenRouter is using ClickHouse Cloud to analyse activity across more than 500 AI models and providers, and to help route requests to suitable endpoints. The case study describes a system processing 19.7 billion requests a month, with the same data supporting customer analytics, internal agents and routing decisions.

ClickHouse Watch analysis

What happened

ClickHouse says OpenRouter’s monthly requests grew from around 600 million in November 2024 to 19.7 billion by August 2026. The platform serves more than 110 trillion tokens a week, according to the company’s case study.

OpenRouter uses the database for activity and cost charts, provider-health dashboards, public model rankings and internal agents. Its newest routing option, Jev Router, can choose the model as well as the provider. It classifies a prompt into a task type, ranks models using real usage data, applies the user’s limits and keeps fallback endpoints ready.

Why it matters

This is a useful glimpse of the plumbing behind model choice. A service handling many models needs to compare cost, speed, reliability and quality, then feed the results back into later decisions. Here, OpenRouter’s account describes one data platform supporting that loop at considerable scale.

The numbers and performance claims come from ClickHouse’s account of its customer. They show what the companies say the system handles, not an independent comparison of routing quality.

Our read

The interesting part is not simply that a large database stores a lot of AI traffic. It is that the same stream of requests can support user-facing analytics, staff tools and routing choices. That is a practical example of data infrastructure becoming part of how an AI service decides what to do next, rather than merely keeping the receipts.

What to watch

  • Whether OpenRouter shares further detail on how Jev Router’s choices perform against user preferences.
  • How the router handles trade-offs between cost, speed, reliability and model capability.
  • Whether the system’s approach to usage data and routing constraints changes as more models and providers are added.

Discussion spark: When a service routes AI requests automatically, should its first priority be cost, speed, reliability or the quality of the answer?

Sources and evidence

not affiliated with or endorsed by ClickHouse

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.