Stripe buys the gateway that meters 10 trillion AI tokens a day

What Stripe is buying is a meter. A layer sitting between 10 million developers and more than 400 models does three jobs on every call (route it, price it, record where it went), and the third one never shows up on an invoice: at more than 10 trillion tokens a day, which averages about a million tokens per developer per day ^[11], OpenRouter can see which lab's new release is actually carrying production traffic well before the rest of the market can, and it says publishing exactly those signals is part of why it exists ^[6].

The volume makes the economics legible even without a disclosed price, which the post does not provide ^[10]. Ten trillion tokens a day annualises to roughly 3.65 quadrillion tokens ^[12], so on that base every cent of gross margin per million tokens is about $36.5m a year and every dime is about $365m ^[14] (illustrative, since OpenRouter publishes no blended take rate here). Run its own growth claim backwards instead: at least tenfold a year since early 2023 ^[4] puts the same gateway near 1 trillion tokens a day twelve months ago and near 100 billion two years ago ^[13]. That is a curve that makes an independence argument easy to write, and OpenRouter wrote it, saying its mission, its neutrality and its market lead made the case for staying alone a strong one ^[8].

This is probably the wrong thing to worry about, but the conflict people will reach for first is the one that is least available here: Stripe does not train frontier models, so the classic gateway failure (the router quietly favouring the parent's model) has no obvious beneficiary, and OpenRouter states the commitment does not bend to any parent company and that routing stays driven by the user ^[7]. The more interesting version is data adjacency. OpenRouter's stated reasons for choosing Stripe include a large customer network and data on how internet businesses grow ^[9]; joined to per-model, per-provider inference volume, that is a picture of which AI companies are compounding, held by the company that also moves their money.

Note what the sale takes off OpenRouter's own roadmap. It is no longer building the case that a neutral layer has to be structurally unowned to be trusted, and it has swapped that argument for reach, plus the inference-adjacent surfaces it says it is expanding into, including AI-native web search and context management ^[16]. Developers get continuity in exchange: same name, same product, same integration, per the announcement ^[5]. Inference, the company argues, is on its way to being the largest line item at every company ^[15], which is the reason the meter is worth owning.

Three observable things would show the neutrality claim held. The public model and provider rankings keep publishing at the same granularity a year from now. Concentration of routed volume in a handful of providers does not rise faster than the model count would explain. And using the gateway still does not require Stripe as the payment path. If all three hold, the sceptical read (mine) was wrong, and the useful part is that all three are checkable from outside, which is the only reason to write the prediction down.