Why OpenRouter Sold at Its Peak

10 trillion tokens a day, 10x year-over-year growth. So why sell now?

The Stripe-OpenRouter deal closed this week. Neither party disclosed the price, but sources put the acquisition premium at close to 8x the previous valuation round, making it the most closely watched AI infrastructure deal of the year.

In Issue #13, Cobo broke down the buyer's logic. For Stripe, acquiring an AI aggregation platform that takes a 5% cut on API calls was about locking in something scarce in the multi-model era: control over which models actually get called. Merge model routing with payment settlement, and a technical API call becomes a metered financial transaction with clearing built in.

But the seller's logic is equally worth examining.

OpenRouter was processing over 10 trillion tokens a day, with inference volume growing more than 10x year-over-year. Alex Atallah, who previously built OpenSea into a dominant marketplace, chose to sell at exactly the moment the numbers looked best. That timing is a tell. Behind the headline metrics, standalone routing gateways face structural pressure from both directions. Cashing out while every indicator is peaking, before the ceiling becomes legible to outside observers, is a highly rational move.

Cutoff from Above: When Routing Logic Moves Upstream

The more immediate threat to OpenRouter came from its own best customers.

In July 2026, Cursor published a technical post introducing Cursor Router, a purpose-built routing classifier trained on over 600,000 real user code requests and validated across millions of production calls. Cursor sits at the entry point of the developer workflow, which gives it something a backend gateway never has: rich, real-time context about what's actually happening. Task complexity, context length, latency sensitivity, subscription tier. By the time a request leaves the editor, Cursor's system has already decided which model handles it. Simple completions go to lightweight local models; complex debugging gets routed to frontier models. That internal dispatch alone cuts inference costs by 30 to 50 percent.

That's what hollows out horizontal gateways. OpenRouter's value proposition depended on being the intelligent layer that evaluated tasks globally and optimized model dispatch. Cursor already does that upstream, with better information. What reaches OpenRouter is whatever Cursor decided not to handle internally, which tends toward lower-value traffic almost by definition.

Meta-orchestration frameworks are accelerating the same shift at the architecture level. Databricks' recently open-sourced Omnigent lets developers hardcode routing logic into their own infrastructure before requests ever reach an external gateway. Step A goes to a cheap open-source model, step B goes to a frontier model, all decided server-side. By the time a request hits OpenRouter, there's nothing left to route intelligently. It's just passing traffic through a pipe.

That's a meaningful valuation problem. A routing layer that actually makes decisions commands a premium. One that executes predetermined instructions is invisible infrastructure, closer to a telecom carrier being bypassed by the application layer than to an intelligent dispatch system. The premium compresses accordingly.

Squeezed from Below: Pricing Power, the Go-Direct Problem, and Revenue Deflation

The foundation model supply side runs on a brutal power law. A handful of closed-source frontier providers hold near-absolute pricing power, backed by capability gaps that open-source alternatives haven't closed. On OpenRouter's platform, roughly one-eighth of models by traffic count captured close to half of total revenue. That concentration compounds in a specific way.

High-volume enterprise customers eventually do the math. Once API call volumes cross a threshold, a 5% routing fee gets hard to justify, and signing directly with the underlying provider becomes the obvious move. The customers most likely to go direct are exactly the ones generating the majority of platform revenue, so each defection hits disproportionately hard.

Token price deflation makes it worse. OpenRouter takes roughly 5% of inference spend flowing through the platform, but per-token pricing has been falling at close to 10x annually. Even as token throughput scales dramatically, dollar-denominated revenue can stagnate or shrink. Volume up, revenue flat.

After high-value traffic migrates to direct contracts, what remains is mostly long-tail open-source traffic, high concurrency at very low per-token prices. A 5% margin on cheap tokens is nearly nothing. Call volume keeps climbing on the dashboard while unit economics quietly deteriorate underneath.

From API Toll Booth to Clearing House for the Machine Economy

OpenRouter was always caught between two squeeze points. Model providers could cut it out with better direct pricing; downstream platforms could build their own routing layers once they scaled. That's a tough position for a standalone infrastructure company trying to justify a high valuation long-term.

Folding into Stripe reorients the growth curve. Thin-margin routing, anchored into Stripe's global payments infrastructure, compliance stack, and clearing networks, starts looking a lot more durable than it did as a standalone business.

Here's what that actually enables. When an agent picks a model and triggers an inference call, Stripe handles wallet authorization and cross-entity accounting within milliseconds. What used to be a network request is now a timestamped, traceable, settled financial transaction. The technical action and the commercial record happen at the same moment.

The direction of travel is clear regardless of what drove the specific deal terms. Routing and settlement are converging into a single stack, and agent-to-agent commerce can't scale until that infrastructure matures. This is the layer it gets built on.