In mid-August, Stripe announced it would acquire OpenRouter for roughly $7.5 billion, according to The New York Times (read the earlier $10B rumor-stage take). Three months earlier OpenRouter had closed a round at a $1.3 billion valuation — a five-fold jump in ninety days. Founded in 2023, the company has just 82 full-time employees, which prices the deal at over $90 million per head.
Here is the part that matters: OpenRouter has never trained a single model. Its only job is deciding which model each request should go to. The platform aggregates 80-plus providers and 400-plus models, serves roughly 10 million developers, and routes over 10 trillion tokens a day.
The buyer is not an AI lab. Stripe is a payments company — a toll collector sitting between buyers and sellers. This is the clearest signal yet that the center of gravity in AI is migrating from the model itself to the entry point that controls default traffic.
The model is commoditizing; the router is not
Think about how ride-hailing worked before aggregators. You installed several apps, compared prices, and re-booked every time. Then aggregator platforms appeared: they compare, dispatch, and settle — without owning a single car. OpenRouter plays that role for AI. It owns no models and no GPUs; it holds the middle layer. Developers get one API, and behind it sit hundreds of models from dozens of vendors.
Before such aggregation, switching models meant rewriting code, re-running tests, and reconnecting billing. With OpenRouter, switching is a config change: pick by price, speed, task complexity, or reliability. If one vendor goes down, traffic fails over instantly. The model is a commodity; the choice layer is where the durable value sits.
The revenue model is telling. OpenRouter does not mark up model calls — a GPT call costs exactly what it costs to buy directly from OpenAI. Its margin lives in the funding/recharge layer: a small fee on top-ups, much like a gift-card or FX spread. So it does not need to bet on a winner. The more vendors undercut each other on price and performance — a price war that signals commoditization — the more tokens flow through the platform, and the bigger its fee base grows.
Sitting in the middle, watching model supply shift
Being the route means seeing the whole map. The share of US-model tokens on OpenRouter slid from roughly 70% a year ago to about 30%, while Chinese open-source models rose to roughly 45%. Anthropic is the extreme case: just 12% of traffic but 46% of revenue — frontier capability commands a premium even when volume is thin. The same platform carries low-margin commodity compute and high-margin frontier models, and whoever controls the router controls where the default requests flow.
What is hard to copy is not the tech — it is the momentum
"Help the user pick a model" is not technically hard; anyone could build it. The moat is scale and timing. OpenRouter entered in early 2023, when almost everyone assumed the model itself was the value and nobody was watching the middle layer. By the time model supply diversified and price comparison became a need, it had already locked in the market. More models make developers dependent on it; more developers make vendors eager to join — a two-sided network effect.
The hard-to-replicate assets are the accumulated ones: model coverage, real traffic data, supplier relationships, and a hard-won reputation for neutrality. The first three can be bought with time and money. The last one cannot.
Neutrality is not a posture here; it is a license to operate. Because OpenRouter trains no models, vendors trust it with their APIs. The moment it builds its own model, it becomes a competitor, vendors withdraw, and developers stop handing over default traffic. Neutrality is a credit instrument you cannot redeem twice — once spent, it is gone.
The tension between neutrality and scale — why the founder sold
This is the uncomfortable part. Raising margins usually means compromising neutrality, and once neutrality breaks, trust collapses and scale loses its foundation. You cannot easily have both, and OpenRouter was hitting that wall even at its peak.
Traffic leadership did not translate into profit leadership. Its ARR is around $100 million-plus against a ~$7.5 billion price tag — a rich multiple that reflects the entry point, not the revenue. Its 5% take rate was being undercut from all sides: Vercel advertises zero markup, Cloudflare passes through upstream prices, LiteLLM charges nothing. Dozens of startups now do some form of routing, and big players — Databricks, Cursor, Meta, Ramp — have started routing in-house, turning what was an indie opportunity into a platform default.
The founder, Alex Atallah, is not new to selling at the top. In 2022 he helped found OpenSea, which once held ~97% of the NFT market before he stepped away as CTO months before the bubble burst. Selling OpenRouter at its highest moment looks like a pattern, not luck.
What this deal really says
Both sides can be right. Stripe believes it bought a future tollbooth: it already processes OpenRouter's payments, so it sees the daily flows, structure, and growth better than any outsider. Pairing a payment rail with a token router lets it touch both ends of a client's cash flow — the money that comes in and the AI spend that goes out.
The deeper takeaway is structural. When the most valuable thing about an AI company is its access point — not its model, not its compute — value has moved from building intelligence to controlling the door that routes to it (developer gateways are already a valued battleground). That is the real signal of this acquisition, and it is worth watching which entry points get acquired next.