Two months ago, I wrote that AI tokens were beginning to look less like API credits and more like a new kind of currency. I didn't expect the analogy to become this literal, this quickly.
OpenAI was already issuing tokens at near-zero marginal cost in exchange for equity in the next generation of AI-native startups—computational seigniorage: issuing cognitive currency, acquiring productive assets.
Now Stripe is reportedly in talks to acquire OpenRouter for around $10 billion—up from a $1.3 billion valuation two months ago. Not because of OpenRouter itself, but because of what layer Stripe would be buying: the infrastructure governing how cognitive currency circulates. Who routes inference. Who measures usage. Who controls billing. Who decides which model receives demand.
Frontier labs increasingly look like central banks, issuing the currency. Hyperscalers resemble commercial banks, distributing the liquidity. Gateways like OpenRouter become the payment network. Which is presumably why a payments company is the one buying it.
The real competitive frontier is quietly moving away from individual frontier models. The scarce layers are becoming compute, routing, identity, governance, billing—and whoever controls them extracts value regardless of which model happens to lead the benchmarks that week.
This starts to look less like a currency exchange and more like grid dispatch. DeepSeek just introduced peak-hour API pricing—double rate at 9am–12pm and 2pm–6pm. Microsoft is quietly swapping OpenAI and Anthropic models for its own MAI models inside Excel and Outlook to cut costs. You don't choose the source. Something upstream arbitrages between cheap and expensive capacity in real time, on your behalf. OpenRouter, Ramp, Cursor, Databricks: a spot market for cognitive capital is already being built, one routing product at a time.
If this goes through, Stripe won't have bought just a model router. It'll have acquired one piece of the layer where the economy of AI is more and more settling.
