What OpenRouter's $7B Sale Teaches Builders About Margin at the Routing Layer

Key takeaways

• Stripe is reportedly paying around $7B for OpenRouter, about 50x its last known $140M annualized revenue, only 90 days after a $1.3B Series B.
• OpenRouter's reported annualized cost to serve its routing product was about $40M, or 28.5% of revenue, implying roughly $100M in annualized gross profit and a gross margin near 70%.
• The company was routing about 250 trillion tokens a month, up from 50 trillion in February, a 5x jump in six months, across roughly 8 million developers.
• That 70% margin is notable because OpenRouter doesn't run its own GPUs, it brokers requests to model providers and keeps a thin markup, so its main cost is infrastructure and engineering, not compute.
• For founders building any kind of wrapper, gateway, or reseller on top of LLM APIs, the lesson isn't the markup percentage, it's serving-cost discipline at volume.

Why did Stripe pay roughly 50x revenue for OpenRouter?

Reports put the deal at around $7 billion, closing just 90 days after OpenRouter raised a $1.3 billion Series B. Its last disclosed revenue figure was $140 million annualized, so the price represents roughly a 50x multiple, a standard-to-rich price for a top-tier AI infrastructure company growing this fast. One detail worth sitting with: reporting on the deal noted that OpenRouter, while much smaller in absolute revenue than some AI coding tools, likely has better underlying economics.

How does a routing layer get to 70% gross margin?

The math is straightforward once you see it. OpenRouter's reported annualized cost to serve its product was about $40 million, which works out to 28.5% of its $140 million in revenue. Subtract that out and you get roughly $100 million in annualized gross profit, a margin near 70%, in the range of high-performing publicly traded software companies. That's unusual for a business sitting this close to raw model inference, and it's explained by what OpenRouter doesn't do: it doesn't own or run GPUs. It routes requests to other providers and keeps a markup, so its cost base looks more like a typical software company's than a model host's.

What's the real lesson for founders pricing an AI product?

If you're not hosting the model yourself, your cost of goods sold is mostly the provider's API bill plus your own infrastructure, not GPU depreciation or data center costs. That changes the whole margin equation: your achievable gross margin becomes a function of how thin a markup the market will bear, and how much volume you can route through that markup. A 1 to 2% markup sounds negligible until you're moving 250 trillion tokens a month.

Should you build your pricing around markup percentage or absolute volume?

Model both, because neither number means much alone. A thin margin at low volume barely covers your infrastructure bill. The same thin margin at OpenRouter's volume produces $100 million in gross profit. For example, say your product routes $10 million a year in underlying token spend: a 5% markup nets $500,000 in gross revenue on that layer alone, before your own infrastructure costs. The volume assumption is doing almost all the work in that outcome, which is exactly why it deserves more attention than the markup number itself.

The story here isn't the exit price, it's that a wafer-thin per-token markup compounds fast once volume is real. Model that curve for your own product with a proper token cost simulator before you lock in a markup number.