Those Who Don't Build Models Are Growing Another Layer on Top of Them

In May 2026, OpenRouter, which routes requests to hundreds of AI models through a single API, closed a $113 million Series B at a valuation of about $1.3 billion. The lead investor was CapitalG, an Alphabet fund, and Nvidia's NVentures was among the followers. Two months later the Wall Street Journal first reported that Stripe was in talks to buy it for about $10 billion, 7.7 times the last round. The deal isn't done, and everyone involved stresses that it may change. The same month, Tokyo's Sakana AI published two papers at ICLR 2026, TRINITY and Conductor, and packaged the engineering into an API product called Fugu. Its published results: 73.7 on SWE Bench Pro against 69.2 for Claude Opus 4.8 and 58.6 for GPT-5.5, and 50.0 on Humanity's Last Exam against 49.8 for Opus 4.8. Neither company trains its own frontier model. OpenRouter gathers more than 300 models behind one entry point, and Fugu makes several models cooperate on a single task. Both sit in the same position between models and users, and the market repriced both at the same moment. Why a zero-markup channel is worth $10 billion OpenRouter's inference is zero-markup, and so is its peers'. In a June 2026 side-by-side comparison, none of OpenRouter, Vercel AI Gateway, Cloudflare, Portkey, or LiteLLM adds to token prices. Vercel's site puts it as "Pay provider prices, never a cent more." OpenRouter's revenue comes from a 5.5% card fee on top-ups and a management fee on BYOK (bring your own API key), and neither is on inference. Tomasz Tunguz's verdict on this structure: reselling inference at cost is a zero-gross-margin business, a payment rail and no software company. The financial basis agrees. OpenRouter's public figure in May 2025 was $100 million in GMV against about $5 million in annualized net revenue, a 20-fold gap. The accounting basis is ASC 606's distinction between principal and agent. A middle layer reselling someone else's compute is an agent and can book only the net amount. Three orders of magnitude separate $5 million in net revenue from a $10 billion asking price, and that distance is what the deal has to explain. One difference that can be checked is data. OpenRouter handles requests of trillions of tokens a day, and every one of them, which model it goes to, which model is used in which scenario, which models are being retired, and which scenarios fail most, passes through it. Anthropic sees Claude's requests and OpenAI sees GPT's, each with only its own segment. The cross-model distribution sits with the middle layer for now.