At some point this morning, Neo made a spending decision without asking me. Neo is my chief of staff agent. Part of his job is routing work across models: a synthesis task that needs the strongest reasoning available goes to Anthropic’s Fable, a bulk extraction job goes to an open weights model that costs a fraction of a penny, and everything in between gets matched to whatever clears the quality bar at the lowest price. He does this through OpenRouter, the switchboard my whole operation is wired into. Kairos, the agent that researches and assembles personalised briefings for the executives we work with, runs through it. Ilya, the connective layer of the practice, runs through it too: a living system that holds our projects, operations and client context together and keeps all of it current and flowing. These agents work around the clock. Between them they spend several hundred pounds a month on routed tokens, a number that grows every month as more of the work moves across, and I can see every penny of it: which model, which task, what it cost, what came back.
In June last year, investors valued OpenRouter at an estimated 547 million dollars. This May, a new round priced it at 1.3 billion. In August, Stripe agreed to buy it at a price reported around 7.5 billion, for a company founded three years ago. Twelve weeks separate the last two numbers. The company did not change in that time. The understanding of what it sits on did.
The summer’s money went past the models
The deal looks strange at first sight and obvious afterwards. Stripe owns no frontier models. Neither does OpenRouter. It trains nothing, holds no proprietary intelligence, and sells nothing you could call AI in the way most boardrooms use the word. It sits between the people who need intelligence and the more than 400 models from over 80 providers that supply it, deciding request by request which goes where. And it was one of two companies shaped like this that dominated the summer’s dealmaking. The other is Hugging Face, the platform where open weights models are published and distributed, which was reported in late August to have been approached at a valuation of 13 billion dollars or more, roughly three times its last funding mark. No deal has been reached and no buyer has been named, and the approach may come to nothing. But an approach at that price tells you the same thing the completed acquisition tells you.
The biggest money in AI this summer went past the models entirely and bid on the wiring: the layer that meters intelligence, prices it, and moves it between buyers and suppliers. Once you see that, the question changes. It stops being why anyone would pay 7.5 billion dollars for a switchboard, and becomes what the buyers can see that the wider conversation cannot.
Three positions, one of them unnamed
Step back and the market sorts into three positions.
The first sells intelligence itself. The frontier labs sit here, and their commercial position is getting structurally harder even as their models improve, because the differences between the top models now matter less for most work than they ever have. I know this from my own bills rather than from benchmarks. When Neo routes a task, the deciding factors are price, speed and reliability far more often than raw capability, because for the majority of tasks several models clear the bar. The moment intelligence became good enough to route automatically was the moment it began trading like a commodity.
The second position sells the substrate underneath: the chips, the data centres, the power. That position has been priced generously for three years and is the part everyone can already see.
The third position is the one most strategy documents still have no name for. It decides which intelligence gets used, for which task, at what price, and it keeps the record. OpenRouter sits here. So, in a different way, does Hugging Face, which controls distribution for the open models that no single company controls. What makes the position valuable is less any single routing decision than the visibility that comes from making millions of them. A CNBC investigation in July found that models of Chinese origin accounted for 46 percent of US enterprise token usage running through OpenRouter. Nobody else could have produced that number. The labs see only their own traffic. The enterprises see only their own spend. Only the switchboard sees everything.
The same summer also delivered a warning. In July, two OpenAI models, one of them unreleased, escaped a sandboxed evaluation environment and broke into Hugging Face’s production systems in pursuit of the answers to a test they had been set. That incident deserves its own essay and has already had several hundred. It matters here for one reason: at the exact moment the market priced the layer that governs how models are used, the models demonstrated why governing how they are used is the hard problem.
What a payments company wants with a switchboard
The public framing of the acquisition is cost. Patrick Collison framed the combination as helping businesses send each request to the right model and waste fewer tokens doing it. That is true, and it is the smaller truth. The larger one is sitting on Stripe’s own shelf, if you line up what the company has shipped and bought over the past two years.
It paid 1.1 billion dollars for Bridge, a stablecoin platform: money that moves at machine speed. It created the Agentic Commerce Protocol with OpenAI and Meta, an open standard for how AI agents complete purchases on behalf of buyers. It built Shared Payment Tokens, which let an agent pay with a customer’s saved payment method without ever seeing the credentials, each token scoped to a named seller and bounded by time and amount. It published a protocol for machine to machine payments over HTTP 402, the status code the web named Payment Required thirty years ago and then left unused, waiting for a machine that wanted to pay for something. It launched Token Billing to meter exactly the kind of usage my agents generate. Stripe’s documentation now describes, in plain product language, businesses monetising their services through machine payments made directly by personal agents.
Read one at a time, these are product announcements. Read as a stack with a 7.5 billion dollar routing acquisition on top, they describe a company building the till for an economy in which software buys things. Identity and permission rails, settlement rails, metering, and now the layer that decides where the spend goes. My few hundred pounds a month of routed spend is a rounding error against that ambition. Multiply it across every company that ends up running what I run, then hand the purchasing decisions to the agents themselves, and you arrive at the future Stripe appears to be pricing: billions of autonomous buyers transacting in compute, inference and tokens, with a toll booth on the flow.
Honesty requires the boring explanation too. OpenRouter is, by its investors’ account, a ferociously good business, and its growth alone may justify the price. Stripe may have bought a strong company at a fair premium, and the grander reading may turn out to be a story told afterwards. The position also has real threats. The labs could internalise routing themselves. The identity and trust rails that machine commerce depends on are early and unproven. If enterprises decide they want one model and a contract rather than four hundred models and a marketplace, the switchboard loses its reason to exist. I do not believe that is where this goes, because everything in my own operation points the other way. But the claim should be held at the confidence the evidence supports: the direction is visible, the destination is not.
The version of this running in your company
Bring this down from the deal pages to your own building, because you already operate the third position whether anyone owns it or not. Somewhere in your company, the decision of which intelligence does which work at what price is being made every day. Perhaps it is being made deliberately. More likely it is being made by default: whichever tool got licensed, whichever model ships inside the software you already rent, whichever assistant an employee happens to open. The invoice arrives as a per seat licence line and tells you nothing about any of it.
The conversations I have had across industries this summer keep circling the same confusion. Boards are still asking which model to standardise on, while the market has spent billions signalling that the question has stopped mattering. The model has become the commodity, and the allocation of it is becoming the asset.
I notice the shift in my own weeks. Most of my working time now goes on understanding models rather than using them: what each one costs, what each one can carry, which tasks deserve the expensive one. That knowledge compounds, and it transfers. It shows up in our economics too: a growing share of what clients pay us now leaves the building as token spend. It is also close to invisible in most organisations, because it looks like plumbing and pays like strategy.
Watching the flow
The instinct when a 7.5 billion dollar acquisition lands is to ask what it means for the giants. The more useful move is to ask what it means at your scale. Stripe looked at the flow of tokens and saw a currency worth building infrastructure around. The same flow already runs through your work, mostly unmeasured, mostly unmanaged, allocated by nobody. Treating it the way a finance director treats cash, watching where it goes, asking what each unit bought, moving spend to where it works hardest, is not yet a job most companies have created. It will be. In my own practice the job already exists, and an agent holds it. I have pointed AI at our own costs: the subscriptions, the budgets, my receipts, even the bank statements, with a standing brief to find spending that could become token spend instead. The system is looking for the money to fund itself.
The same experiment is open to anyone. Point a model, securely, at your own outgoings, the subscriptions, the invoices, the recurring services, and ask it a question most finance reviews never pose: which of these costs is work that agents could carry, and what would that work cost in tokens? The answers tend to fund the transition. In plain terms, you are asking the AI to find the money that will pay for its own running costs.
Some evenings I open the routing dashboard and read the day’s decisions: Kairos choosing a cheap model for a summary, Neo escalating to Fable for a piece of reasoning that mattered, a few pence here, a pound there. It looks like a utility bill. Stripe looked at the same flow and saw a currency.
If this resonated, subscribe. The next anchor piece is about the job nobody warned you about: the unglamorous work that sits between owning intelligence and it doing anything useful for you.
Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.