The Broker Layer: Why Stripe Would Pay $10B for the Meter, Not the Model

On July 23, The Information reported that Stripe is in talks to acquire OpenRouter for something close to $10 billion. Two months earlier, in May, OpenRouter had raised a $113 million Series B led by CapitalG at a $1.3 billion valuation. Nothing about the company changed in those two months except the number of people who understood what it was.

The reflexive read is that this is a bubble artifact — a payments company with too much paper reaching for an AI logo. I do not think that is what happened. I think Stripe looked at the AI stack, asked which layer will still have pricing power in five years, and concluded that it is not the layer everyone is building data centers for.

It is the layer that counts.

What you are actually buying when you buy a router

OpenRouter is, mechanically, unglamorous. It is a single API endpoint that sits in front of more than 400 models from dozens of providers. You write your code once against OpenRouter, and it decides — or you tell it — which model actually serves each request. If a provider goes down, it fails over. If a cheaper model would do, it can route there. It handles the billing for all of it and takes a markup of roughly five percent on the inference spend that passes through.

That is the whole product. A switchboard with a cash register attached.

The volume through that switchboard is what makes it interesting. As of its May 2026 raise, OpenRouter was processing roughly 25 trillion tokens per week. Six months earlier that number was 5 trillion. It quintupled in half a year, and it did so during a period when the price per token was falling, which means the underlying growth in units was even steeper than the revenue curve suggests.

Fifty times in eighteen months, on a five percent take rate, during a period of collapsing unit prices. Work backward through the take rate and you get the number that actually matters: if roughly $50 million of revenue represents a five percent cut, then something on the order of $1 billion of annualized inference spending was flowing across that switchboard.

I want to be careful with that figure, because it is a derived estimate and not a disclosure. The revenue number is a third-party estimate pinned to early 2026. The token number is from May 2026. The take rate is reported as approximate and certainly varies by provider and contract. Blend them and you get a back-of-envelope average of well under a dollar per million tokens across the whole book — which is entirely consistent with a mix dominated by cheap open-weight models, and which is itself a clue about what has been happening to this market.

The point of the arithmetic is not precision. It is scale. A meaningful fraction of the world's independent inference spending — the portion not locked inside a hyperscaler contract or a direct enterprise agreement — passes through a company that, until this week, most engineering leaders could not have picked out of a lineup.

The three jobs hiding inside one API call

Before going further it is worth separating what a router actually does, because the bear case and the bull case are arguments about different jobs, and people talk past each other by conflating them.

The first job is abstraction. One API, many providers, consistent request and response shapes, so that swapping a model is a configuration change rather than a rewrite. This is genuine engineering value and it is also the most easily replicated of the three. Every gateway does it. Several open-source projects do it well and charge nothing.

The second job is arbitrage. Given a request, pick the provider that serves it best on whatever axis matters — cost, latency, availability, context length, regional compliance. This is where the intelligence lives, and it is worth something, but its value is inversely proportional to how different the options are. When one model is dramatically better for a task, choosing well is valuable. When forty models are within noise of each other and priced within noise of each other, choosing well is worth approximately nothing, because any choice is fine.

The third job is aggregation. Consolidate demand from many buyers into one relationship with each supplier, carry the credit and the float, and produce a single reconciled bill. This is the least technically interesting job and the only one that gets structurally harder for a competitor to copy as it grows, because its input is volume and volume is not a feature you can ship.

Almost every argument about whether routers are a real business is an argument about job two. Commoditization does destroy job two — that part of the bear case is simply correct. But job three is the one that scales into a franchise, and job three is strengthened by exactly the conditions that kill job two. The more interchangeable the suppliers, the more the buyer stops caring which one served the request and the more they care only about the aggregate bill. Which is another way of saying: the buyer stops relating to the model and starts relating to the meter.

Hold that distinction, because the rest of this argument depends on it.

The bear case is completely correct

Here is the part that makes the price look indefensible, and it deserves to be stated at full strength rather than strawmanned.

OpenRouter's core function is to put every model on the same shelf, sorted by price and capability, and let developers pick. Doing that job well is identical to commoditizing your own suppliers. The better the comparison shopping, the less any individual model can charge for being itself. And in a world where models are interchangeable and cheap, what exactly is the router for?

The evidence for that erosion is not speculative. It is visible in OpenRouter's own traffic.

Chinese open-weight models went from about 2 percent of platform traffic to about 45 percent in roughly a year. Meanwhile Anthropic accounts for something like 12 percent of traffic but 46 percent of revenue. Read those two facts together and you get a precise picture of a bifurcating market: an enormous and rapidly growing volume of cheap tokens doing undifferentiated work, and a small, expensive tier doing the work people will actually pay a premium for.

OpenRouter's own analysis found that the relationship between a model's benchmark scores and its actual market share on the platform is close to an inverse correlation. The leaderboard winners are not the volume winners. Cheap wins, and it wins overwhelmingly, and the router is the instrument that makes cheap findable.

Layer the macro price trend on top. The Stanford AI Index measured the cost of querying a model at a fixed capability level — roughly 64.8 percent on MMLU, the GPT-3.5 threshold — falling from about $20 per million tokens in November 2022 to about $0.07 per million tokens by October 2024. That is a reduction of more than 280-fold in under two years, for a constant amount of capability. Meanwhile the capability spread at the top has compressed, with the leading models on major leaderboards separated by a narrow band of points rather than generations.

So: suppliers converging in quality, collapsing in price, and increasingly open-weight. A marketplace whose function is to make that convergence legible. And a five percent tax on the whole arrangement.

The bear case writes itself. As models become fungible, choosing between them stops being hard. When choosing stops being hard, the thing that chooses for you stops being valuable. OpenRouter is digging its own grave with admirable efficiency, and Stripe is reportedly paying $10 billion for the shovel.

I think every sentence of that is true. I also think it leads to precisely the wrong conclusion.

Commoditization does not destroy the margin, it relocates it

Here is the flip.

When a supply market commoditizes, pricing power does not evaporate. It moves. It moves away from the thing that got commoditized and toward whatever is adjacent to it and scarce. This is one of the oldest patterns in industrial economics and it has played out so many times that we have stopped noticing it.

Electricity generation commoditized; the grid operator and the utility that owns the meter did fine. Long-distance bandwidth collapsed to near zero; the companies that owned the exchange points and the billing relationship did fine. Generic pharmaceuticals commoditized; the pharmacy benefit managers who sit between the maker and the payer became enormously profitable precisely because the drugs became interchangeable. Interchangeability is what creates the switching decision, and whoever owns the switching decision owns the customer.

The bear case assumes the router's value is the routing. It is not. The routing is the acquisition mechanism. The value is what accumulates underneath it: a per-request record of what every customer spends on intelligence, which models they use for which workloads, how their consumption is trending, and — critically — a billing relationship that survives any change in which model is currently winning.

Consider what that means for switching costs, which is where most people misjudge this. Switching models on OpenRouter is trivially easy; that is the product. Switching off OpenRouter is not, and the gap between those two facts is the entire business. Once your spending controls, your provider failover, your cost attribution across teams, and your single invoice all run through one integration, moving is not an API change. It is a finance migration.

That is the shape of durable infrastructure. Easy in, easy to change anything inside it, painful to leave. It is the same shape as the payments business Stripe already runs.

There is a useful historical rhyme here that is worth being precise about, because the loose version of it is misleading.

In the 1990s and 2000s, as long-distance telephony and then bandwidth commoditized, the money did not stay with the carriers who owned the fiber. It concentrated in two places: the interconnection points where traffic changed hands, and the billing and settlement systems that decided who owed whom for which minute. Carriers fought a brutal price war over the transport itself and mostly destroyed each other's margins doing it. The settlement layer was comparatively boring and comparatively durable.

The rhyme is not exact — inference is not transport, and models are differentiated in ways minutes never were. But the structural mechanism is the same one operating here. When the underlying unit becomes fungible and its price falls fast, the participants who compete on producing the unit compete away their margin, and the participant who sits at the exchange point keeps a small percentage of an enormous and growing flow. A small percentage of an enormous flow is a very good business. It is, specifically, the business Stripe is already in.

Why it is a payments company and not a cloud

The identity of the buyer is the most informative part of this story, and it is the part getting the least attention.

The obvious acquirers for a model router are the clouds and the data platforms. Databricks reportedly held early talks. Any hyperscaler could have made the case internally that owning the routing layer complements their inference business. None of them is the one reportedly closing.

Stripe is. And Stripe's position makes the logic clear.

Look at that sequence as one strategy rather than six events.

Stripe has spent two years building the rails for the agent-to-merchant transaction: the agent buys the sweater, the merchant gets paid, Stripe settles it. The Agentic Commerce Protocol is that thesis written down as a standard.

OpenRouter is the other half. It is the rails for the agent-to-model transaction: the agent consumes intelligence, the lab gets paid, someone settles it.

Own both and you are not in the AI business at all. You are the settlement layer for an economy in which software is both the buyer and the seller — collecting on the intelligence an agent consumes and on the goods that agent then purchases. That is a coherent, enormous, and genuinely defensible franchise, and it is worth considerably more than a routing convenience.

It also explains the multiple. Stripe is not underwriting $50 million of marketplace revenue growing quickly. It is underwriting the possibility that inference settlement becomes a payment network — and payment networks are valued on the volume they touch, not the margin they currently extract from it. At roughly $1 billion of annualized inference flow today, in a market whose token volume quintupled in six months, a $10 billion price is a bet on the flow compounding into something that looks like a rail rather than a reseller.

Whether that bet is correct is a separate question. But it is not a confused one.

The meter only matters if it is the meter of record

The thesis has a hard requirement, and it is worth naming precisely because it is where the deal could fail.

Being a router is not enough. Plenty of things route. The category is crowded with capable alternatives: LiteLLM if you want an open-source gateway inside your own infrastructure, Portkey if you want production-grade routing with guardrails and budgets, Vercel and Cloudflare if you want a managed edge service with zero operations, Bedrock or Vertex if you have already committed to a cloud and want routing inside its compliance boundary. Several of these are free. Some of them are better at specific jobs than OpenRouter is.

What none of them has, at OpenRouter's scale, is position as the financial system of record for inference spend across an entire market.

The distinction is not academic. It is the difference between a business that gets competed to zero and one that gets more entrenched every quarter, and OpenRouter today sits somewhere between the two. Its 400-model catalogue and its 25 trillion weekly tokens are settlement-layer assets. Its five percent markup on easily-comparable commodity inference is gateway economics, and gateway economics erode.

Stripe's contribution to that gap is the obvious one. Stripe knows how to turn a flow of transactions into a financial position: underwriting, treasury, risk, reconciliation, revenue recognition, and the unglamorous compliance apparatus that makes a finance department comfortable letting a third party stand between it and a major cost center. That is the machinery that converts a router into a network, and it is machinery OpenRouter would have taken years and considerable capital to build alone.

This is also why the buyer is not a cloud. A hyperscaler acquiring the routing layer inherits an obvious conflict: it would own both a marketplace that is supposed to route to the cheapest option and an inference business that wants the traffic. Neutrality is the product, and only a buyer with no model of its own can credibly keep selling it. Stripe has no model. That is a feature.

The neutrality argument, stated carefully

Neutrality is doing a lot of work in this thesis, so it deserves scrutiny rather than assertion.

A neutral intermediary is only valuable if buyers actually believe the routing recommendation is not corrupted by the intermediary's own economics. And a five percent markup on pass-through spend is not perfectly neutral either: a broker paid a percentage of spend has a mild structural interest in spend being higher, which is the same conflict that has dogged percentage-of-spend advertising agencies and benefit managers for decades. Buyers eventually notice this. Some of the sharpest criticism of pharmacy benefit managers — the closest structural analogue to what a mature inference broker becomes — is precisely that their incentives diverge from the payers they nominally serve once they are large enough to shape the market rather than just navigate it.

So the honest form of the neutrality claim is narrower than the marketing form. The router is credibly neutral between suppliers, because it does not make a model and does not care which one wins. It is not automatically aligned with the buyer on total spend. That gap is manageable while the market is growing and buyers are focused on capability, and it becomes a live issue the moment AI budgets stop growing and procurement starts auditing. Any acquirer paying $10 billion for this position is buying that future argument along with the asset — and Stripe, whose entire business model is a transparent percentage of volume that customers have long since made peace with, is unusually well practiced at having it.

The four ways this thesis breaks

I would rather state the failure modes plainly than pretend a $10 billion bet on a rumored deal is a sure thing. There are four, and they are not equally likely.

The labs disintermediate deliberately. The most direct threat is that the frontier labs, watching 46 percent of a marketplace's revenue flow through their models while a broker takes a cut and actively shops customers toward cheaper substitutes, decide the arrangement is unacceptable. They can price direct relationships below the routed price, withhold newest capabilities from aggregators, or attach terms that make resale unattractive. This is the risk every marketplace runs with a concentrated supplier base, and OpenRouter's supplier base is very concentrated at the revenue-generating end.

The hyperscalers bundle it to zero. Routing inside a cloud is a feature, not a product. If AWS, Google and Microsoft each ship a competent neutral-enough router included with a commitment an enterprise has already signed, the standalone router is left selling to the segment that has not committed to a cloud. That segment is real, growing, and disproportionately made of the startups building agents — but it is not the whole market, and it is the segment most sensitive to a five percent markup.

Margin compression outruns volume growth. The five percent take rate is under structural pressure from both directions: suppliers want more of the spend, and buyers can see exactly what the markup costs them because the underlying list prices are public. A router's markup is uniquely legible compared to, say, an interchange fee. If the rate compresses toward two percent while unit prices keep falling, volume has to grow faster than both to keep revenue rising.

The relationship never actually consolidates. The whole thesis assumes the router becomes the system of record. But many organizations will run models through several paths at once — a cloud contract for regulated workloads, a direct lab agreement for the flagship model, a router for everything experimental. If routed spend stays a minority of the bill, the router never becomes the meter of record for the enterprise, and it stays a useful tool rather than a financial position.

None of these is fatal on its own. All four operating at once is the bear case made real, and the honest assessment is that the first two are already underway. The bet Stripe is reportedly making is that volume growth and settlement entrenchment compound faster than those forces subtract — and that a neutral, financially-sophisticated intermediary is structurally advantaged against both a lab that wants to lock you in and a cloud that wants to bundle you.

What this means if you are shipping on top of a router

Strategic analysis is worth very little if it does not change what you do on Monday, so here is the practical translation for anyone with production traffic running through a gateway.

Treat your router as a financial dependency, not a technical one. The technical dependency is trivial to unwind; the financial one is not. Before your routed spend becomes a material line item, make sure you can answer: can I export a complete per-request spend history, can I reproduce my cost attribution independently, and do I have a direct contractual fallback with the two or three providers that account for most of my bill? If the answer to any of those is no, you are building on a position you do not control, and the acquisition announcement is a good prompt to fix it. This is the same discipline I argued for in my analysis of how token pricing is engineered to feel cheap, applied one layer up the stack.

Separate the routing decision from the settlement decision. These get bundled because one vendor sells both, but they are different problems with different lifespans. You may want intelligent routing forever and want to change who bills you next year. Architecting so that your provider abstraction and your billing relationship are not the same integration costs very little up front and preserves an option that is expensive to buy back later.

Expect the markup to become negotiable, and negotiate it. Five percent on commodity inference is a price set when nobody was looking. As routed spend becomes a visible line item in AI budgets, it will get the same scrutiny as every other percentage-of-spend fee eventually gets. If you are pushing meaningful volume, that rate is a conversation, not a constant — particularly during an acquisition, when the acquirer wants retention numbers.

Instrument your own meter. The single most valuable defense against any intermediary becoming your system of record is having your own. If your telemetry already attributes token spend to teams, features and workloads independently of your vendor's dashboard, then switching intermediaries is an inconvenience rather than a loss of institutional memory. The control plane that agent gateways are becoming is a governance asset, and you want to own a copy of it.

Model the run, not the call. Every dynamic in this article is amplified by the shift from single requests to long-running agents, where a single dispatched task can consume billions of tokens across hundreds of calls. The intermediary's cut compounds with the number of calls, which is exactly the number that is growing fastest. I worked through that cost structure in detail in the run-cost era, and it is the reason the routing layer's volume curve looks the way it does.

The layer nobody was watching

There is a pattern worth naming here, because it keeps recurring and the market keeps being surprised by it.

Every few months the AI industry has an argument about where the value is going to settle. The candidates are always the visible ones: the frontier labs, the chips, the clouds, the applications. Enormous sums move on those debates. The compute allocation turn was one version of that argument; the endless model-versus-application framing is another.

And then the answer turns out to be a layer nobody was arguing about, because it was too boring to have a faction. Not the model. Not the chip. The switchboard with the cash register.

That is not a coincidence, and it is not luck. It is the consequence of commoditization doing what commoditization always does. When the exciting layer gets cheap and interchangeable, the value migrates to whatever is adjacent to it and hard to replicate — and "hard to replicate" in a market of interchangeable suppliers means the relationship, the aggregated demand, and the ledger. Those three things are what a broker owns. They are unglamorous, they are invisible from the outside, and they are the reason a payments company is reportedly willing to pay frontier-lab money for a piece of middleware.

The deal may still fall apart; the reporting is explicit that talks could collapse, and rumored acquisitions fail all the time. But the price has already done its work as information. Somebody with $1.9 trillion of annual payment volume and an unusually good view of how autonomous software spends money looked at this stack, and decided the durable asset was not the intelligence.

It was the meter that counts it. I expect more of these acquisitions before the end of 2027, and I expect the buyers to keep coming from outside the AI industry, for the same reason this one did: they are the only ones who can stay neutral, and neutrality is the product.

Further Reading

• The Free Sample: how AI token pricing is engineered to feel cheap — the buyer-side view of the same meter
• The Run-Cost Era: why agents cost more to run than to build — the consumption curve driving routed volume
• The Control Plane Arrives: how agent gateways govern production AI — the governance half of the same layer
• My prediction on gateway and router consolidation through 2027