Audio editionListen to this book6 chapters · 50 min
01
A chief executive says tokens are a currency. His own company's products are built as though they are not.
On 19 August 2026, Stripe announced it had agreed to acquire OpenRouter.c-001 The press release described the target in a single line — a platform that routes and optimizes token usage across more than four hundred models from more than eighty providersc-011 — and then handed the microphone to Patrick Collison, who explained what the company thought it was buying.
Tokens are the central currency for companies building with AI, and it’s clear that the real-world economic potential will depend on making good use of scarce compute resources.c-009
It is a good sentence, and I suspect it is the sentence that sold the deal internally. It is also wrong in a specific way that the rest of this book is about — and the strange thing is that Stripe’s own products appear to know it, because not one of them is built as though it were true.
Consider the thing Stripe just paid billions for. OpenRouter runs a credit system: you load money in, you spend it down on inference. In what unit is that balance held? The documentation answers without ceremony. OpenRouter uses a credit system in which the base currency is US dollars.c-015
Not tokens. Dollars.
The competition does the same. Vercel’s AI Gateway sells credits too, and bases its rates on the provider’s list price, with charges deducted from a dollar balance.c-047 And Stripe’s own AI billing product, which shipped before this acquisition, is the clearest tell of all: it lets a business define an “AI Credit” and then configure how usage of each model converts to credits.c-018 Converts. From one thing into a different thing, at a rate somebody has to set. That verb is carrying the entire argument.
Every serious system in this market meters in tokens and settles in dollars. Nobody has built one that does otherwise, and it is not for want of ambition. The token is the meter. The dollar is still the money.
What a currency has to do
The reason is less about finance than about measurement, and it starts with a property so basic that it usually goes unstated: a unit of account has to mean the same thing to both parties at the moment they agree to a trade.
A dollar is a dollar when you commit to spend it. A kilowatt-hour is a kilowatt-hour, which is why your electricity bill works — you do not know in advance how many you will burn, but you know what one is, and you know its price before you use it. That is the whole trick of a meter. The quantity floats; the unit does not.
A token fails this, and it fails it in four different ways at once, none of them exotic and all of them documented by the vendors themselves in the pages they expect their own customers to read.
The first is that a token’s price depends on who serves it. For open-weight models the weights and the hardware come apart: a lab publishes a model, and then dozens of companies run it on their own machines, at their own prices, with their own throughput. The output is identical. The invoice is not. OpenRouter’s routing documentation puts the spread in its own worked example — Provider A at one dollar per million tokens, Provider B at two, Provider C at three.c-025 A factor of three, for the same tokens out of the same model.
The second is that you do not get to choose which of them serves you, and neither, quite, does the router. You might reasonably assume it picks the cheapest and moves on. It does not. The documented default is to load balance requests across providers, prioritizing pricec-024 — balance, note, not minimize — and the balancing is probabilistic, with candidates weighted by the inverse square of their price. In the documentation’s own example that makes the cheapest provider nine times likelier to be selected than the dearest.c-026
Nine times more likely still sends roughly one request in ten to a costlier supplier. The die is loaded; it remains a die. Which means that at the instant you press send, the price of that request is not yet a number — it is a distribution, and it will not collapse into a number until after the work is done.
This is the load-bearing fact of the book and it is worth slowing down for, because it is stronger than the thing people usually say. The complaint is not that AI pricing is hard to predict, or that it varies by workload. It is that the price does not exist yet at the moment of purchase. It gets settled afterwards, by a draw you neither control nor observe.
There is a good reason for the design, and it is worth stating plainly rather than treating the nondeterminism as a flaw somebody should fix. Routing every request to the single cheapest provider would pile load onto one supplier, wreck tail latency, and hand that supplier the power to reprice at will. Spreading traffic across a price-weighted field is how a router keeps its own supply side honest. The nondeterminism is deliberate, and it is load-bearing for the market’s health. It is simply fatal to using the token as a unit of account.
The third failure is that not all tokens are the same kind of token. Prompt caching lets a provider hold part of your context so that repeated calls do not reprocess it, which makes it the single largest lever on real-world inference cost — and which splits one token into several priced categories. On recent OpenAI models, cache writes are charged at 1.25 times the price of the original input pricing.c-027 Cache reads are charged at a quarter or a half of it, depending on the model.c-028 So a token going into the cache costs a quarter more than an ordinary input token, and the same token coming back out costs a half or a quarter of one. Same word, same model, same request, three prices, decided by which side of a cache boundary it happened to fall on.
If those categories were at least consistent between vendors, a conversion table would rescue the situation. They are not, and that is the fourth failure. The same documentation records that some providers — Anthropic among them — carry a negative discount on cache writes and a positive one on cache reads.c-029 A negative discount is a surcharge. One vendor’s cache write is a rebate and another’s is a penalty, with no shared sign convention to appeal to. And because caching and routing are entangled — the router pins a session to a provider only when that provider’s cache-read price beats its ordinary prompt pricec-030 — the caching behavior changes the routing, which changes the price, which changes whether the caching was worth doing.
Where this bites
Put the four together and you arrive somewhere uncomfortable: you cannot quote a fixed price for a request before it is routed. Not because the tooling is young. Because there is no fixed price there to quote.
Anyone who has tried to reconcile an AI bill has already met this, usually without having a name for it. The most honest acknowledgment I have found sits in the documentation of LiteLLM, the open-source gateway that a great many teams use to track their spend. It carries a section headed, without euphemism, “Cost does not match your provider bill?” — and beneath it a debugging workflow instructing you to align time ranges, compare token categories including cache, and then decide whether the gap is ingestion, formula, or model-map pricing.c-031
Read that last clause again. Three named ways your own meter can disagree with your supplier’s, catalogued by a mature tool that does this for a living. Metering AI spend is a genuinely hard measurement problem rather than a solved one waiting on a better implementation, and the software that does it best is candid about that.
The cost of getting this wrong is rarely a catastrophe. It is a slow tax, and it comes due in four places. Budgets drift, because a forecast built on average cost per token moves as the routing mix moves — and the drift is invisible, since the token count was right all along. Margins erode quietly: if you resell inference at a fixed price per unit while your own cost is a distribution, you have written somebody an option and handed it over for free. Reconciliation becomes permanent, a monthly ritual of comparing two numbers that will never agree and deciding how much disagreement is tolerable. And optimization goes after the wrong quarry, because cutting token count is not the same as cutting cost once cache reads are a quarter price and cache writes are a surcharge. Teams tune the meter when they meant to tune the bill.
None of that is exotic. It is just what happens when you denominate something in a unit that will not hold still.
What got built instead
So the interesting question is not why nobody has minted a token currency. It is what the people closest to the problem built once they worked out that they could not.
They built a conversion layer. Stripe’s billing product does not sell tokens; it lets a business define an abstract AI Credit, specify how each model’s usage converts into that unit, and then sell prepaid packs of it.c-018 The metering underneath is fine-grained enough to carry every distinction this section has been describing, segmented by model and by token type including cached tokens.c-019
That is the right architecture, and it concedes the whole argument. The credit is a synthetic unit rather than a token, with an exchange rate against real tokens that is configured per model and can be changed by whoever is selling it. The complexity does not vanish. It moves behind an interface, where one business absorbs it so that its customers never have to look at it. At the time of writing, that product was not generally available; access sat behind a waiting list.c-020 The idea has run further ahead than the shipping.
If you take one thing into your own systems from this section, make it the discipline of keeping the two units apart. Store budgets in dollars and derive tokens, never the reverse — any internal system holding a budget as a token count has a bug that will stay hidden until the routing mix shifts under it. Stamp an access date on every price you rely on, because vendors reprice without changelogs and a price page is evidence about a day rather than a standing fact. Capture the cached-versus-uncached split in your own telemetry from the beginning, since retrofitting it after a surprising invoice means reprocessing logs that may never have kept the field. And decide in advance how much drift between your meter and your supplier’s you are willing to accept, then write the number down — teams that skip this end up settling it during a month-end close, where the answer gets decided by whoever is most tired.
02
Four reported prices, none of them disclosed, and one fee structure that explains the interest better than any of them.
Nobody knows what Stripe paid for OpenRouter.
That is worth stating flatly at the outset, because the number gets repeated in conversation as though it were settled, and it is not. Stripe disclosed no terms.c-003 What exists instead is four figures from four outlets, each citing a source it does not name. Bloomberg reported more than seven billion dollars.c-004 The New York Times, via CNBC, put it at about $7.5 billion, with $1.5 billion of that allocated to OpenRouter’s founders.c-005 Semafor said eight billion.c-006 And earlier, while the talks were still running, The Information had the figure at ten billion.c-007
The temptation is to average these and move on. Resist it. They are not four measurements of one quantity, and averaging them would produce a number that nobody reported and no document supports. They are snapshots of a negotiation taken at different moments by different people with different access. If you need a number for an argument, use a range and say whose it is. If your argument only works at eight billion, it is a weaker argument than you think, because eight billion is the loosest of the four.
What is not in dispute is the shape of the jump. Less than three months before the acquisition, OpenRouter had raised $113 million at a valuation of about $1.3 billion.c-008 Whatever the closing figure, the company was worth something on the order of five to six times more in August than it had been in May, and nothing about the product had changed in the interval.
One more thing was true at the time of writing and may not be by the time you read this: the acquisition had not closed. It remained subject to customary closing conditions, with the parties expecting to close within weeks.c-002 Everything in this book describes an agreed deal rather than a completed one.
So the question is what the buyer saw.
The shape of the take
The usual answer is distribution: OpenRouter sits in front of a great deal of traffic, and traffic is valuable. That is true but incomplete, and it misses the detail that should make a payments company sit up.
Look at how OpenRouter actually earns. It does not mark up inference. Its documentation is explicit that there is no markup on inference pricing, and that it charges a fee when you purchase credits instead.c-012 What you pay for tokens is what the underlying provider charges for tokens. The business is instead a percentage taken at the moment money enters the system: the published platform fee on the pay-as-you-go plan is 5.5%.c-013 For customers bringing their own provider keys the arrangement shifts but does not change character — free up to $25,000 of list-price inference per month, then a 5% fee above that.c-014
Sit with the structure rather than the numbers. A software company charges for the software. A reseller buys wholesale and sells retail, and lives on the spread. OpenRouter does neither. It passes the goods through at cost and takes a percentage of the money moved.
That is a take rate. It is the economic signature of a payment processor, and it is a genuinely unusual thing to find in an infrastructure company. The fee does not scale with how much compute you consume, or how many models you call, or how many seats you have. It scales with how much money you put in. Whether you spend that balance on one enormous batch job or ten million small requests, whether you burn it in a day or leave it sitting for a quarter, the 5.5% was collected at the door.
A company whose entire business is understanding money at the door would recognize that instantly.
The economics that come with it are the kind payments people find familiar. OpenRouter was reported to be generating roughly $140 million in annualized revenue by mid-2026,c-016 with the cost of serving the routing product at 28.5% of revenue and gross margin around 70%.c-040 That margin is not the product of clever pricing on inference, because there is no pricing on inference. It is what falls out of taking a percentage of transaction value while paying only for the infrastructure that moves the transaction.
The part that was already there
The strongest evidence that Stripe understood this is that Stripe had already started building it.
The press release says so, in a clause that reads like housekeeping and is not: since the previous year, Stripe had been working to help companies optimize their token costs and route tokens efficiently, having launched products including Token Billing.c-017 Before any of this, Stripe was already metering AI tokens as a billing primitive. It had shipped the meter. What it lacked was the thing that knows, request by request, which of eighty providers served the work and what that provider charged — and that knowledge is not something you can build from the outside, because it lives in the routing decision itself.
Read the acquisition that way and it stops looking like a payments company wandering into AI infrastructure. It looks like a metering business acquiring the instrument its meter was missing.
There is a further clue in how Stripe describes the difficulty. The release frames the problem not as cost but as combinatorial complexity — which model to use for which tasks, at which speed and at what price — and notes that the challenge is exacerbated by the pace at which models are released and repriced.c-075 That is a description of an unstable unit of account written by someone who has clearly stared at the problem. And the stated ambition is to help businesses maximize profitability by routing their requests intelligently and spending their tokens efficiently.c-010 Revenue and cost. Money in and money out. The two sides of a ledger, which is the one thing Stripe has always sold.
Where the argument gets uncomfortable
All of which supports the reading that this was a payments acquisition, and I think that reading is right. But it does not support the further claim that gets attached to it — that the token is therefore becoming a currency, and that owning the router is owning the mint.
The same company that charges a payment processor’s take rate settles it in dollars. OpenRouter’s credit balance has US dollars as its base currency.c-015 The 5.5% is levied on dollars entering, not on tokens leaving. At no point in the flow does a token function as money. It functions as the thing being measured so that dollars can be moved correctly, which is exactly what a meter is for.
Everything here about the rationale comes from parties with an obvious interest in the rationale sounding coherent. A press release is a first-rate source for the fact that a deal happened and a poor one for why. Both companies had every reason to present a legible strategic story in August 2026, and neither had any reason to say that they paid a large multiple because the asset was scarce and someone else might have taken it.
The take-rate structure is different, and it is why this section leans on it. That is not a story anyone told me. It is on a price page, it was there before the deal, and it would be there whether or not the acquisition made sense.
What follows from it is a question rather than a conclusion. If the position is genuinely a payments position, the 5.5% should be defensible — rents at a chokepoint usually are. Section 4 tests that, and finds the picture considerably messier than the number suggests. Before that, though, it is worth understanding what the router is actually doing when it decides where your request goes.
03
A component that selects suppliers by price, at volume, is not observing a market. It is setting one.
A router looks like plumbing. Requests arrive, requests leave, and the interesting work is assumed to be happening at the ends. That framing survives right up until you notice that the router is choosing, on every single request, which supplier gets paid — and that it makes the choice on price.
Once a component selects suppliers by price and controls enough volume, it stops being a passive observer of a market. Supplier revenue is now a function of the router’s weighting. That is the argument of this section, and the way to see it is to look closely at what the routing decision actually involves, because there are far more priced dimensions in it than most people building on top of it realize.
More dials than anyone expects
Start with the one from Section 1: provider selection weighted by the inverse square of price, so that the cheapest candidate is nine times likelier than the dearest in the documentation’s own three-provider example.c-026 That alone makes the cost of a request a distribution rather than a number.
But price is only the default sort. A caller can filter on measured performance first and then sort on price, and the filtering is empirical rather than advertised — OpenRouter will select hosts that have sustained at least 40 tokens per second for 90% of requests over the last five minutes, and route to the cheapest provider within that qualified group.c-064 That is a rolling five-minute window over a ninetieth percentile. The set of providers eligible to serve you changes continuously, based on how they have been performing against other people’s traffic in the last few minutes. Two identical requests, sent ninety seconds apart, can face different candidate sets and therefore different price distributions, without anything in your code changing and without any provider having announced a price change.
Then consider what else carries a multiplier once you look beyond the router at the vendors underneath.
Latency tolerance is priced. Amazon Bedrock offers selected models for batch inference at a 50% lower price than on-demand inference.c-063 Half price, for agreeing to wait. If your workload can tolerate a delay and you have not told anyone, you are paying twice what you need to.
Geography is priced. Anthropic applies a 1.1x multiplier across all token pricing categories when a caller specifies US-only inference.c-060 Not a surcharge on a line item — a ten percent uplift across input, output, cache writes and cache reads alike. A compliance requirement, expressed as a coefficient.
The cache is the economy
Section 1 gave the OpenAI figures: cache writes at 1.25x input on recent models, cache reads at a quarter or a half. Those figures belong to one vendor rather than to the industry, and the others differ.
Anthropic’s published table runs on three multipliers rather than two, because it sells cache duration as a product. A five-minute cache write costs 1.25x base input price. A one-hour cache write costs 2x. A cache read costs 0.1x.c-058 Bedrock, meanwhile, prices cache reads at 75% below its on-demand input pricec-062 — a 0.25x multiplier, which is neither Anthropic’s 0.1x nor the 0.5x that some OpenAI models carry.
Three vendors, three different discounts for the identical operation of reading a cached token. There is no convergence here and no sign of any, which means a tool that wants to predict cost across vendors cannot hold a single number for “cache read.” It has to hold a table, and keep it current, against vendors who reprice without changelogs.
Anthropic, to its credit, publishes the break-even directly: caching pays off after one cache read for the five-minute duration, or after two cache reads for the one-hour duration.c-059 That is an unusually honest piece of documentation and it is worth dwelling on, because it makes the decision structure visible. Whether to cache is not a performance question with a cost side effect. It is a bet on reuse. Write to the five-minute cache and you need one hit to break even. Buy the hour and you need two. Get the reuse pattern wrong and caching costs you money — which is the opposite of how almost everyone describes caching in every other part of computing.
Now couple that back to routing, because the two are not independent. A cache lives on a specific provider’s hardware. To benefit from it, subsequent requests have to reach the same provider, which is why the router pins sessions — and pins them conditionally, activating sticky routing only when the provider’s cache-read price beats its ordinary prompt price.c-030 The router is running the break-even calculation on your behalf and changing its routing policy based on the answer.
The entanglement produces failure modes that look like nothing at all in a diff. Changing a model slug mid-conversation — appending a latency-oriented variant suffix, say — invalidates the sticky key.c-065 A one-word change to a configuration string, made to improve latency, silently discards a warm cache and re-routes the conversation. The latency probably does improve. The bill moves in a direction nobody predicted, because the person who made the change was reasoning about speed and the system was also reasoning about cache residency.
What this makes the router
Assemble the dials: price-weighted probabilistic selection, rolling throughput filters, batch tiers, geographic multipliers, three-way cache multipliers that differ per vendor, and sticky sessions that engage conditionally on a price comparison. Every one of those is a lever on what a request costs, and the router holds most of them.
A component holding that many pricing levers over that much volume is not a neutral pipe. When a router weights by inverse square of price, it is not merely finding the cheap provider — it is manufacturing demand elasticity. A provider that drops its price does not simply win the requests where it is cheapest; it wins a larger share of a probabilistic draw across the whole eligible set, immediately, with no sales cycle and no contract renegotiation. And a provider that raises its price does not lose its customers, because it never had customers in the retail sense. It loses weight in a distribution.
That is a real market mechanism, and it is a better one than the alternative in most respects. It is also, from the perspective of anyone trying to answer “what will this cost,” an active adversary. The very properties that make the routing layer good at disciplining suppliers are the properties that make a fixed price impossible to quote.
Everything above describes documented mechanism, not measured outcome. I have not seen data on how much the price distribution actually moves in practice for a typical workload, and I am not aware of any published study that measures it. It is entirely possible that for most applications the variance is small enough to ignore, and that the practical answer is to use a trailing average and stop worrying.
The mechanism is certain; the magnitude is not, and nobody should take a claim about magnitude from this book, because this book does not have one.
What the mechanism does establish is where the leverage sits. If a supplier’s revenue depends on a weighting function held by an intermediary, the intermediary is the most commercially valuable component in the chain — regardless of whether it charges anything for the privilege. Which raises the question the next section has to answer: given that position, what does the market actually charge for standing in it?
The answer turns out to be four incompatible things at once.
04
Five firms, one function, and prices that range from 5.5% of everything you spend to nothing at all.
If the routing layer is as valuable a position as the last two sections suggest, the market should agree on what it is worth. Markets usually do converge on a price for a well-understood function, even a contested one — card acquiring settled into a band, cloud storage settled into a band, and you can predict what a new entrant will charge to within a few basis points.
The AI gateway market has not converged. It has not even nearly converged. The same function — sit between an application and a set of model providers, route the request, meter the spend — is monetized four incompatible ways by five serious companies, and the prices are not close.c-069
OpenRouter charges a share of the money you load: 5.5% on the pay-as-you-go plan.c-013
Vercel charges nothing on the tokens. Its documentation states plainly that the AI Gateway charges no markup and no platform fee on tokens, with the customer paying the provider’s list price and Vercel deducting from a credit balance.c-021
Portkey charges a subscription. Its production plan is $49 a month with a quota of 100,000 recorded logsc-067 — a software license with usage-based logging, entirely decoupled from how much inference flows through it. Below that sits a permanently free tier, explicitly not suitable for production workloads.c-068
Cloudflare charges nothing at all. Its AI Gateway’s core features are offered free on all plans, and the core it names is not a stub: dashboard analytics, caching, and rate limiting.c-066
And beneath all of them, LiteLLM gives the metering away as open-source software you run yourself, tracking spend for keys, users, and teams across more than a hundred models.c-023
Five providers. A percentage of funds, zero, a flat monthly fee, zero again, and free source code.
What a spread like that means
A price dispersion this wide is not noise. It is a signal that the participants disagree about what business they are in.
Take a percentage of money loaded and you have declared yourself a financial intermediary whose revenue scales with customer spend. Charge a flat subscription and you have declared yourself a software vendor whose revenue scales with customer count. Charge nothing and you have declared the gateway a feature of something else you sell — hosting, in Vercel’s case; the broader network, in Cloudflare’s. Publish the source and you have declared it infrastructure that should not be a business at all.
Those are four different businesses wearing the same coat. And it means the 5.5% is not the market price for routing. It is one firm’s answer to a question the market has not settled, and it is the highest answer on the board by a wide margin — infinitely higher than three of the four alternatives, which charge nothing for the gateway itself.
I want to be careful about what this does and does not establish. It does not establish that the 5.5% is doomed. Prices persist above zero-priced alternatives all the time, for reasons that are perfectly rational: switching costs, breadth of model coverage, reliability, the fact that the free options make you operate them yourself. OpenRouter’s fee buys access to a catalogue and an operational burden somebody else carries, and plenty of teams will consider that a bargain at 5.5% of spend. The presence of a free competitor does not refute a paid one; if it did, nobody would sell databases.
What it does establish is that the fee is a business-model choice rather than a property of the position.c-051 Occupying the routing layer does not entitle you to a take rate. Two of the largest firms in this list occupy the same layer and have decided the correct rent is nothing, because they monetize elsewhere and would rather have the traffic.
And it complicates the payments framing from Section 2 in a specific way. A payment processor’s take rate is defensible because moving regulated money at scale is genuinely hard, genuinely licensed, and genuinely not something a competitor can give away as a loss leader. A routing fee has none of those protections. Cloudflare can zero it out on Tuesday and absorb the cost inside a network business, and there is no regulator, no license, and no settlement obligation standing in the way.
The other pressure
The competitive squeeze is the visible half. The other half is that the thing being charged for does not work perfectly, and its own practitioners say so.c-054
Section 1 quoted LiteLLM’s documentation on reconciliation. Read in the context of this section that admission carries more weight. If metering were both hard and reliable, it would be a defensible service. If it were easy and reliable, it would be a commodity but at least a working one. What the documentation describes is a capability that is available for free and does not fully work — which is close to the worst possible configuration for anyone hoping to charge rent on it.
Sacra’s analysis of OpenRouter names the pressure directly, observing that open-source alternatives such as LiteLLM let enterprises self-host routing infrastructure, reducing willingness to pay for standalone routing.c-045 That is a private-market research firm modeling a company it does not have access to, so treat it as a considered opinion rather than a finding. But it is the same conclusion the price list arrives at from the other direction.
I have five prices, gathered because they were reachable, and no systematic survey of the gateway market. There are more gateways than these — hyperscalers bundling routing into existing contracts, observability vendors extending into it, enterprise platforms building it internally. I do not know the shape of the whole distribution, only that its visible members disagree violently.
A proper market survey would strengthen or overturn the argument here, and it has not been done.
What can be said with the evidence in hand is narrower and still useful. The fee at this layer is contested, several well-capitalized firms have set it to zero, the underlying capability is available as free software, and that software concedes it does not always agree with the invoice. Whatever Stripe bought, it was not an uncontested toll booth.
Which leads directly to the strongest argument against this entire book — the possibility that everyone involved already knows all of this, and that the routing fee going to zero was never the point.
05
The strongest case against this book, presented without a rebuttal appended, because I could not build one.
Everything so far has built toward a conclusion: that the routing layer matters because it is where compute becomes a billable quantity, and that a payments company bought it for that reason. This section sets out the strongest case that the conclusion is wrong, or at least that it is right for reasons that make it far less interesting than it sounds.
I am not going to answer it. I could not, with the evidence I gathered, and appending a rebuttal I cannot support would be worse than leaving it standing.
The case that routing is meant to be worthless
Start with the version Semafor published two days after the announcement, which is the sharpest reading of the deal I have seen anywhere.
The observation is that OpenRouter’s business could be destined for commoditization, and that this is actually kind of the point.c-041 The reasoning follows from the market structure described in Section 4 rather than contradicting it. If routing collapses into finding the cheapest provider that clears a quality and reliability bar, then routing is a solved problem with a shrinking margin, and everyone in it is heading for zero. But zero is survivable — attractive, even — if you are not trying to earn there: Stripe can afford to let routing become low-margin, because it can monetize everything surrounding it, from billing and tax and fraud through to stablecoin settlement, treasury and possibly financing.c-042
Take that seriously and the thesis of this book inverts. The routing layer is not the valuable position; it is the entry point to a set of valuable positions, none of which is routing. Stripe does not need a take rate on inference because it has billing, tax, fraud, settlement and treasury waiting on the other side of the relationship, all of which are businesses it already knows how to run at scale. Under this reading, buying OpenRouter is buying distribution into every AI-native company’s financial stack, and the fact that the acquired product happens to route tokens is almost incidental.
Notice what this does to Section 4’s argument. I presented the four-way price dispersion as a problem for the payments framing — evidence that the fee is contested and might go to zero. Semafor’s reading takes the identical fact and calls it the plan. Under this account the 5.5% is not a rent to be defended but a legacy price that a rational owner will happily compete away, because the customer relationship is worth more than the fee ever was. The same evidence, pointing the opposite direction. I do not have a way to distinguish the two from outside.
The case that it is an expense-management deal
The second reading is less dramatic and possibly more likely.
TechCrunch noted that Stripe’s large acquisitions had historically been about helping companies collect and manage incoming money, and that this one looked different — a move to the other side of the ledger, into expense management, beginning with AI expenses.c-043
That reframes the deal as being about cost control software, which is a real and large category with nothing much to do with payment rails. And the competitive evidence supports it, because that category was already filling up before Stripe arrived: Databricks developed its own AI gateway, Rippling launched one focused on employee AI spend, and Ramp launched one for AI expense management too.c-044
Rippling and Ramp are not infrastructure companies. They are spend-management companies, and they built AI gateways because AI spend is a line item their customers cannot see clearly. If the natural competitor set for this product is Ramp rather than Cloudflare, then the position Stripe bought is a seat in the expense-management market, and the elaborate argument about units of account is a description of a technical problem rather than an explanation of a purchase.
The uncomfortable part, for me, is that this reading explains the price better than mine does. Expense management is a proven category with known multiples and an obvious buyer synergy. Owning the settlement layer for a future compute economy is a story that requires several things to happen that have not happened yet.
The case that the price was simply too high
And then there is the arithmetic, which is unkind whichever story you prefer.
At the figure floated during the talks, Stripe would have been paying roughly seventy times OpenRouter’s annualized revenue.c-038 Seventy times. The same reporting describes that price as high compared to other AI acquisitions when calculated as a multiple of forward revenue.c-039
The defense is that the unit economics are unusually good for the category, and they are — Section 2 set out the margin structure. But a strong margin on $140 million of revenue does not obviously justify a price several dozen times that revenue unless you believe the revenue is about to become very much larger, which brings you back to needing the strategic story to be true.
There is also a piece of evidence about the buyer’s state of mind that is hard to categorize and too good to omit. A leaked letter from Stripe’s founders to investors explained the company’s operating posture by saying it had decided that 1 January marked the beginning of the singularity and it had been operating on that basis.c-076 The reporting is careful to note the phrase is tongue-in-cheek and that Patrick Collison had used it in that spirit publicly.c-077 I am not going to build an argument on a joke in a letter. But a firm operating on an explicitly stated assumption that the technological curve has gone vertical is a firm whose acquisition multiples should be read in that light, and it is at least worth knowing that the assumption was written down.
What would have to be true for this book to be right
Set the three counter-readings side by side and the honest position is that all of them are consistent with the evidence I have.c-052
For the argument in this book to hold, the metering-and-settlement position has to be worth more than the expense-management seat, and the router has to remain the place where the metering happens rather than being reduced to a commodity component inside somebody else’s billing stack. Neither is established. Both are assertions about a market that is roughly eighteen months old.
What I can say is that the counter-arguments are about where the value accrues, not about the mechanism. Nothing in this section disputes that the token is a poor unit of account, that prices are nondeterministic at submission, or that the credit is a conversion layer over dollars. Semafor, TechCrunch and PYMNTS are arguing about who captures the value created by that mess. They are not arguing that the mess is imaginary.
So the mechanical claims of Sections 1 and 3 survive this section intact. The commercial claim of Section 2 — that this was a payments acquisition in the meaningful sense rather than the metaphorical one — does not survive it cleanly, and I would not want a reader to finish this book believing it had been demonstrated. It is the reading I find most persuasive. It is not the only one the facts allow.
One small thing cuts the other way, and it deserves a mention because it is the only piece of post-deal evidence anyone has. OpenRouter’s own announcement commits to neutrality in unusually specific language: routing decisions will remain driven by one thing, which is what is best for the user.c-046 If routing were merely a funnel into a financial-services relationship, that commitment would be an odd thing to make and an expensive one to keep.
06
Prepaid credits, gifted compute, and the regulatory line that explains why you cannot buy a friend a hundred thousand tokens.
The practical question underneath all of this is whether you can sell someone a quantity of AI compute the way you can sell them a quantity of anything else. A gift card. A prepaid balance. A coffee bought for a stranger on the internet.
The answer turns out to be yes, with a conversion layer, inside a closed loop — and the three qualifications are doing all the work.
The unit everybody invents
By now the pattern should be familiar, but it is worth seeing how far it extends, because the conversion layer is not a Stripe peculiarity. It is what everyone builds when they need a stable number.
Stripe’s billing product, as established, lets a business define an AI Credit and configure how each model’s usage converts into it, then sell prepaid packs of that unit.c-018 The metering underneath tracks model and token type including cached tokens,c-019 and the product was not yet generally available at the time of writing.c-020
Anthropic has already shipped the same idea, and almost nobody has noticed. When Claude is sold through the AWS or Azure marketplaces, Anthropic rates the token usage in USD at standard per-model rates, applies any negotiated discount, and converts the result to Claude Consumption Units at one cent each.c-061 A CCU is worth a penny. It exists because a cloud marketplace needs a single countable line item and cannot express a charge such as 1.25x base input price for a five-minute cache write, so the vendor rates everything in dollars, converts to a synthetic unit at a fixed rate, and reports a quantity.
That is the whole thesis of this book, shipped as a production billing integration by a model vendor, for reasons that have nothing to do with any acquisition. Dollars are the unit of account. The synthetic unit is a presentation layer over dollars, invented because the underlying quantity would not sit still. And note the direction of the conversion — usage is rated in USD first, then converted. The dollars are not derived from the units. The units are derived from the dollars.
Anyone building a credit product is going to reinvent this, and should. The design question is not whether to have a conversion layer but who absorbs the risk in it. If you sell a credit today and the underlying provider reprices tomorrow, either your margin moves or your customer’s balance quietly buys less than it did. Someone eats that. Fixing the exchange rate is a promise about a price you do not control, and Section 3 is a long argument for why you should be nervous about making it.
The expiry problem
There is a second feature of every prepaid AI product currently shipping, and it is one customers dislike: the value goes away.
OpenRouter’s terms provide for expiry of unused credits after one year of purchase.c-036 Anthropic’s gift subscriptions expire 365 days after the purchase date.c-035 These are not outliers; they are the standard shape.
Expiry is where prepaid products stop being a pricing decision and start being a regulated one. Money taken now for a service delivered later is a liability on the seller’s balance sheet, and unredeemed value is revenue only under specific accounting conditions. Consumer protection rules in many jurisdictions constrain when prepaid value may expire and what fees may be charged against it.
I gathered no sources on the accounting or the consumer-protection treatment. This is a prompt to get advice, not advice.
Buying someone else compute
Which brings us to the consumer version, and the reason it is harder than it looks.
The patronage model is well established and boringly profitable. Buy Me a Coffee charges a 5% transaction fee, creators keep 95% of the earnings, and payouts go directly to a bank account.c-032 It is a thin layer over a payment processor, and the processor is Stripe: the platform states that payment processing is handled by Stripe and Wise.c-033
So the obvious product suggests itself. Instead of buying a writer a coffee, buy them the inference to write with. Both legs already exist and both already terminate at the same company — the patronage rail runs on Stripe, and Stripe supplies the token-metering primitive.c-055 There is no missing infrastructure.
And yet nobody sells it, while two adjacent things do ship. Anthropic will sell you a gift, but the gift is a plan: you choose a tier and a duration, and the recipient receives full access to the features of the selected plan.c-034 It is a subscription with a bow on it, not a balance. You cannot gift someone eighty thousand tokens and let them spend it where they like.
The reason is not technical, and I think it is not squeamishness about the product either. It is that a transferable balance spendable across unaffiliated vendors is a different kind of object in law from a subscription to one vendor’s service.
The line that actually matters
Here I have to correct a framing I nearly used, because it is the sort of error that survives in a book for years.
The natural place to reach for is FinCEN’s guidance on stored value, which held that the agency does not currently interpret the definition of stored value to include closed system products such as a mall-wide gift card program.c-037 That is a clean statement of the closed-versus-open distinction, and it is tempting to quote and move on.
It is also from 2003,c-056 the same document warns that FinCEN intends to engage in further rulemaking on the definition of stored value,c-057 and — decisively — that rulemaking happened. FinCEN’s 2011 final rule amended the Bank Secrecy Act regulations by renaming stored value as prepaid access, defining that term, and deleting the terms issuer and redeemer of stored value.c-070 The vocabulary the 2003 ruling operates in no longer exists in the regulations. The rule was framed as addressing regulatory gaps that had resulted from the proliferation of prepaid innovations over the preceding twelve yearsc-073 — precisely the category a cross-vendor compute credit would fall into.
The 2011 regime imposes suspicious activity reporting and recordkeeping requirements on both providers and sellers of prepaid access, plus registration for providers,c-071 while exempting certain lower-risk categories from certain requirementsc-072 — and it retains closed loop prepaid access as a defined category.c-074
The shape of the answer is therefore visible even from outside: closed loop is treated as lower risk and carries lighter obligations; open loop does not. A gift card for one vendor’s service sits comfortably on one side of that line. A transferable balance spendable across eighty inference providers sits on the other, or at minimum sits close enough to it that you would need a lawyer before you shipped.c-053
I have read a 2011 final rule and the definitions it amended. I have not researched the exemption thresholds in detail, state money-transmitter licensing, the CFPB’s prepaid account rules under Regulation E, the CARD Act’s gift-card provisions, or anything outside the United States. Several of those could be decisive.
What I can say is narrow: the reason you cannot buy your friend a hundred thousand tokens is not that the payments plumbing is missing. It is that the object you would be selling them is a regulated instrument, and the closed-loop products that ship today are shaped the way they are because that shape avoids the question.
What would change my mind
The argument of this book comes down to one claim: that tokens are a meter reading, that every credit is a conversion layer over dollars,c-049 and that the interesting position is therefore the exchange desk rather than the mint.
Here is what would falsify it, and what I would watch.
If a vendor began quoting a fixed price per token guaranteed at submission time — absorbing the routing variance itself, the way an insurer absorbs risk — the central mechanical claim would weaken considerably.c-048 The variance would still exist; it would simply have found someone willing to be paid to hold it, which is a real product and one I would expect to see before I expect to see a token currency.
If a transferable, cross-vendor balance shipped and survived, the argument in this section is wrong, and the constraint I have attributed to regulation was really just nobody having bothered.
If the cache multipliers converged across vendors — if 0.1x and 0.25x and 0.5x became one number the way transaction fees converged into a band — then a common unit becomes tractable, and the mess described in Section 3 turns out to have been an artifact of an immature market rather than a structural feature.
And the single most informative thing, the one that costs nothing to observe: whether OpenRouter’s commitment that routing decisions will remain driven by what is best for the user holds under its new ownership.c-046 If in three years the router quietly prefers providers that settle through its parent’s rails, then the exchange desk was the point, and it was worth more than the eight billion dollars nobody will confirm was paid for it.
If it does not, then Section 5 was right, and I will have written a book about the plumbing of a business that turned out to be about something else. That would be a fine outcome. The plumbing is real either way, and the bills still have to be reconciled by someone on a Tuesday afternoon, holding two numbers that do not match.
Anuj Sadani builds AI systems and the teams that build them. He spent a decade at NVIDIA, has worked inside AI-first organizations, and has led multicultural engineering pods across Europe — sixteen-plus years, most of them on the AI era's unglamorous half: turning hype into systems that ship, and systems that ship into outcomes that hold up in production. That work has fed into industry recognition from Gartner and Everest Group, and an Innovator of the Year nod.
He is the author of The Clean Vibe Coder: A Code of Conduct for Programmers in the Age of AI Agents and Borrow the Line. Own the Move., and writes about engineering, leadership, and what stays human when the tools get good. He still believes the best technology is the kind that makes the people around it braver.
anujsadani.in
tech.anujsadani.in