Twelve and a half billion dollars. That's what every company on earth spent on frontier model inference last year. And nobody can tell you who spent it.
The single most requested number in enterprise AI, and it arrives with a hole in the middle of it.
Daniel's been reading about the infrastructure side of this, the serving question, and now he wants the money side. His words: put some actual dollar figures on enterprise consumption of proprietary models. GPT, Claude, the frontier labs.
He wants the brackets. Startups at one end, Microsoft at the other.
He wants to know what a typical organization actually pays, how far that spreads, and which public data — if any — tells us how many organizations sit in each category.
And then the interesting half.
What generates the bill in the first place. Employee use, customer-facing products, coding agents, automated workflows. And one thing he's firm about: keep API inference separate from software subscriptions, and separate again from training.
Three buckets, and the headline numbers mix all three.
Which is why the headline numbers are useless.
The whole product line spans nine orders of magnitude. You can open an account this afternoon and pay four cents for a million tokens. You can also be a company with a nine-figure annual commitment to the same product family. Same pricing page.
Same pricing page, different planet.
So let's set the denominator first, because everything else hangs off it. Menlo Ventures does the annual enterprise survey, and their 2025 report puts total enterprise generative AI spend at thirty-seven billion dollars. That's up from eleven and a half billion in 2024. A little over three times.
Thirty-seven billion total.
Of that thirty-seven, eighteen went to infrastructure. And inside the eighteen, twelve and a half billion was foundation model APIs. That twelve and a half is the number. That's the closest thing anyone has to a hard figure for aggregate enterprise API inference spend.
The actual question Daniel asked, in other words. The market for the thing.
The other nineteen billion went to applications. Copilots, coding tools, vertical AI. That's largely seat-based subscription spend, and it has to stay separate or you double-count. And training infrastructure is a further four billion, also separate.
So Daniel's boundary lines are exactly right. Twelve and a half is the API line, nineteen is the software line, four is the build-it-yourself line.
Three different businesses that get reported as one number whenever a lab announces revenue.
And the twelve and a half is where the public data starts getting thin. Almost immediately.
Immediately. Neither lab publishes a histogram of customers by annual API spend. Not OpenAI, not Anthropic, not Google. The market size is published. The share is published. The distribution underneath it is not.
So everything after this point is inference from disclosed floors and reported contracts. Not a published distribution.
There is exactly one disclosed cut. Anthropic said in April that more than a thousand customers spend over a million dollars a year. And that number had doubled from five hundred in about two months.
A thousand customers above a million. That's a floor, not a ladder.
It's a floor. It tells you the top of the pyramid is real and it's growing fast. It tells you nothing about the shape below it. And Anthropic also reports three hundred thousand plus business customers as of September 2025.
Hold on. Three hundred thousand customers, and a thousand of them are above a million a year.
Do the arithmetic on the total. Twelve and a half billion across three hundred thousand business customers is a mean of roughly forty thousand dollars a year. But nobody publishes a median. And the mean is doing something dishonest there.
Because one customer at fifty million drags up a hundred small ones.
If a thousand customers are above a million, and some of those are well above, you can account for a very large share of the twelve and a half billion with a few hundred accounts. Which means the median is far below the mean. My guess is the median enterprise API spend is a few thousand dollars a year.
A few thousand, and it's the number that describes most customers.
Most customers by count. By dollars, the distribution is upside down. It's a barbell with a very thin handle.
So give me the brackets, then. Clearly labeled as estimates.
Estimates, all of them. At the bottom, self-service startups and small teams. Hundreds of dollars a year to low thousands. That's consistent with OpenAI's published tiers, where you can run a real workload for the price of a lunch. Then medium businesses and scale-ups, where you're past the free tier and into committed spend. Tens of thousands to low hundreds of thousands a year.
That's where most of the three hundred thousand sit, I'd guess.
Probably. Then smaller Fortune 500 companies and large enterprises. Low millions. They've got a procurement process, they've negotiated a rate card, they've got a committed spend contract. Then the enormous multinationals, which is tens of millions to hundreds of millions.
And the top of that bracket is Microsoft, per Daniel's prompt.
Microsoft is the interesting one, because the biggest number associated with Microsoft and AI isn't an inference bill at all. Executive testimony in May put Microsoft's cumulative spend on the OpenAI partnership at over a hundred billion dollars.
On the partnership.
Building capacity, funding the relationship, the whole arrangement. That is not a bill for tokens. It's the cost of owning the supplier.
That's a completely different line item wearing the same headline.
And it's the mistake the coverage makes constantly. Anthropic has a five hundred eighteen billion dollar decade-long compute commitment. Roughly eighty percent non-cancelable. Two hundred billion to Google Cloud over five years. Fifteen billion a year to SpaceX through May 2029. Those are compute purchases. Anthropic is buying capacity, not receiving an invoice for inference.
So the scariest numbers in circulation are Anthropic's shopping list.
Its shopping list. And there's a second confusion stacked on top. What the labs pay each other and what customers pay the labs are different flows entirely. OpenAI paid Microsoft seventeen point two billion dollars in 2025. Ten and a half of that was R&D and training, six billion was cost of revenue, which is inference, and about half a billion was sales and marketing.
OpenAI paying Microsoft.
Meanwhile Microsoft paid OpenAI three hundred and three million dollars. Fifty-six to one in OpenAI's direction. Neither of those is a customer's bill. That's two companies settling up for compute and for a partnership.
So we've got a market of twelve and a half billion, one disclosed floor at a thousand customers above a million, and no verified figure — anywhere — for what a single organization actually pays in a year for inference.
Not one. The closest thing is OpenAI's own cost of revenue, which was seven and a half billion in 2025. That's what OpenAI spent serving everything, not what any customer spent.
Every number that sounds like an answer is a lab's cost, not a customer's bill.
Which is the honest summary of part one.
So the brackets are guesses, the distribution is undocumented, and the biggest figures people quote are compute purchases. That's the state of the public record.
The state of the public record is a denominator and a floor. Everything between them is estimation.
Then the more useful question. What actually generates the bill?
Coding. Coding is the first real killer use case in the enterprise, and it's not close. Enterprise coding spend was four billion dollars in 2025. The year before it was five hundred and fifty million.
Eight times in a year.
And that four billion is fifty-five percent of all departmental AI spend. More than half of what enterprises buy, at the departmental level, is coding tools.
Why coding? Why not customer service, why not legal?
Because of what a coding agent actually does per task. It makes many model calls for a single job. It reads files, it reasons, it edits, it runs the tests, it fails, it retries. Often with large context windows. Compare that to a chat copilot, which is one call and one answer.
So the same seat generates a very different bill depending on what it's doing.
An order of magnitude different, probably more. A copilot seat is a fairly predictable monthly cost. A coding agent is metered, and its meter runs every time it has a thought.
Including the wrong ones.
Especially the wrong ones. A retry loop is a billing event. And that's why Anthropic's enterprise lead is a coding story. It holds fifty-four percent of coding against OpenAI's twenty-one. That's the whole reason it took forty percent of enterprise API spend in 2025, up from twenty-four the year before and twelve the year before that.
And OpenAI went from fifty percent in 2023 down to twenty-seven.
Google's at twenty-one. Those three are eighty-eight percent of enterprise API spend between them.
Cheap to be the fourth one.
Now the horizontal copilots, the general assistants. That's seven point two billion, and it's eighty-six percent of the horizontal AI category. Big, real, but seat-based. A predictable per-head number, and it scales with headcount, not with ambition.
That's a budget line a CFO can live with.
The metered one is not. And here's the part that surprised me. Only sixteen percent of enterprise deployments are true agents. Sixteen percent. The other eighty-four percent are fixed-sequence workflows wrapped around a single model call.
So the agentic fleet everyone's writing about is a rounding error on the bill.
Today's bills are seat-based copilots and coding assistants. The autonomous multi-step thing that runs unattended — that's a small slice. It's growing, but it is not what's generating the twelve and a half billion.
Which Daniel asked about directly. Automated workflows.
Automated workflows are mostly fixed-sequence. Do this, then this, then stop. That's cheap per run. The expensive shape is the one that decides for itself how many times to try.
So where does that leave employee use versus customer-facing products?
Employee-facing tools are the bulk. Copilot seats and coding assistants, pointed at your own staff. Customer-facing AI products are growing and they're real, but they're a smaller share of the twelve and a half. And the reason is that a customer-facing product has a margin problem — every call a customer makes costs you money directly, so you're motivated to keep the model small and the calls few.
The internal tool has no such discipline, because the value is harder to measure.
There's no revenue line attached to a coding agent. There's just a bill.
So now the question underneath Daniel's question. Why do the bills climb when the price per token keeps falling?
This is the structural part, and it's the best thing in the whole picture. OpenAI's gross margin fell from about forty percent to thirty-three percent, while revenue tripled.
Revenue triples, margin falls.
Because the mix shifted toward reasoning models, and a reasoning model burns far more compute per query. It generates a chain of thought before it answers. Some of these models will spend thousands of tokens thinking about a question the user asked in twenty.
And the customer sees one answer.
And pays for the thinking. So you get a compound effect. Price per token falls, but tokens per task rises faster. Coding agents make many calls, reasoning models burn more per call, context windows keep growing because everybody wants the model to see the whole file. The meter you're reading got cheaper. The number of meters running went up.
And that's why the twelve and a half billion climbs even though the pricing page looks generous.
Every efficiency gain gets reinvested in a harder task. That's the pattern.
What about the billing path? Who actually invoices the customer?
Forty-seven percent of Anthropic's 2025 revenue was billed by Amazon and Google. Up from thirty-two percent in 2024. So when an enterprise says it's spending two million a year on Claude, close to half the time that's an AWS or a Google Cloud line item.
On an Amazon invoice.
On a cloud invoice, and the cloud owns the relationship. It knows the customer, it holds the contract, and it will happily sell you Gemini or a Llama endpoint off the same shelf.
So the labs are partly renting their customers.
And the customer gets a single cloud bill with the model line buried in it. Which is another reason no one can produce a clean figure for what a company spends on inference. It's inside a cloud bill that nobody breaks out.
The last piece. How safe is that revenue?
Anthropic's draft prospectus says two customers each accounted for twelve percent of 2025 revenue. And it notes that many large customers lack long-term commitments.
So the biggest bills are re-won every month.
Consumption pricing means the customer decides each month how much to consume. There's no three-year lock-in on the usage side. You can have a nine-figure account and lose it in a quarter if the team moves to a different model.
That's a strange business. Enormous revenue, enormous concentration, renewed monthly.
The Decoding Discontinuity teardown put it well. Their revenue is rent. Mostly metered and re-won month by month. Their costs are leases.
Rent in, leases out.
And the leases are the five hundred eighteen billion. That's the asymmetry. The upside is meters running in a customer's cloud account. The downside is a decade of non-cancelable commitments.
There's a detail in that teardown I keep coming back to. The revenue is concentrated in the customers who can leave fastest. The two accounts at twelve percent each — those are the ones with the leverage to renegotiate every quarter. So the stability of the whole business rests on the least stable part of it.
Which is why the labs push so hard on multi-year commits at the top. The smaller accounts are sticky by inertia. The large ones are sticky only by contract, and the contracts are short.
So the concentration risk and the churn risk are the same risk.
They always were. It just shows up in two different lines of the filing.
Raj, you've been sitting there the whole time. What have you got?
Gentlemen, this is not a market. This is a screenplay. Three hours, two intervals, one villain — the meter. I am seeing Herman as the numbers man, very strong, very strong, and Corn as the conscience, the one who asks the question nobody wants to answer. We open on the pricing page. Four cents a million tokens. Beautiful. Then smash cut to the boardroom, and the bill is eleven feet long. Interval twist — the cloud was the customer all along.
Raj, when you say the cloud is the customer, do you mean the billing relationship, or ownership of the account?
Both, neither, it doesn't matter, we fix it in the edit. And here is the part I love. We put a song in the retry loop. The agent tries again, and again, and each time the orchestra gets bigger. By the ninth retry the whole Mumbai Philharmonic is on its feet.
So the bill is the score.
My friend. And the last shot, on the CFO's face, as the monthly invoice arrives. No words. Just the invoice, and one tear, and the music swells. I have the option papers in my bag. Excuse me, another booking.
The orchestra gets bigger every time it fails. That's roughly the coding-agent economics, honestly.
So the picture is this. A thousand customers above a million, three hundred thousand customers total, and a median that's probably a few thousand dollars a year.
And the twelve and a half billion is carried by a small number of enormous bills. The tail isn't the tail. The tail is the market.
One more thing worth having for Daniel's list. The two labs' headline numbers aren't comparable.
The fifty billion versus seventy billion gap that was in the news last week. That's a basis difference, not a restatement. OpenAI counts what it keeps, net. Anthropic grosses up the sales billed through the clouds.
So one company's number includes money that goes straight to Amazon.
And the other's excludes money that goes to Microsoft. You cannot put those two figures side by side and declare a winner. It's an accounting convention wearing a scoreboard.
Every "who's winning enterprise AI" chart is built on that.
Every one I've seen.
Herman, you told me you'd read something overnight that you weren't going to bring up.
I did, and I'm going to bring it up. Sacra's estimate has OpenAI's inference costs going from eight point four billion this year to fourteen point one billion next year.
So the cost line grows faster than the customer count.
Substantially. And that's before the agentic share moves off sixteen percent.
Which it will.
How much does a token cost when nobody's watching it?
That's the whole problem in one line.
Hilbert, I think you've solved the episode.
I didn't ask it to solve anything. I asked it because I've been on the phone with my nephew since you started talking, and he's got a number for me, and the number is not what he said.
He works in procurement, then.
He works in procurement at a hospital group. Twelve hospitals, all regional, all paper files. Four years ago they set up their own department for the AI line items. Two people. He's one of them. His job is to go through the bills and find out what's actually on them.
What did he find?
Their annual invoice from the model provider was sixty-eight thousand dollars and change. Sixty-eight thousand four hundred and something. He read it twice because he thought the decimal was in the wrong place. It was his first month on the job. And the number that mattered wasn't the sixty-eight thousand. It was what sat next to it on the same statement, from the same cloud, for storing the files the model reads. That was two hundred and eleven thousand.
The storage cost more than the inference.
The file storage was two hundred and eleven thousand and the model was sixty-eight. So he goes to his director and says, we're paying three times as much to keep the files as we are to think about them. And the director says, well, that's the storage line, that's a different budget.
Different budget. Same invoice.
Different budget. His department can only see the sixty-eight. The storage sits under IT infrastructure in a code he doesn't have access to. Two years he's been asking for the code. Every quarter the request gets closed as resolved with no note. He now knows the number by heart and he cannot get a copy of the line. His manager told him to stop raising it in meetings because it makes the AI programme look bad.
The storage is doing the honest accounting his department won't.
He asked me to guess the total for all twelve hospitals and I gave him a figure and he laughed and said it was eleven percent high, which he knew without looking it up.
Which side of it is inference, from his point of view?
None of it. In his filing it's one line item called "cloud services" and it's been the same code since before the AI programme existed. That's why he can't find the model spend and the storage spend separately. They were never separated. He's reconstructing it from purchase orders.
The four years, that's the whole time the department's existed.
Four years, two people, and a code he can't get.
The sixty-eight thousand number is what he pays for the model...
That's one hospital group's number and it's already buried in a code nobody can see.
The only clean figure any of us has heard this whole episode came from a man reconstructing it from purchase orders.
It's smaller than the storage sitting next to it. Which changes the shape of Daniel's question a little. If the storage line routinely outruns the model line, then the model spend isn't where the argument is. The argument is that it's invisible.
Hilbert, the director — did he ever look at the sixty-eight thousand?
He called it encouraging. Said it showed the programme was disciplined. He's not wrong. It's the storage he can't see.
Right. Because the inference metric everybody's building charts about is the smallest number on that statement.
The smallest number, and the only one anyone asked about.
He sent me the invoice as a photo. Four pages. He's keeping it in a drawer.
A drawer somewhere in a hospital group, and that's the state of enterprise inference accounting.
That's the most honest figure in the episode. Sixty-eight thousand, and it's smaller than the thing next to it that nobody counts as AI.
Which puts the twelve and a half billion in a new light. The aggregate is real, the bills are real, but a huge amount of enterprise model spend is sitting in a code called cloud services, next to storage, and the only person who's looked at it closely had to fight for the access.
When the next earnings call cites enterprise momentum, that number includes a lot of line items that were never labelled.
Last thing before we close, because Daniel asked for the most credible picture, and this is it. There's a gap between the two labs' headline revenue figures that isn't a gap.
Fifty against seventy. One of those numbers is net of what the cloud takes. The other grosses it up. Put them side by side and it looks like a race. It's two different rulers.
The "who's winning" question is answerable and unanswerable at the same time.
Answerable on share, where they're measured the same way. Unanswerable on the headline, where they're not. And the headline is the number everyone quotes.
Which is why Daniel's instinct to separate the buckets was right. Subscriptions, training, inference. Separate them and the picture gets smaller and truer. Add them together and you get a number you can't defend.
Small and true beats big and blurry.
Where does the bracket structure go from here?
Up, if the per-query compute keeps rising and the agentic share moves off sixteen percent. The bottom of the distribution probably stays where it is. The top gets further away. Which is the shape of every enterprise software category, honestly.
The bottom stays a few thousand dollars a year and the top stays a small number of enormous accounts. The brackets don't move. The gap inside them widens.
The median is a few thousand dollars and mostly unpublished, which is the least satisfying answer we've ever given anyone.
It's the honest one. The most credible picture is that the twelve and a half billion is carried by a handful of very large bills, that the median enterprise spends almost nothing by comparison, and that the mega-contracts in the headlines are compute purchases, not inference revenue.
Which is what Daniel was really asking. What does an organization actually pay, and the answer is it depends on which organization, by four or five orders of magnitude, and nobody publishes the ladder.
The storage next to the model probably costs more.
In at least one hospital group it does.
Thanks to Hilbert Flumingtop, our producer, for his nephew and his drawer.
For the sixty-eight thousand.
For more along these lines, there's episode forty-eight, Renting vs. Building; episode three, Safetensors or something else; and episode ten, How ASR Went From Frustration To ... Whisper Magic. This has been My Weird Prompts, the human-AI collaboration podcast. If you want to put a figure on your own inference bill, or you've spotted one we can't, send us your own prompt on Telegram at t dot me slash MWP listener bot. We read everything.
We do. And if you enjoyed this one, a review helps more than you'd think.
See you soon.
