Let’s talk about how a company can spend millions on AI without a single meeting where anyone approved AI.
Not a shadow IT story. Nothing stolen, nothing hidden. The money was approved. It was approved eighteen months earlier, in a cloud contract, for something else entirely.
Picture the renewal meeting. Two people from finance, one architect, a procurement lead, and a slideshow with a number on it: committed spend. Four million over three years, and if you don’t draw it down you lose the discount, and in most of these contracts you lose the remainder.
Now fast forward a year. Somebody in the business wants to buy AI. And the reason the deal closes in a week instead of a quarter is not that the model is good. It’s that there’s unused commitment sitting on the table with an expiry date.
That is the engine nobody is describing when they talk about enterprise AI adoption.
The numbers came out this week and they’re worth sitting with. Bloomberg reported on 26 October that Anthropic’s IPO filing puts cloud marketplace sales at 48% of its revenue in the first half of 2026 — Claude bought through Google, Amazon and Microsoft cloud accounts. In 2023 that channel was 11%.
Same filing, the other side of the ledger: at least $114.5 billion committed to Google through 2033, $110 billion to Amazon through 2036, $44.9 billion to Microsoft. Call it $269 billion of minimum compute spend owed to the same three companies that carry half the revenue.
And the expense line tells you what the channel costs. Sales, marketing and partner-related spend went from $120 million in the first half of last year to $3.2 billion in the first half of this one. Partner payments alone up $2.7 billion.
So: a company paying a fee to sell through partners, and committing hundreds of billions to buy from them. Money leaving twice. That’s not a scandal, it’s just a structure — and structures shape behaviour whether or not anyone intended them to.
Here’s the behaviour it shapes on your side of the table.
You don’t buy AI. You draw it down.
There’s no new budget request, so there’s no line item, so there’s no owner. There’s no competitive tender, because the platform is already an approved supplier and the security review was done two years ago for object storage. There’s no unit cost, because the invoice has one number on it and it’s not the model’s.
Which is exactly why it’s pleasant to buy and impossible to price.
Now look at what the buyers actually know. CloudZero published its October pulse on a same-store panel of 422 organisations, and the median company’s AI spend hit 3.02% of its cloud bill in September — fourteenth straight monthly increase, the biggest jump on record, about four and a half times where it was a year ago. The seventy-fifth percentile is 12.55%. Nearly 30% of those companies have AI above 10% of their cloud spend, and 8.6% are above a quarter.
Then the same report asks 252 finance leaders to estimate their AI share of infrastructure spend. Median answer: 21 to 35%. Fifty-four percent said above 20%.
Billed median, 3%. Perceived median, somewhere north of 20%.
Some of that gap is definitional — bundled AI inside SaaS doesn’t show on a cloud bill. Fine. But the survey found that leaders who can only see a total bill were just as confident in the 20%-plus number as leaders with team-level allocation. Confidence without visibility. That’s the tell.
And 72% of the people who think they’re spending above 35% say they spend more than they can justify. They’re also the most committed: 59% plan to increase more than 20% next year. Of the teams that can’t justify what they’re spending, 93% are restructuring the budget process.
Read that again. The people who can’t explain the bill are the ones voting to make it bigger.
That is not stupidity. That’s a channel working exactly as designed.
Because a marketplace purchase doesn’t feel like spend. It feels like using something you already paid for. That’s the oldest trick in subscription economics and it works on CFOs, not just consumers.
Now the pattern, because this is the part worth writing down.
Every platform era ends the same way: the billing relationship becomes the distribution. App stores didn’t win because the apps were better. They won because they held the credit card. The marketplace didn’t beat the vendor’s website on user experience. It beat it because the buyer had already signed the master agreement and the purchase order was a formality.
Cloud marketplaces are that same machine, one layer up the stack. And AI walked straight into it, because AI is the first product in twenty years that every department wants and no department has a budget for.
So it goes on the cloud bill. Where it becomes invisible, unallocated, and — this is the important part — very hard to leave.
Think about what leaving involves. Your contract isn’t with the lab. It’s with the reseller. Your committed spend is a three-year commitment to the landlord. Your security review cleared a platform, not a model. Your data sits where the platform put it. You can swap the model in an afternoon. You cannot swap the bill in an afternoon, and the bill is what the renewal conversation is actually about.
There’s a second incentive stacked on top, and it’s the one nobody says out loud. Those three marketplaces sell their own models. They are competitors to the thing they’re distributing. That’s not illegal or unusual — every retailer has private label — but it means the shelf is not neutral, and the ranking you see was decided by somebody with a P&L.
And a third: most of this is consumption-priced. Per token, per action, per conversation. Which means the vendor earns more when the machine says more, and the marketplace takes a percentage of the saying. Nothing sinister in the arithmetic. Just don’t pretend the incentive points at brevity.
So what does an operator do. Five things, none of them clever.
Split the line. AI gets its own cost centre and its own owner, even if it arrives on one invoice. If you can’t split it, you can’t manage it, and the 3% versus 21% gap is what that looks like from the inside.
Price it per finished job. Not per seat, not per token. What did this cost per claim processed, per ticket closed, per quote drafted. One number per outcome. That’s the only figure that survives a budget review and the only one that lets you compare a model to a person or to nothing.
Contract with the lab where you can, and if you go through the marketplace for the discount, know exactly what that discount costs you in flexibility. Get the exit terms in writing before the pilot, not after.
Run a second provider in production. Not a bake-off. Production, on real work, small volume. The point isn’t to switch. The point is that the day you need to switch, you already know how, and the other side of the table knows you know.
And keep your own eval set. Your ten ugliest real cases, with the right answers written down by a human who does the job. Then when the model version changes under you — and it will, on a date you didn’t pick — you can tell in an afternoon whether your work got better or worse.
That last one is the whole thing, really.
Because the reason the cloud bill works so well as a channel is that it removes the moment where somebody has to decide whether this is worth it. No approval meeting means no argument about value. No line item means no comparison. No comparison means the only question left is how much, and the answer arrives after the fact, in a number nobody can decompose.
You end up renting judgement — a machine that decides things inside your business — and being unable to say what that rental costs, or what it’s worth, or who agreed to it.
That’s the sovereignty problem in its most boring form. Not deepfakes. Not robot overlords. A line item that got absorbed into a bigger line item, and a decision that quietly moved to whoever holds the contract.
The principle worth keeping is simple enough to put on a wall: you should be able to name the price of the thinking you rent, and the person who approved it should be the person who understands it.
Everything else — the models, the benchmarks, the demos — is a detail that changes every six months. That one doesn’t.
You’ve got this. Get the number out of the cloud bill first.