Skip to content
Threadbaire
← Blog

Booked, Paid, Used - The Missing Number in AI's Demand Claims

Lida Liberopoulou ·18 November 2026 · CC BY-SA 4.0

Independent research on AI economics and infrastructure · Commissioned research

Since Jensen Huang resigned as Nvidia's chief executive last week, the coverage has followed the scandal. But it has mostly not followed the money. For two years, Huang was the industry's main witness for AI demand. Hundreds of billions of dollars of loans have been made on the strength of that demand, partly with savers' money. This piece asks what the demand claims actually count, and finds that the number needed to test them, how much of the capacity is actually being used, is the one nobody publishes.

To go straight to what's missing, skip to the missing number.


The Nvidia scandal currently dominating the news is full of colourful stories. We have heard about the affairs, an alleged spy, and the chips that went missing in Southeast Asia. All of them raise fair questions about trust. But there is a bigger trust question underneath them, and it has less to do with the people involved than with the numbers, and the money attached to them.

Jensen Huang didn't just run Nvidia. He vouched for the whole industry's demand: in keynotes, in interviews, on earnings calls and in deals, some of them agreed on his word before a line was written down. Elon Musk wrote on 14 November that when Huang promised him computing power, “we’d put together a purchase order at some point later.” And money for these deals comes increasingly from lenders: banks, private credit funds, and behind them insurers and pension funds. We have heard plenty of big promises about AI's future, and all of that is open to debate. But these loans are not only bets on the future. They are being made now, partly on the strength of claims about how well AI is doing today.

The man who vouched for the demand is gone. But the question he never had to answer is still here. When the industry says "demand", what is it counting? And how much of what has been sold is actually being used?

What do they mean by demand?

First, the scale. The biggest cloud companies are expected to spend nearly $800 billion on data centres and equipment this year and $1.3 trillion next year, by Nvidia's own figures. Nvidia sold $108 billion of data-centre products in the quarter to October, and Huang says a single gigawatt of AI data centre costs around $50 billion to build. Microsoft alone has signed $365 billion of data-centre leases that haven't started yet. Amazon's long-term borrowing has more than doubled this year, to $173 billion. Anthropic has committed to about $900 billion of computing. And Nvidia has offered up to $105 billion of credit support for one data-centre campus in Ohio, and helped set up a platform to raise $500 billion more from Wall Street, which, Bloomberg reported last week, has yet to produce a deal.

By the industry's own account, what justifies all this spending is demand. "With demand, everything will be resolvable," Huang said in October.

But demand, as the industry uses the word, isn't usage. When Huang or Amazon's chief executive Andy Jassy talk about demand, they may well have usage in mind. But the numbers they give for it count wholesale demand: AI labs booking capacity from cloud companies, or cloud companies ordering chips. In short, they count contracts between companies in the middle of the chain. When they say demand, it sounds like usage. What they count is something else.

The distinction matters here because the cash AI actually produces comes only from usage. Usage is the chips switched on and running an AI model because someone is actually using it to write code, answer customers or analyse documents. It is the work the hardware does when a task is given to an AI.

Normally, when a business uses AI for that kind of work, it pays the AI provider for it, either through a subscription with set limits or per request, by the token. The AI provider uses that money to pay the cloud company for the computing it needs to do the work. The cloud company uses it to repay the loans that bought the chips, and to cover the rest of its expenses. Every step depends on the first one: someone, somewhere, using AI and paying for what it does.

That's what I'll call paid usage: work done by the chips and paid for by a customer out of its own money.

How it's supposed to work

In a normal market based on paid usage, the money flows one way, and it should look something like this:

customers pay for AI → AI providers → cloud companies → chipmakers

The whole system would rest on actual money paid by people and companies using AI. Everyone in the middle would depend on the customers at the end continuing to pay for a service they use. AI providers and cloud companies could still take out loans, like any other business. And as with any other business, the lender would look at the paying customers, and decide from those numbers whether the business can repay. Whatever the arrangement, what the chips are worth would be decided by how much the people using the AI are willing to pay for what it does.

How it actually works

Today, what we can actually see is a lot of the money entering in the middle.

  • Investors fund the AI companies: global venture funding for AI passed $400 billion in the first half of 2026, and Nvidia's finance chief said roughly 70% of it is spent on computing.
  • The cloud companies invest in the AI labs that rent from them: Amazon, Google and Microsoft in Anthropic, Microsoft and Amazon in OpenAI. In the last quarter alone, Amazon put $21.3 billion into OpenAI.
  • And Nvidia invests in the companies that buy its chips (nearly $50 billion in the AI labs alone, by its own account), promises to rent their capacity if they can't sell it, and stands behind their loans.

customers' money + investors' money + suppliers' money + borrowed money → AI providers → cloud companies → chipmakers

This isn't unusual. It's how booms get financed, and it lets companies grow fast enough to keep up with a growing market. But it also means payments up the chain can be made before paid usage arrives, and every number in the middle (bookings, contracts, chip sales) can keep growing without telling you whether paid usage is growing with it.

And that is the main problem. The spending figures are precise to the billion, in filings and on earnings calls. But listen to how the same companies describe whether the chips are being used.

"The demand is just extraordinary," Huang said in an interview on 21 October. "2027 is largely sold out," said Andy Jassy, Amazon's chief executive, two days later. And on Nvidia's earnings call in August, its finance chief Colette Kress said: "NVIDIA compute is fully utilized across every cloud we serve."

They give the money in figures while giving the usage in adjectives. To be fair, Jassy is talking about bookings for next year. But Huang and Kress are describing today: not that the infrastructure will struggle to keep up with AI use in the future, but that it can't keep up right now.

And this is where the real difference lies. A forecast is an opinion: anyone can be wrong about next year. A description of today is a claim of fact, and claims of fact can be checked.

So let's see what these "extraordinary demand" claims are actually based on.

What does "sold out" actually count?

When Jassy says “sold out,” he means Amazon’s computing capacity: large blocks of chips or data-centre power that customers reserve, often years ahead. When Kress says "fully utilized," she means Nvidia's chips inside the clouds that buy them. Words like these run through almost every earnings call in this industry: "sold out," "fully utilized," "supply constrained," demand that "exceeds available supply." They sound like they all mean the same thing: the chips are busy. But there are three different things a company can be counting when it uses them.

Think of a gym that sells 1,000 annual memberships for a room with 200 machines. It has booked 1,000 members. It has been paid for by 1,000 members. How many are using the machines at 7pm on a Tuesday is a different number. And that is the number the gym never talks about.

In the same way, the AI industry's demand numbers count the first two. Some companies also say how much capacity is switched on, like the number of machines plugged in. None counts the third.

Booked means a customer has reserved computing: a number of chips, or a share of a data centre's power, for a set period, often starting months or years later, sometimes in buildings that aren't finished yet. The only thing this tells you is that someone has reserved that capacity for a period of time. Nothing has necessarily been paid, and nothing has been used.

Part of what is booked is sometimes reserved by the chip supplier, not by an AI provider. In its latest quarterly report, Nvidia says its cloud partners "have strong customer demand and robust sales pipelines." But in the same paragraph, it explains that it has promised to rent their capacity itself if they can't sell it: $36 billion of promises, typically lasting six years. To be fair to Nvidia, the cloud companies can stop renting to it and sell to other customers "at more advantageous rates," so Nvidia is a safety net, not an overpaying customer. But when a cloud company's capacity counts as "sold," the buyer of that capacity might be Nvidia.

Paid for. OK, so if money is actually being paid, the chips must be in use, right? No. CoreWeave's latest quarterly report says 98% of its revenue comes from contracts where the customer pays whether it uses the capacity or not. That is real revenue. But it only tells you a customer is committed to pay for the capacity it reserved. It doesn't tell you the chips are working.

Used means the chips are actually switched on and doing work for someone. It's the number that matters, but none of these companies publishes it. The closest anyone comes is Kress's "fully utilized," which has no definition and no figure attached. Huang described what happens when a customer isn’t ready to use the chips it ordered: “we’ll ship the computers to somebody who can use it.” So Nvidia clearly knows that booked isn’t the same as ready. But “can use it” here means ready to plug in: a finished building, power, money. Whether the chips then do any work is a different question, and Nvidia doesn’t give a number for that either. It may not have one. Nvidia sells the chips, but the cloud companies run them, so a real usage figure would have to come from them.

And there is the question of who vouches for the demand. Asked about it in October, Huang pointed to his customers: Elon Musk's SpaceX ("demand is many times the capacity he has"), CoreWeave, Nebius and Nscale. Nvidia has invested in all four of these companies, and sells them the chips. Its latest share-holdings filing, published on Monday, 16 November, values its stakes in the first three at about $24 billion at the end of September. Nvidia has also invested in Nscale, a private company, in two funding rounds; the size of that stake isn't disclosed.

None of this means there is no demand. But it means the main evidence for it comes from companies whose value depends on people believing the demand exists. That's why usage, the number that doesn't depend on anyone's word, is what really matters here.

That number isn't published. But there's another way to test a claim like "demand is extraordinary." If you can't count the people using the gym, you can still look at the car park. If demand today really outran supply many times over, it would leave traces in the numbers that are published.

If demand were that big, what would we see?

There would be few idle chips. Lenders would be lining up to fund these companies. And paid use would be growing in proportion to the spending. None of these proves anything on its own, but together they should point one way.

So here is what we actually see.

Idle chips. An idle chip is one that has been made, sold and paid for, but isn't doing any work. It might be sitting in a warehouse. It might be installed but not yet switched on. Or it might be switched on with nothing to run. In every case, someone has spent the money and nobody is earning from it yet.

If demand today really outran supply many times over, you might expect idle chips to be rare, with every chip put to work the moment it arrived. So how many are idle?

In the same interview where he called demand extraordinary, Huang said that between shipping and switching on, a data centre can take four to six months, and, talking about AI spending on the scale of a trillion dollars, described "a few hundred billion dollars sitting idle while you're waiting for revenues." Alan Waxman, chief executive of the investment firm Sixth Street, said that by his firm's numbers about $200 billion of chips are sitting in warehouses because the data centres aren't ready. At Alphabet, buildings and equipment not yet in use rose by 85% in nine months, to $145.3 billion, while what is in use grew by 28%. And last month Microsoft said it would make up for delayed capacity by "finding unused capacity" in hardware it already runs.

To be fair, a chip waiting to be powered isn't a chip nobody wants. There can be a shortage of switched-on capacity even when the chips exist: what keeps them from being used may be a lack of money, power, memory to run alongside them, or buildings to house them. So idle chips don't disprove the demand. What they do show is how much has been spent ahead of any revenue, and that carries a cost: idle chips earn nothing, lose value every month, and someone is paying interest on them.

Lenders. The specialist cloud companies, the ones that do little but rent out chips, have huge costs to cover, mainly the chips they buy. The costs are so large that even funding rounds in the billions aren't enough, so they need someone to lend them money. And this is the test that matters most. Lenders don't count on promises of a bright future. They get their money back with interest and nothing more, so they look hard at who will pay. If customers were queuing to rent these chips, lenders would be queuing to fund them.

Lenders have lent hundreds of billions of dollars for AI this year, but much of it to giants like Microsoft and Google, or to deals they stand behind. For the specialists, lenders aren't queuing. Instead, in many of the biggest deals, it is Nvidia that has to stand behind the cloud company before lenders will sign. Nvidia explained why on its August call. Cloud companies, Kress said, would normally need "a single long-term offtake guarantee that lenders typically require": one big customer who promises to pay for years. Instead, Nvidia provides "a take-or-pay commitment on a portion of the facility's capacity, a minimum revenue guarantee that gives lenders the confidence to underwrite the project." Then she added: "We're not making loans." That's true: Nvidia isn't lending the money. So in plain English, lenders want a buyer who has promised to pay, and to get the deal done, Nvidia, the supplier, becomes that buyer.

Kress also said that the AI labs "for which we expect to leverage our balance sheet" would account for roughly a quarter of Nvidia's business next year. And Huang, speaking to CNBC in Berlin in October, put it most simply: demand for the neoclouds is "off the charts," but "what is missing in the middle is the working capital to fuel that growth." Lenders' own rule, as Kress described it, is that a customer signed up to pay for years makes a project easy to fund. If the demand came with enough of those customers, money would be easier to find.

The form of the support says something too. Nvidia isn't offering cheaper loans or better collateral. It is offering to be the customer. That suggests lenders' doubt was at least partly about demand: who will pay, and for how long.

But there may be another explanation. Maybe lenders don't doubt the existence of paying customers but the chips themselves, because they lose value too fast. Huang says Nvidia's chips stay useful for "probably eight to maybe more" years. If he's right, the chips keep earning for years, and the lenders' hesitation is about something else, such as demand or timing (though a chip that still works isn't necessarily one that resells for much). If chips stop earning their keep before six years, the companies spreading their cost over six years are reporting better profits than they should. Amazon shortened part of its server lives from six years to five in 2025, citing how fast AI is moving. Lenders may well doubt both, and that is the worst case for the industry's story: demand that hasn't been shown, and profits that look better than they are. Either way, the industry's own numbers don't show that demand today far outruns supply. I looked at what a chip's useful life depends on in Compute Watch No. 1.

Paid use in proportion to spending. People pay real money for AI, and a lot of it. Anthropic took in $33.2 billion of revenue in the first nine months of the year, and spent $18.8 billion on computing and infrastructure. It says its annualised revenue reached about $81 billion at the end of October. Its computing commitments now come to about $901 billion, of which about $452 billion is expected to be spent before the end of 2030. About $186 billion of the total is subject to termination rights or to conditions that must be met first, such as the other side securing financing. Even leaving those aside, its revenue has to keep growing to cover the rest. But Anthropic is a single company, and we can only see its numbers because it is going public. The question is whether paid use keeps pace with the spending across the whole chain, and that can't be checked. As we saw earlier, a large part of what is spent on computing is investors' money, not customers'. And none of the biggest cloud companies reports AI revenue as a figure of its own. Microsoft has even stopped reporting how many people pay for Copilot.

Even Sam Altman, whose company is among the biggest buyers, allows for the gap. Asked on Monday whether the debt behind the build-out is a problem, he said no, pointing to how fast revenue is growing. But he raised the question himself: "Is it possible that the buildout can get ahead of how quickly the economy can digest it? Yes, that could happen. I don't see that level of recklessness happening right now." That is a judgement, not a number.

One company, its own numbers

If you wanted to test all this on one company, CoreWeave is the obvious choice. Renting out Nvidia chips is its whole business, so unlike Amazon or Microsoft, its numbers aren't mixed in with shopping, software or advertising. Its chief executive, Michael Intrator, told investors last week that CoreWeave has "10x as much qualified demand as available capacity through the end of 2027." He didn't say what "qualified" means. And Nvidia owns about $4 billion of its shares. If any company could show how much of its capacity is used, it would be this one.

It has contracts for 4.6 gigawatts of power, and about 1.7 gigawatts of that is switched on. Two customers made up 54% of its revenue last quarter. But it does not disclose how much of that switched-on capacity is being used. On its latest call, the only mention of utilisation came from an analyst's question, and the finance chief's answer gave no figure. As in every other case here, no number was given.

Here is my best guess at what that missing number might look like, and I label it as a guess. Think of a hotel with 100 rooms open and a posted rate of €200 a night. If every room were let at that rate, it would take in €20,000 a night. If it takes in €10,000, either half the rooms are empty, or they're going cheaper than the posted rate.

CoreWeave had between 1.5 and 1.7 gigawatts switched on last quarter. Its finance chief says pricing for new long contracts is going up to $20 to $25 million per megawatt a year, and short contracts pay more. If just 1.5 gigawatts were rented at the long-contract rate, it would bring in $30 to $37 billion a year. Its actual revenue is running at about $14.6 billion, less than half of that.

There are innocent explanations for this: new capacity starts billing late, older contracts are cheaper, and CoreWeave may count "switched on" differently from how I do. And even this guess measures money, not use: most of CoreWeave's revenue is paid whether the chips work or not. The published numbers can't tell you which explanation is right. That's the closest an outsider can get, and it's exactly the kind of question a usage figure would answer.

On Nvidia's August call, Huang said the computing hours these companies rent out are "insanely profitable," and that he had "heard the other day" that they earn back their cost in less than a year. He was mostly talking about the big cloud companies, which don't report the profits of their AI business separately. Where those profits can be checked, at a company that does nothing but rent out chips, the picture is different. Last quarter CoreWeave made $207 million from running its business and paid $846 million in interest, for a net loss of $709 million. It expects its interest bill to pass $1 billion this quarter. The one company in this chain whose AI profits can be clearly measured is the one selling the chips: Nvidia keeps about 75 cents of every dollar of sales after the cost of making what it sells, its gross margin.

One line in CoreWeave's report deserves an answer. During the quarter, $1.9 billion of prepaid contracts "became cancellable." The report doesn't say which, or why.

What number is missing?

The industry has a fair answer to much of this. Capacity is booked years ahead because it takes years to build. Long contracts are what make it possible to finance. Some chips have to wait because the buildings and power aren't ready, and a customer can have good reasons to reserve more than it uses every hour. And the rent on AI chips, which I track each month in the Compute Watch, is still rising, which is consistent with a real shortage, at least in the part of the market the indexes cover. All of that can be true. But none of it tells you how much of the switched-on capacity is doing work, how much that work earns, or how much of the money comes from outside the chain.

So, again, everything in this piece comes back to one missing number: how much of the switched-on capacity is actually being used. And there is a simple way to measure it.

Chips draw far more electricity when they work than when they sit idle. So the electricity meter is a rough usage meter: not a perfect one, since a data centre also powers cooling and networking, but one that already exists. Some companies already tell you how many megawatts they have switched on: CoreWeave says about 1.7 gigawatts, for example. But that is the size of the capacity, like the number of rooms a hotel has open. What they don't tell you is how many megawatt-hours were actually used: how much electricity the chips drew, over the quarter, doing work. That's closer to the number of nights guests actually slept in the rooms. Every data centre already has this number. It's on the electricity bill.

Other industries publish theirs. Power plants report a capacity factor: how much electricity they produced against how much they could have. Airlines report how full their planes are. Hotels report occupancy. In telecoms, where I used to work, an operator that advertises 99% 4G coverage has to say what coverage means, the regulator tests it, and the advertising watchdog can make it withdraw the claim. We are at the point where a €20 mobile plan gets more scrutiny of its claims than an industry borrowing hundreds of billions of dollars.

Why is the number missing? Companies have to disclose what they owe and what they have promised, so every filing is full of commitments, guarantees and stakes. But nothing requires them to disclose whether the chips are working. Every search through the filings turns up the financing, and none turns up the usage. So how can this be fixed?

What should they publish?

A usage figure alone wouldn't settle everything: a chip can be busy all day for an AI lab that pays its bills with investors' money, sometimes from the same companies that sold it the chips, rather than with money from its customers. So four numbers would answer most of this, and the companies here already hold the data for most of them:

  1. Use against capacity: chip-hours spent running work against chip-hours available, each quarter, and, separately, megawatt-hours used against megawatt-hours available.
  2. How much capacity is rented by suppliers or investors, including a supplier renting back its own customers' capacity.
  3. Revenue from customers outside the chain: customers not funded by the seller, the chipmaker, or their investors.
  4. For Nvidia: how much of the $36 billion rent-back it reported for the quarter to July has been used up by outside customers, how much by Nvidia itself, and how much is new.

Nvidia reported its results on 17 November, its first without Huang. The language hasn't changed. "Our demand is accelerating," Kress said, and Nvidia's chips remain "fully utilized across every hyperscaler and neocloud we serve." Again, no figure was provided. But two things on the call were new. Nvidia's outlook now takes into account "our customers' access to capital," and Kress said chips are allocated according to whether customers' buildings, power and money are ready, and, for smaller customers, whether they have a "solid offtake": someone lined up to pay. And asked about a new line in its disclosures, she explained that Nvidia's guarantees only start once the cloud company has raised its own money and built the data centre. If it is delayed, "our guarantee doesn't exist anymore." The demand is described as more than anyone can supply. But who gets the chips is decided by who has the money and the buildings ready to install them. That suggests the limit isn't only chips. It's money and readiness too. Nvidia, which normally files its quarterly report on results day, says it will file this one "in the coming weeks."

Whoever runs Nvidia next will be asked about China, about the board, about what Huang knew. Those are fair questions. The one I'd add is shorter.

What does "fully utilized" mean, and how do you know?


Primary sources