AI is Burning Billions. Who Gets the Bill?

One company's capital expenditure is another company's revenue. That works beautifully—until the money has to come from somewhere outside the loop. How does a capital cycle this large eventually turn back into real economic value?

Part 1 of a series on the economics of the AI buildout: what it costs, who finances it, and who ends up owning what.

---

Start with the least disputed fact in the entire AI argument.

For 2026, Amazon has guided to around $200bn of capital expenditure. Alphabet to $175-185bn. Meta raised its range to $125-145bn. Microsoft is tracking $110-120bn. Together that is somewhere near $630-700bn in a single year, against roughly $388bn the year before.

Nobody disputes these numbers, because the companies published them. What people disagree about is what they mean.

Where the money goes

Follow it one step and something becomes obvious.

Nvidia's financial year 2026 ended on 25 January. Revenue was $215.9bn, up 65%. Data centre alone was $197.3bn, up 68% from $115.2bn, and now more than nine tenths of the company.

That is not two separate booms. It is the same money, counted twice, at two points in its journey. What the hyperscalers book as capital expenditure, Nvidia books as revenue. Nvidia's revenue supports Nvidia's valuation. That valuation is part of the market's confidence in the whole category, which is part of what makes the next round of spending fundable.

Money goes round, and each lap makes the lap look justified.

The transfer window, but for GPUs

If that sounds abstract, there is a system most of Europe already understands in its bones.

Football clubs buy players from each other. In 2025 they spent a record $13.11bn on international transfer fees, more than half again what they spent the year before, across more than 86,000 transfers. Every one of those fees is simultaneously one club's cost and another club's income. A club that sells well books a profit, that profit funds the next signing, and the fees keep climbing because the money keeps moving.

The accounting rhymes as well. A transfer fee is not treated as a cost in the year it is paid. The player's registration goes onto the balance sheet as an asset and is written down across the length of his contract, so an 80 million pound signing on a six-year deal shows up as roughly 13 million a year. The fee is enormous; the annual charge is manageable. Clubs choose contract lengths knowing exactly that.

And yet no amount of clubs trading with each other has ever made football richer. The money that genuinely enters the sport comes from outside it: broadcasters, sponsors, people buying tickets and shirts, and, increasingly, states and private fortunes of widely varying provenance, the all works. Transfers move that money around at high speed and in public. They do not create it.

That last category is where the comparison stops being a comparison. Saudi Arabia's Public Investment Fund owns Newcastle United. Qatari state investment owns Paris Saint-Germain. Abu Dhabi money owns Manchester City. Those same funds are now among the largest outside investors in AI infrastructure. Abu Dhabi's MGX, a vehicle of Mubadala and G42, is a Stargate partner and has closed a $49bn AI fund, and it led a $40bn purchase of Aligned Data Centers alongside BlackRock's infrastructure arm, Microsoft, Nvidia, Kuwait's investment authority and Temasek. The Saudi fund put $36.2bn into AI-related deals in 2025. Sovereign funds together put roughly $66bn into AI and digital infrastructure that year.

Which complicates the question this series is asking. When a growing share of the outside money belongs to a state, the return being sought may not be financial, and "does it pay for itself" stops being the only test it has to pass.

Hold onto the accounting detail. The equivalent question about AI hardware, how long the thing you bought counts as an asset, turns out to be the most consequential number in this entire industry, and it is part two.

This is not a scandal

It is worth stopping here, because this is where a certain kind of article reaches for the word "bubble" and stops thinking.

Nothing described above is fraudulent. It isn't even unusual. Every capital cycle in history has worked this way. Railways, electrification, fibre: someone spends enormous money on infrastructure long before anyone can prove what it will be worth, and the spending itself creates real revenue for suppliers, real jobs, and real assets.

The circularity isn't the problem. The circularity is what a buildout is.

The question is narrower and harder. A loop like this has to terminate somewhere outside itself. Chip revenue justified by cloud spending justified by chip revenue is a closed system, and closed systems don't repay capital. At some point the money has to come from someone buying a product because it made them better off, in an amount larger than the infrastructure cost to serve them.

That hasn't been demonstrated yet at anything like this scale. It also hasn't been disproven. It is genuinely open, and most writing on the subject pretends otherwise in one direction or the other.

Where the money comes from

For most of the last decade this was a boring question. The companies doing the spending were among the most cash-generative businesses in history, and they paid for infrastructure out of operating cash flow. Boring is a feature, because money you already earned comes with nobody attached to it. Borrowed money does. A loan carries covenants: conditions the lender writes in, such as keeping debt below some multiple of earnings or not selling particular assets. Break one and they can demand the money back early. Nobody can do that to you over cash that was already yours.

That has changed, and the International Energy Agency states it plainly:

"Data centre investments have grown too large to be funded from company balance sheets alone, and large amounts of funding from capital markets will be critical for their buildout."

Then the part worth reading twice. The pace of data centre growth, the IEA says, "will be sensitive to market sentiment, including expectations for returns on investment in data centres and AI deployment, as well as to broader macroeconomic and financing conditions." The build rate is now coupled to the mood of the bond market.

The numbers behind it are large and recent. Amazon has raised roughly $53bn of debt this year, including a $37bn dollar offering and about 14.5 billion euros in March. Alphabet issued $31.8bn of foreign-currency notes in the first half of 2026 alone, spread across sterling, Swiss francs, euros, Canadian dollars and yen. Analysts at BofA raised their forecast for hyperscaler debt issuance in 2026 to $175bn, from $140bn. Estimates for AI-related issuance across the wider ecosystem, including chipmakers, developers and utilities, run from roughly $300bn to $570bn.

Issuing in five currencies is not a treasury preference. It is what you do when no single market can absorb what you need to raise.

Some of it does not sit on the balance sheet at all. Put the data centre and the loan against it inside a separate company, and the parent's own accounts stay cleaner, which protects its credit rating and its room to borrow again. This is legal, disclosed and old. It also means the balance sheet you can see understates what has been promised.

Here is why the shift matters more than the size of any single number. Equity-funded mistakes fail quietly, over years, and mostly punish the people who chose them. Debt-funded ones have dates attached. Debt has to be repaid or replaced on a fixed date, whatever is happening that month. That calendar does not care whether the technology eventually works. It cares whether it works in time.

The word doing the heaviest lifting

There is one place the argument gets slippery, and it is worth learning to spot.

When one of these companies reports demand, it can mean two entirely different things. It can mean consumption: people used the service, and the meter ran. Or it can mean commitment: somebody signed a contract promising to buy compute later.

Both are real. But they are not the same asset.

As of 31 March 2026, OpenAI's purchase commitments were reported at roughly $665bn, covering chips, power and data-centre capacity from Microsoft, Oracle, Amazon and the Stargate projects. Oracle reported a contracted backlog of about $523bn in April 2026. The Oracle-OpenAI arrangement alone is reported at $60bn a year for five years, beginning in 2027.

Every one of those is a promise about years that have not happened yet.

This is not hidden. Contracted-but-undelivered revenue has an accounting name, remaining performance obligations, and it sits in the filings so you can find it. But it gets reported, discussed and priced as though it were the same thing as customers paying today, and it isn't. A backlog is only worth what the counterparty can eventually pay.

Something you can check yourself

Here is the arithmetic, and it takes about five minutes per company.

Open the most recent 10-Q. Find remaining performance obligations. Divide by the last twelve months of revenue.

That ratio tells you how many years of current business the company has already booked as promises. A modest number means contracts are a normal part of the operation. A very large number means the valuation depends on a future that is contractually described but not yet delivered.

Then ask the second question, which matters more: who owes it? A backlog spread across thousands of customers is a forecast. A backlog concentrated in a handful of counterparties who are themselves funding their obligations from capital markets is something else. It is a bet on those specific companies, wearing the clothes of a bet on the technology.

Do that for the largest AI infrastructure providers and you will learn more than any argument about whether the technology is real.

What this series is going to do

The technology works. That question is settled well enough here, and it is the wrong question anyway.

The open question is whether the value it creates compounds fast enough to repay the capital, the energy and the infrastructure being committed to it now. That depends on things that are measurable, disclosed quarterly, and almost never discussed: how long the hardware stays economically useful, who captures the value once it exists, what the physical constraint costs, and which of the promises turn into cash.

Those are the next parts. None of them require predicting anything.

For your cloud provider: what is its remaining performance obligation, and who owes it?

---

Sources: Sovereign investment figures (MGX's $49bn fund and Stargate partnership, the $40bn Aligned Data Centers consortium, PIF's $36.2bn of 2025 AI transactions, roughly $66bn of sovereign-fund AI and digital infrastructure investment in 2025) are as reported. Club ownership is a matter of public record. FIFA Global Transfer Report for 2025 international transfer spending of $13.11bn across more than 86,000 transfers; player registrations are capitalised and amortised over contract length under IFRS, and the Maguire illustration is the standard worked example. IEA, Key Questions on Energy and AI, for both quoted sentences on capital-market dependence. Debt figures (Amazon's approximately $53bn including a $37bn offering and about 14.5 billion euros in March 2026; Alphabet's $31.8bn of foreign-currency notes in H1 2026; BofA's $175bn hyperscaler forecast; $300-570bn ecosystem-wide estimates) are as reported and move monthly. 2026 capital expenditure figures are company guidance as reported and have been revised upward repeatedly during the year; treat the ranges as of publication. Nvidia FY2026 results (year ended 25 January 2026) are from the company's annual report: revenue $215.9bn, data centre segment $197.3bn. OpenAI purchase commitments of roughly $665bn as of 31 March 2026, Oracle's contracted backlog of roughly $523bn as of April 2026, and the reported $60bn-per-year Oracle arrangement beginning 2027 are as reported rather than read from a filing by me, and the totals differ between accounts, so verify against the current 10-Q before relying on any of them. Accurate as of 3 September 2026.

This is a personal analysis of public filings, not investment advice.