The AI bill is landing on those who never invested in it

Hyperscalers are the companies that build and operate the largest computing capacity on the planet, such as Amazon, Microsoft, Alphabet, Meta and Oracle. For two decades they paid for their data centres with the cash they generated. With artificial intelligence that is no longer enough, and the wait until the revenue arrives is starting to be paid for by people who never asked to take part.

On Thursday Amazon reported the best operating quarter in its history. A record margin of 13.9%, and AWS, its cloud business, growing by 39%. And, a little further down the earnings release, trailing twelve-month free cash flow of minus $42.4 billion. A year ago it was positive to the tune of $14.8 billion. The following day the shares were down around 7% before the open, and not because the business is doing badly. They fell because the market finally did a sum it had been putting off for months.

The explanation lies in a single line of the release itself. Purchases of property and equipment rose by $94.8 billion in a year, “primarily reflecting investments in artificial intelligence”. Between July and September alone Amazon spent $77.3 billion on capex, when analysts were expecting $58.7 billion. And its president and chief executive, Andy Jassy, has already warned that in 2027 it will spend significantly more than this year’s $220 billion.

The fibre outlived those who laid it

To my mind, this goes far beyond an earnings story. For twenty years the big technology companies were cash-generating machines that bought back shares with whatever they had left over. Today four of the five biggest AI investors are burning cash, and Bank of America estimates that the group’s combined free cash flow will reach minus $144 billion in 2027. Artificial intelligence demands the outlay now and promises the revenue later, and in between lies a gap that someone has to finance. The interesting question is no longer whether AI will make money, which it probably will, but who bears the wait. I will give you my answer up front. More and more people who never chose to be part of this bet are in it up to their necks.

Amazon is not an isolated case, nor even the most striking one. Alphabet, Google’s parent company, bought back $45.7 billion of its shares in 2025. In the first nine months of 2026 it has not bought back a single one. Instead it has done something it had never done before: sell shares for cash, $84.75 billion in June, with Berkshire Hathaway putting in $10 billion. Its debt has risen tenfold since the end of 2024, to $126.1 billion. In February it placed a sterling bond maturing in 2126. What can I say? When a company that lives off advertising borrows for a hundred years, something has changed in the nature of the business.

Meta recorded its first negative free cash flow in the third quarter, with capex doubling this year to around $140 billion. It has stopped buying back shares, has created a class of non-voting shares so that it can issue equity without Mark Zuckerberg losing control, and is preparing the largest share issue in its history, of up to $97.5 billion. The cost of insuring its debt against default hit a high on 22 October. Nobody thinks Meta is going to go bust. The insurance is rising because bondholders are discovering that they are lending to a company that now needs outside money to sustain its plan.

The electricity bill is the part of the wait that nobody chooses to pay

Microsoft is the exception, and it deserves a nuanced look. It still generates free cash flow, $18.6 billion in the quarter, albeit 27% less than a year ago, and devotes 68% of its operating cash flow to capex, compared with 121% at Alphabet. Two days before its results it offered voluntary redundancy to 60,000 employees, 27% of its global workforce. I would not go so far as to say that one explains the other, but the order of priorities seems fairly clear.

At the opposite extreme are Oracle, downgraded by S&P to BBB-, one notch above junk, and CoreWeave, the GPU rental company that in mid-October was paying more than 15% on its debt. When the market demands 15% from a company on which part of Microsoft’s capacity depends, what it is doing is putting a price on the wait.

The advocates of this investment have a powerful argument, and I acknowledge it. Demand is not an assumption. AWS has a contracted backlog of $578 billion, Google Cloud $561 billion and Microsoft $677 billion. With those numbers, building looks like the prudent thing to do.

The problem lies in who is signing those contracts. In January, Microsoft’s chief financial officer acknowledged that around 45% of its commercial backlog came from OpenAI, and the specialist press attributes close to half of Oracle’s to OpenAI. According to the plan it presented to its investors, published by the Financial Times, OpenAI expects to burn $278 billion between 2026 and 2030. The customer that justifies a good part of its suppliers’ spending needs to raise capital without pause in order to pay them. This very week, according to press reports, OpenAI cut its funding round from $30 billion to $20 billion and its target valuation from $1.4 trillion to $1.25 trillion, and as of 31 October it had commitments of around $11 billion.

Amazon is earning more than ever and generating less cash than ever

Anthropic, the other big lab, is in a better position, with an operating profit in the second quarter. But almost half of its revenue in the first half was billed through the Amazon, Google and Microsoft platforms, and it has committed to buying more than $110 billion of computing from each of the first two, which are also its shareholders. The money goes full circle before it shows up as revenue.

And then there is Nvidia, which sells chips to everyone. It has invested in OpenAI and is guaranteeing up to $105 billion for a data centre in Ohio that OpenAI will lease and fill with Nvidia chips. Its average collection period has gone from 45 to 60 days. When a supplier gets paid later and guarantees its customers’ leases, it is taking on part of their risk. We have seen this film before. In 2000 Lucent had committed $8.1 billion of financing to its customers, 24% of its sales, so that they would buy its equipment. Three years later its shares were worth 99% less.

If negative free cash flow were the whole bill, it would be manageable, because these companies have enormous balance sheets. What worries me a good deal more is what does not appear on them.

Meta has not recorded on its balance sheet the $27.3 billion of debt for Hyperion, its large campus in Louisiana. It was issued by a special purpose vehicle in which Blue Owl funds hold 80% and Meta 20%, and Meta has committed to leasing the centre and to guaranteeing part of its residual value. The bond received an A+ rating, the kind that pension funds and insurers can buy. In October it traded as low as 91 cents. Oracle does something similar on a larger scale. It has $288 billion in signed data centre leases that have not yet commenced, 2.3 times its debt, and almost half of its operating cash flow in its latest quarter came from customer prepayments, money paid in advance by those who have not yet received the service.

The customer that justifies the spending needs financing to pay for it

The Wall Street Journal estimates that nine large technology companies have accumulated some $3 trillion in off-balance-sheet commitments, between leases yet to commence and agreed purchases of chips and energy. That is three times the debt and leases they do record. The wait has been sliced up and shared out among bondholders, private credit funds, insurers and savers who probably do not know they are holding it in their portfolios.

Add to that an unresolved accounting question. The big players depreciate servers and GPUs over five or six years, while Nvidia launches a new generation every year. Michael Burry calculates that, with more realistic useful lives, profits from 2026 to 2028 would be overstated by $176 billion. There are counter-arguments. Older GPUs are reused for inference, and renting them has even become more expensive. But the doubt explains why the market has stopped focusing on earnings per share and is looking at cash flow, which does not depend on how many years a chip lasts.

So far, those bearing the wait have at least taken on the risk, even if through a fund. There are others who have not.

The electricity bill is the clearest example. On the PJM grid, which serves 67 million people across thirteen US states, the price of electricity capacity has gone from $28.92 per megawatt-day in 2024-2025 to $329.17 in 2026-2027, the cap allowed by the regulator. The market’s independent monitor attributes 63% of the 2025-2026 increase to data centres, some $9.3 billion passed on to all consumers. It is no coincidence that this Tuesday American voters will decide 34 ballot measures on data centres. In Aragón, Red Eléctrica already puts the capacity of data centres with grid connection permits at 3,700 megawatts, more than double the peak demand of the entire region. The financial risk stays in the United States; the pressure on the grid and on water stays here.

Whoever pays in advance is also lending

The second channel is interest rates. The minutes of the Federal Reserve’s September meeting note that AI investment has replaced tariffs as the main driver of underlying prices. The Fed raised rates in September and the market is pricing in further hikes until mid-2027. The mechanism is perverse. Investment pushes up inflation, inflation forces rates higher, and higher rates make the debt that finances the investment more expensive. Kristalina Georgieva, managing director of the IMF, summed it up a few days ago with an image that is hard to improve on: an economy standing on a single leg, that of AI.

Are we, then, looking at a bubble about to burst? I do not know, and I distrust anyone who claims to. Investment in computers and data centres is running at around 1.3% of US GDP, almost the same as telecommunications in 2000. That fibre was not useless. It ended up carrying streaming video and the mobile internet, but it did so in the hands of those who bought it cheaply after the collapse of those who had laid it. And there is a difference that works against today’s builders. Fibre lasts for decades, whereas a GPU can be superseded in three or four years, and that leaves much less time for demand to catch up with supply.

A company that lives off advertising now borrows for a hundred years

I would keep an eye on two things in the coming weeks. Whether OpenAI closes its funding round, and on what terms, because half the chain of contracts depends on its ability to raise money. And the prepayments line in Oracle’s December accounts. If customers stop paying in advance, it will show up there before it does in any press release, and it will be the sign that nobody wants to finance the wait any longer.