They became the bank
Alphabet, Meta and Amazon now spend more on AI than their businesses generate in cash, and the gap is being filled by bonds, new shares, and Nvidia.
On Thursday Microsoft put out a short statement saying it would sell $25 billion of bonds. It was ten tranches, running from 2028 all the way out to 2066, and closing on 4 November. The stated purpose was the usual phrase companies use when they don't want to be pinned down, "general corporate purposes", which can mean working capital, buybacks, acquisitions, paying down old debt, and capital expenditure. Everyone reading it knew which of those it meant. Microsoft is aiming to spend $170 billion on capital expenditure this year, most of it on data centres for the AI labs, Anthropic and OpenAI among them. Bloomberg Law called it the company's first US bond sale since 2017.
A day earlier, on the earnings call, Amy Hood, Microsoft's finance chief, had more or less trailed it. "There's a form of financing called leasing, and we've been using that one as a vehicle," she told analysts. "We're fortunate to have access in the debt markets, given where our ratings are, and we haven't yet tapped that." By Thursday night they had.
On its own that is not much of a story. Microsoft has a AAA rating from S&P, and a company with that rating borrowing money cheaply is about as ordinary as finance gets. What makes it worth slowing down for is everything else that happened this week. Taken together, I think it's the week the AI build-out stopped being paid for out of profits. The money now comes from lenders, from shareholders, and increasingly from the companies selling the equipment.
The cash ran out first at the richest companies
Start with the numbers that came out of the earnings filings, because they are the plainest part of this and they don't need any interpretation.
Alphabet reported on Wednesday. Its operating cash flow for the third quarter was $42.3 billion, which is an enormous amount of money for one company to generate in three months. It spent $51.4 billion on property and equipment in the same quarter. Free cash flow, the cash left after paying for the buildings and chips, was minus $9.1 billion. The quarter before that it was minus $5.9 billion. That's two quarters in a row in which the company that owns Google Search spent more on infrastructure than its whole business brought in. It kept its full-year capex guidance of $195 to $205 billion and said 2027 would be "significantly" higher, with details in February.
Meta went first, a week earlier. Its third-quarter free cash flow was minus $2.1 billion, as $34.9 billion of capex overtook $32.8 billion of operating cash flow, and the stock fell nearly 7%. Axios ran the headline "Meta joins Big Tech's negative free cash flow club", which is a sentence that would have read as satire two years ago. Amazon reported on Thursday and was negative too. Over the past twelve months its free cash outflow was $42.4 billion, on $210.7 billion of capex, and it said 2027 spending would be "meaningfully more" than this year's $220 billion. On the same day, Andy Jassy told analysts that AWS, growing at 39% and now running at $184 billion a year, could become a trillion-dollar revenue business in five to ten years. He also said Amazon would have to keep spending heavily to get there.
Microsoft was the exception, positive on free cash flow, and it went to the bond market anyway. Oracle, which is not positive and doesn't expect to be until the second half of its 2028 fiscal year, announced $10 billion of new debt the same day.
I don't want to make this sound like distress, because it isn't, not at these companies. Alphabet and Meta and Amazon are borrowing against businesses that print money, and Meta is still sitting on about $82 billion in cash and securities. But it changes the character of the thing. For most of the last three years, the honest answer to "who is paying for all this?" was that the four or five richest companies in history were paying for it out of their own pockets, and if it turned out to be a mistake, it would be their mistake to eat. That answer is no longer quite true. Now it's partly the bond market, and partly the people who buy shares.
Meta asked its suppliers to buy in
The shareholders part arrived on Monday, when Meta said it would raise $95 billion in new equity, $35 billion through underwritten offerings straight away and another $60 billion through an at-the-market programme over 2027. The immediate offering was later sized at $37.5 billion, split between common stock and mandatory-convertible preferred shares. Meta is one of the most profitable companies on earth, and it's issuing tens of billions of dollars of new stock to pay for data centres.
And then on Tuesday, the detail that tells you most about where we are. Nvidia said it would put $3 billion into Meta's share sale. AMD said it would put in $1 billion. Meta's own announcement had said Nvidia, AMD and Broadcom were interested in buying up to $5 billion between them. The chip suppliers are buying shares in their customer so that the customer can afford to keep buying chips. Nvidia's press release folded it into an expanded partnership, more Vera Rubin and Vera CPU deployment and closer work on Muse, Meta's personal agent. Jensen Huang supplied the line about agents being "a new computing paradigm", and Mark Zuckerberg said he was excited to keep working together. Nvidia's stock was flat that day. Nobody seemed surprised.
They shouldn't have been, because Nvidia has been doing this for a while. According to Dealroom, as reported by CNBC, it has taken part in 121 investment rounds so far this year, up from 13 in 2022. AMD has done 25, up from two. Nvidia has stakes in OpenAI, Anthropic, xAI, Intel, Mistral, Wayve, CoreWeave and Nebius. And beyond the equity, it has started doing something closer to what banks do.
The supplier as lender of last resort
The clearest example is CoreWeave. Nvidia invested $2 billion in the company and committed to buy up to $6.3 billion of CoreWeave's unsold computing capacity through April 2032, if nobody else does. That's a promise to be the customer of last resort for a business whose main asset is Nvidia chips. Nvidia's own filings show about $36 billion of these cloud backstops in total, plus a guarantee of up to $105 billion for OpenAI's planned data centre in Ohio.
Then, late on Thursday London time, The Information's Phoebe Liu reported that Nvidia is in talks to give Nscale, the British-founded AI cloud company it has invested in, billions of dollars of credit guarantees. The money would back debt for the power infrastructure and construction of Nscale's campus in West Virginia, the one Anthropic agreed in August to rent for about $45 billion over six years. It's roughly 460 megawatts, meant to run on Nvidia's Vera Rubin chips and come online in late 2027. The report says the deal isn't signed and the terms could change, and Nvidia, Nscale and Anthropic didn't comment to Reuters. But the shape is clear enough. Nvidia would stand behind the loans that pay for the building that holds the Nvidia chips that Anthropic will rent.
On X the reaction split along the usual lines. The market accounts posted it in capitals, and one small macro account titled its post "The AI Family Business". The most useful framing I've seen of these deals came from Gil Luria at D.A. Davidson, speaking to AFP before this week's news. "If Nvidia gives CoreWeave money to turn around and buy Nvidia chips, that's not as good as Microsoft buying Nvidia chips because they have a customer in OpenAI that has a customer in an enterprise technology that's willing to pay for it," he said. In the same piece he called the approach "unprecedented, but it appears to be working". Michael Cusumano at MIT Sloan put it more simply: "They're essentially becoming a bank."
Microsoft, as it happens, is becoming a bit of a bank too, though more quietly. Its quarterly filing on Wednesday showed $10 billion of revenue from its commercial arrangements with OpenAI in the September quarter, about 11% of the whole company's revenue and more than Bing or Xbox. It also showed $7.2 billion of accounts receivable from OpenAI at the end of September, up from $6.0 billion at the end of June. Receivables are bills sent and not yet paid. A 20% rise in three months doesn't mean anything is wrong, and OpenAI is not short of investors, but it does mean Microsoft is extending more trade credit to a company it owns about a quarter of. When the supplier, the landlord and the lender are the same few firms, it gets harder to tell from the outside how much of the demand would still be there if they stopped.
Where the risk actually ends up
None of this would matter much if money were cheap and patient. It is getting less so. CNBC's line on the Microsoft deal was that investors have turned more skittish about companies with tall debt piles "as central bankers raise interest rates to fight inflation". And October is when the weakest part of the chain showed what that looks like.
Early in the month, Firmus pulled what would have been one of the largest listings in the history of the Australian stock exchange, a $44 billion float, after investors baulked at the valuation. The following week CoreWeave, the original neocloud, had its worst week in nearly a year. Its shares fell 15.5% on one Wednesday and about a quarter over the week, and the yield on its 2031 bonds went past 16%. Bonds it sold as recently as June had come with a 9.675% coupon. The cost of insuring its debt against default rose to between 1,000 and 1,200 basis points, which Reuters said implied close to a 63% chance of default within five years. Elon Musk replied "yikes". Mike Intrator, CoreWeave's chief executive, told staff in a memo that the company's backlog, $104 billion at the end of June plus more than $25 billion of new commitments since, more than covers its debts, and that markets can "misunderstand and act irrationally". He may well be right about the backlog. The trouble is that a backlog gets paid out over years, while the interest on $50 billion of debt is due every quarter. By the end of the month, according to AFP, CoreWeave, Nebius and IREN were each down at least 30% for October.
This is the bit I think people get wrong when they argue about whether there's an AI bubble. The argument is usually framed as whether the demand is real, and the honest answer from the supply side this week was yes. Samsung reported a record quarterly operating profit of 107.4 trillion won on Thursday. Its high-bandwidth memory sales more than tripled from the previous quarter, and it said its entire 2027 HBM allocation is already committed under fixed contracts. Even the sceptics quoted this week, like Paulo Carvão at the Harvard Kennedy School, say the market for AI chips is still limited by supply, not demand. Nobody who sells shovels is short of buyers.
But demand being real isn't the same as every borrower being safe. The labs will rent the capacity, but they'll rent it on a schedule, and it has to be built, powered and financed before that, at interest rates that are rising. The richest companies can carry that gap on their own balance sheets, or borrow at AAA rates to cover it. The neoclouds can't, so they borrow at 10% and then 16%, and when that stops working the supplier steps in with a guarantee. So the risk moves down the chain to the firms least able to hold it, and then, when it gets too heavy for them, back up to the one firm with the most to lose if the orders stop. Carvão told AFP that "the closer these relationships are, the harder it is for one to assess true sustainable demand", and that "if AI demand disappoints, spending cuts could hit chipmakers, datacentre builders and their financiers at once". Nvidia's own annual report says two direct customers made up 22% and 14% of its revenue last fiscal year.
Someone has already said who won't pay
Which brings me to the strangest document of the week. It wasn't from a company at all.
On Thursday, Greg Casar, who chairs the Congressional Progressive Caucus, and Maxine Waters, the senior Democrat on the House Financial Services Committee, put out a resolution opposing any federal bailout of AI companies if the bubble bursts. Eleven other Democrats signed on. It says the government should oppose emergency lending or capital injections "to prop up insolvent A.I. companies' operations or prevent their orderly resolutions through ordinary bankruptcy processes", and that executives who fail should hand back the mansions, the planes and the yachts. "A.I. companies that fail should go bankrupt, and executives who commit fraud should be held accountable. Federal help should go to workers and retirees," Casar said. Waters invoked 2008 by name. The resolution itself claims that AI-related firms hold "trillions of dollars in debt, much of which is raised in opaque and unregulated private credit markets".
It's non-binding, it comes from the minority party, and at least one fact-checker noted that as of Thursday it hadn't shown up in the House record with a number and a committee referral, so it may be closer to a statement of intent than a filed bill. I wouldn't read it as policy. I'd read it as a marker. A year ago, the idea that Congress would need to say in advance that it won't rescue the AI industry would have sounded absurd, because the AI industry was four companies with more cash than most countries. The fact that a dozen members of Congress felt the need to say it this week tells you that the financing has changed enough for people outside finance to notice.
What I'd watch
The easy way to tell this story is as a bubble about to pop, and I don't think that's right, or at least not yet. The labs' revenue is real. Microsoft booked $10 billion from OpenAI in a single quarter, and AWS is growing at 39%. The chips are sold out, and the memory is sold out a year ahead.
The more accurate version, I think, is less dramatic and more lasting. The build-out has grown past what even the richest companies can fund from what they earn. The shortfall is being filled by bonds, by new shares, and by the suppliers themselves, and those suppliers are now quietly acting as the industry's lender of last resort. That works fine while the orders keep coming and the money stays cheap. It gets uncomfortable if one of those stops. So the things I'll be watching aren't the model launches. They're the Nscale terms if they're signed, CoreWeave's bond yields, Microsoft's receivables from OpenAI next quarter, and Alphabet's 2027 capex number in February.
Nvidia has become the bank because nobody else was willing to lend on these terms. Whether anybody stands behind Nvidia is a question the House Democrats have, at least for now, answered in advance.