AI infrastructure spending has reached a scale where the interesting question is no longer whether the demand is real. It is who is paying for it, in what form, and what happens to the wider economy if the payments stop. Three separate pieces of reporting over the summer of 2026 answered one part of that each, and read together they describe something none of them states on its own.
The first is a fundraising announcement. The second is an argument about debt. The third is a municipal accounting exercise that is, as far as we can establish, the only place anyone has put a number on the downside.
$690–760bn: Estimated 2026 capital spending by Alphabet, Amazon, Meta and Microsoft combined. About 77% above the $410bn they spent in 2025. By August, Alphabet was selling a 100-year bond to help pay for its share.
The number that has changed shape
Start with the spending itself, because the headline figure is both enormous and genuinely uncertain.
| Company | 2026 guidance or estimate |
|---|---|
| Amazon | $200bn |
| Alphabet | $175bn – $185bn |
| Meta | $115bn – $135bn |
| Microsoft | Estimates diverge sharply: see note |
| Combined, all four | $690bn – $760bn depending on the source |
Company figures are guidance or analyst estimates as reported through 2026; they are not all drawn from a single filing and are not directly comparable. Microsoft is left blank deliberately: published estimates for its 2026 calendar-year capex range from about $120bn to $190bn, a gap too wide to report as one number. The combined range spans roughly $690bn (Futurum) to $725bn and $760bn (variously reported), against about $410bn in 2025.
The $725bn figure implies a 77% increase in a single year. Whichever end of the range proves right, AI infrastructure spending at this scale cannot be financed out of operating cash flow alone, and in 2026 it stopped being financed that way.
What Alphabet raised, and what it was for
On Monday 1 June 2026, Alphabet announced an $80bn equity raise. On Tuesday 2 June it priced and upsized the deal to $84.75bn. Almost every account of this described it as Alphabet raising money for AI. Both halves of that description need qualifying, and the company’s own filings do the qualifying for us.
| Instrument | Size | Terms as filed |
|---|---|---|
| Class A common stock | $9bn | 25.46m shares at $355.20 |
| Class C capital stock | $9bn | 25.46m shares at $351.80 |
| Series A depositary shares | $8.375bn | 167.5m shares at $50 |
| Series B depositary shares | $8.375bn | 167.5m shares at $50 |
| Underwritten subtotal | $34.75bn | Upsized from the $30bn announced on 1 June |
| At-the-market programme | $40bn | Class A and Class C |
| Private placement | $10bn | Berkshire Hathaway |
| Total | $84.75bn | As stated by the company |
Figures from Alphabet's Form 8-K exhibits of 1 and 2 June 2026. The four underwritten legs sum to $34.75bn; adding the $40bn programme and the $10bn placement gives the $84.75bn total the company reports. The depositary shares represent interests in 6.25% mandatory convertible preferred stock converting in May 2029.
That was equity, not debt, a distinction often lost. Alphabet did not borrow this money in June; it sold ownership, which is why the shares fell about 4% on the announcement: dilution, not leverage. The debt came later, and separately.
The $30 billion that is not for AI
Buried in the use-of-proceeds language is a detail that materially changes the headline. Alphabet states that approximately $30bn of the at-the-market programme is intended to address 2026 calendar-year tax obligations arising from employee equity vesting.
That is about 35% of the $84.75bn going to something that has nothing to do with AI compute. It is an entirely ordinary corporate need and there is nothing improper about it. But “Alphabet raises $84.75bn for AI” and “Alphabet raises $84.75bn, of which roughly $55bn is available for AI and general corporate purposes” are different sentences, and only the second one matches the filing.
What Berkshire paid
The $10bn private placement was priced below the public book, which is normal for a block of that size and worth quantifying rather than glossing.
Berkshire paid $351.81 for Class A shares against a public offering price of $355.20, a 0.95% discount, and $348.20 for Class C against $351.80, a 1.02% discount. On $5bn per class, that is worth about $48m and $51m, roughly $99m in total. That is our calculation from the prices in Alphabet’s 8-K exhibits: the value of the pricing difference, not a gain Berkshire has booked.
From equity in June to a 100-year bond in August
The shape of the financing changed over the summer, and that shift is the real story the three sources circle. Having sold equity in June, Alphabet went to the bond market in August, seeking up to $25bn in a multi-currency sale that included a rare 100-year note.
Between those two events, in July, the company reported its first ever negative free cash flow and raised its annual capital expenditure forecast for the second time in the year. A company generating less cash than it spends, increasing what it intends to spend, and moving from equity to century-dated debt inside ten weeks is a company whose funding model is under active revision.
The debt question
By late August the financing had become the story. A New York Times opinion piece published on Wednesday 26 August 2026 under the headline “Here’s How AI Debt Binge Sinks the Economy” put the argument in its strongest form. The specific figures below are not from that piece, which sits behind a paywall we did not read; they come from reporting published in the same window.
The borrowing backdrop, as reported by Fortune on 29 August 2026
- US federal debt: $40 trillion, with a projected $2 trillion deficit this fiscal year and about $1 trillion a year in debt servicing.
- Investment-grade corporate bonds: $1.7 trillion issued in the year to July, about 27% above the previous year’s pace. 2026 is projected to pass $2 trillion for the first time.
- Less visible AI borrowing: An estimated $1.65 trillion. The least firm number here, presented as an estimate.
The mechanism being described has a name: reverse crowding out. The traditional worry is that heavy government borrowing squeezes private companies out of the credit market. The argument now is the inverse: that corporate borrowing on this scale pulls capital away from Treasuries, so Treasury yields must rise to clear the market, raising the government’s own cost of carrying $40 trillion of debt.
Treasury Secretary Scott Bessent, quoted by Fortune, described AI companies’ attitude to borrowing costs as “almost yield-agnostic.”
That phrase is the part worth sitting with. A borrower indifferent to the interest rate is a borrower whose plans do not adjust when money gets more expensive, which removes the mechanism by which credit markets normally slow an investment cycle down.
Who gets paid while this happens
The clearest read on how markets have judged all this came on the same day Alphabet priced its raise.
On Tuesday 2 June, the day Alphabet priced its raise, Broadcom closed at a record, up nearly 5%, while Alphabet fell about 4% in a broadly rising market. Broadcom was up close to 40% for the year and about 90% over twelve months; Alphabet was up about 120% over twelve months, against about 27% for the S&P 500, according to Investopedia. Broadcom’s gains followed an April extension of its partnership with Google to co-develop custom AI chips.
Both companies have substantially outperformed the index. But on the day the buildout’s funding was announced, the market marked down the company doing the building and marked up the company selling it the parts. That is a coherent verdict: the equipment revenue is contracted and near-term, while the return on the infrastructure is neither.
New York has already costed the downside
Almost nobody has put a number on what happens if AI infrastructure spending slows. On Thursday 21 May 2026, New York City Comptroller Mark Levine did, in a report modelling five scenarios adapted from national work by Moody’s Analytics.
Scenario names and probabilities are from the comptroller's press release of 21 May 2026 and sum to 100%. The two job figures are as reported by ABC News from the full report; figures for the other scenarios were not in the material we could access.
Start with the odds. The downside scenarios (AI falls flat, job replacement and AI shockwave) carry a combined 50% probability in the comptroller’s own weighting. Then look at what triggers the worst of them. “AI falls flat” is not defined as the technology failing. It is defined as a drop-off in AI investment and an accompanying stock market slide.
This is the connection between the three stories. The financing risk and the jobs risk are the same risk seen from two ends: the argument about debt is about whether the investment can keep being funded, and the comptroller’s 25% scenario is what one city’s labour market looks like if it cannot. Neither report says so, because neither was looking at the other.
Levine’s recommendation follows from the spread of outcomes rather than from any single one: raise the city’s rainy day fund from 8.5% of tax revenues to 16%. That is a near-doubling of a buffer, justified not by a prediction but by the width of the range.
There is no city in America more exposed to both the promise and peril of artificial intelligence than New York City.
Mark Levine, New York City Comptroller, 21 May 2026
What would count as evidence
Most commentary on this subject is a bet dressed as an analysis. A few things would genuinely move the argument, and they are worth naming in advance so that the next round of headlines can be read against them rather than absorbed.
- Whether capex guidance is revised down. Alphabet raised its forecast twice in 2026. A cut, rather than a slower increase, would be the first real signal.
- Whether the bond market keeps absorbing the paper. Issuance above $2 trillion is only a problem if it starts pricing badly. Watch spreads, not volumes.
- Whether free cash flow turns positive again. Negative free cash flow during a buildout is a choice; negative free cash flow after one is a condition.
- Whether supplier order books lengthen or shorten. Broadcom’s backlog is a forward read on hyperscaler intent that does not depend on anyone’s forecast.
None of this establishes that an AI investment slowdown is likely. The comptroller’s most probable scenario is the benign one, at 35%, and chip demand through 2026 has repeatedly beaten expectations. The claim is narrower: the buildout is now being financed rather than funded, that changes who carries the risk, and one city has done the arithmetic on what the bad case costs.
From paying cash to borrowing
AI infrastructure spending in 2026 crossed the line from something the largest technology companies could pay for out of pocket to something they have to raise money for. Alphabet is the clearest case: $84.75bn of equity in June, roughly a third of it earmarked for employee tax obligations rather than compute, followed by its first negative free cash flow in July and a bond sale including a 100-year note in August. That is not a company in trouble. It is a company changing how it pays for things, in public, inside a single quarter.
The market has read this consistently, rewarding the suppliers and scrutinising the spenders. The credit commentary has read it as a question about who ultimately bears the cost when corporate issuance above $2 trillion meets $40 trillion of federal debt. And the New York City comptroller, alone among the three, has translated it into the only currency most people actually experience: about 52,500 jobs, at a probability of 25%.
The useful conclusion is not a prediction. It is that the three conversations (the fundraising, the debt and the jobs) are one conversation, and that the trigger the comptroller specifies for his bad scenario is precisely the thing the credit market is arguing about. AI infrastructure spending does not have to fail for New York to lose those jobs. It only has to stop being fundable.
