Big Tech now borrows to pay for the AI buildout

For the first three years of the AI boom, the largest technology companies paid for their data centers the way a wealthy family pays for a kitchen renovation: out of current income, with some grumbling at the dinner table and no call to the bank. That arrangement ended in 2026. In the second quarter, Alphabet’s free cash flow was negative $5.9 billion, Amazon’s trailing-twelve-month free cash flow was negative $7.6 billion, and Meta’s free cash flow fell to $784 million against $31.1 billion of capital spending. In June, Alphabet sold $49.6 billion of new stock and convertible preferred to pay for, in its own words, “capital expenditures to scale AI infrastructure and global compute.” Oracle, which in its August quarter spent $28.5 billion on capex against $19.3 billion of revenue, now sits one notch above junk at S&P.

I don’t think the data can tell you whether AI is a bubble, and I make no claim about that. The data does show a change in how the buildout is paid for. During 2026 it stopped being an operating expense of very profitable companies and started being financed like infrastructure, with long-dated bonds, equity issuance, project-finance vehicles and debt-heavy specialist clouds, the way railroads and telecom networks were financed before it. This change matters because it changes who gets hurt if demand disappoints. When a capex cut is funded from cash flow, the pain lands mostly on shareholders and on Nvidia’s order book. When it is funded with 40-year bonds and residual value guarantees, the list of people exposed gets much longer, and some of them, like electricity customers in the mid-Atlantic, never chose to be in the trade.

Cash flow ran out before the ambition did

After the late-July earnings calls, the four biggest spenders guided to $720 billion to $745 billion of capital expenditure for 2026: Amazon about $220 billion (raised from about $200 billion, partly on the cost of memory), Alphabet $195 billion to $205 billion, Meta $130 billion to $145 billion, and Microsoft about $175 billion for the calendar year. MLQ tallied the same range in early August, and adding the guidance myself gives the same answer.

Microsoft’s additions to property and equipment for the fiscal year ended June 2026 were $115.9 billion, against $64.6 billion a year earlier. On the July call it cut its calendar-2026 capex estimate from about $190 billion to about $175 billion, because data centers will now be depreciated over 25 years instead of 15 and more future leases will fall outside the capex line. The CFO said, as CFO Dive reported, “Outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged.” The cut reflects a change in accounting policy and does not mean a slowdown.

In the June quarter, Alphabet generated $39.1 billion from operations and spent $44.9 billion on property and equipment, twice what it spent a year earlier. Amazon spent $54.2 billion in the quarter and $173.0 billion over twelve months, more than the $161.4 billion its operations produced. Meta’s operating cash flow of $31.9 billion almost exactly matched its capex plus finance-lease payments. For the first time in this cycle, three of the largest cash generators in the world were spending all of it, or more, on construction.

To cover the gap, these companies went to the capital markets. Alphabet’s June equity raise came alongside $20.3 billion of senior notes in the quarter and a $40 billion at-the-market stock program, which the company says is mainly for tax obligations on employee equity grants. Oracle launched a $25 billion bond and a $20 billion equity distribution agreement in February, then sold $19.9 billion of stock through that program in the June-to-August quarter alone. Its borrowings stood at about $125 billion at the end of August, and S&P cut it to BBB- in July. The price of insuring Oracle’s debt against default for five years went from 144 basis points at the start of the year to about 203 in July, according to LSEG data cited by the Financial Times, and a Seeking Alpha report republished on TradingView put it at a record 227.15 basis points in late September.

Bond buyers as a group have started to ask for more. Alphabet, Amazon, Meta, Microsoft and Oracle sold about $220 billion of bonds in 2026 through August 10, according to BNP Paribas data reported by Reuters, and 78 of the 91 hyperscaler bonds issued this year with comparable data were trading at higher yields on July 28 than when they were sold. (Other trackers count this issuance differently.) Amazon’s $25 billion deal in July drew an order book of 2.5 times the offering, down from 3.2 times in March, and Apollo’s chief economist measured the hyperscaler cover ratio this way, as Fortune quoted: “It was nearly 5x in February 2026 but tumbled to below 2x in July.” Demand is still more than enough to clear these deals, but buyers now want a higher yield on each new one than they did in the winter.

Some of the borrowing does not show up as debt. Meta’s Hyperion campus in Louisiana is financed through Beignet Investor, a vehicle that sold a $27.3 billion bond, with Blue Owl holding 80% of the joint venture’s equity. Meta leases the buildings in four-year blocks and gives a residual value guarantee for 16 years that pays bondholders back if it walks away. S&P rates the bond A+ once the campus is operating, on a debt service coverage ratio of 1.12 times. The vehicle sits off Meta’s balance sheet while the risk stays with Meta, and IFR’s summary puts it well: “Beignet is Meta credit in project-finance clothing.”

The case against worry is strong, and none of these companies is a distressed borrower. Alphabet’s free cash flow over the trailing twelve months was still $53.3 billion, and Meta held $90.3 billion of cash and securities at the end of June. Selling stock at a high valuation is what a good CFO is supposed to do. The spending also rests on signed contracts. Oracle reports $664 billion of remaining performance obligations, and Amazon’s chief executive said on the July call that the company “will still not have enough capacity to meet all the demand we have in 2026,” as TIKR reported. Utilities borrow against contracted demand all the time. Once a company borrows against its backlog, the quality of the customers behind that backlog becomes a concern for its lenders too, and in AI those customers are a short list of young companies.

Construction, imports and supplier revenue all rose sharply

The physical boom shows up clearly in three primary data series I track, and all three moved hard this year.

The first series is construction. The Census Bureau’s private construction table for August 2026, released October 1, puts data center construction at a seasonally adjusted annual rate of $84.95 billion, up 73% from $49.05 billion in August 2025 and more than five times the $15.6 billion rate of January 2023. General office construction, over the same twelve months, fell 9.7% to $45.8 billion. Census files data centers under “office” in its press release, so the headline says private office construction rose 29.8% in a year, which on its own would suggest a return to the office. Data centers are now about 63% of that office category, while total US construction was down 1.7% year on year. I chart the monthly series at /buildout/spending/.

The second series is imports, because much of what goes inside those buildings is made abroad. The August trade release, published October 6, shows US imports of computers of $255.0 billion from January through August, against $130.9 billion in the same months of 2025, and computer accessories of $158.5 billion against $89.5 billion. Goods imports from Taiwan for those eight months were $188.7 billion, almost exactly level with China’s $190.3 billion. Monthly imports from Taiwan first passed China’s in December 2025. Tariffs have not visibly slowed any of this, partly because the 25% Section 232 tariff that took effect on January 15, 2026 covers “a very narrow category” of advanced chips, in the words of White & Case’s summary, and it exempts chips imported for use in US data centers. The monthly import data is at /machine/imports/, and the Taiwanese server makers’ sales are at /machine/sales/.

The third series is the supply chain’s own revenue. Nvidia reported $89.0 billion of data center revenue in the quarter ended July 26, up 117% from a year earlier, and guided to $108 billion of total revenue for the current quarter. TSMC’s September revenue was NT$511.9 billion, up 54.6% year on year. At the 32 NT dollars per US dollar it assumed in its July guidance, its July-to-September revenue comes to about US$46.7 billion, according to Neoteo’s tally of the monthly reports, which is above the US$44.6 billion to US$45.8 billion range it guided to.

This spending shows up in GDP, though less than the capex totals suggest, because so much of the hardware is imported. The Bureau of Economic Analysis revised second-quarter growth up to 2.2% in its September 30 third estimate, from 1.5%, and said the upward revision to fixed investment was led by nonresidential structures, “mainly data centers.” ING’s economists, working from the earlier vintage of the data, estimate that tech investment accounted for between 36% and 50% of year-on-year growth in the second quarter depending on the definition, and they prefer the 36% figure because it nets out imported hardware. Jason Furman’s famous estimate that information-processing equipment and software made up 92% of growth in the first half of 2025 is a gross figure for an earlier year. My summary is that roughly a third of current US growth leans on this investment.

Lab revenue is growing fast, and the labs count it differently

Revenue growth is the strongest argument against worry, and I take it seriously. Anthropic’s annualized revenue run rate topped $65 billion at the end of July 2026, according to a person who spoke to Reuters, up from about $9 billion at the end of 2025. I can’t think of a software business that has grown like that, and I’ve tried.

That number is hard to compare with OpenAI’s. On September 29, Axios reported that OpenAI was on pace for annualized revenue of “almost $70 billion”. On October 8, Axios reported a figure of $50 billion and explained the gap: Anthropic records the full value of sales made through cloud partners as revenue and books the partner’s cut as an expense, while OpenAI records only its share, and the $70 billion figure had been an attempt to restate OpenAI on Anthropic’s basis. Both methods are said to comply with GAAP. A run rate is also one month multiplied by twelve, which is a fine way to describe momentum and a poor way to service a 2056 bond. The same report noted that Anthropic’s reported second-quarter operating profit excluded stock-based compensation, so even the profit figure comes with its own adjustment.

Analysts’ estimates of the revenue the buildout needs are far higher. JPMorgan estimated last November that the industry needs about $650 billion of annual revenue, in perpetuity, to earn a 10% return on the investment it modelled through 2030. Bain’s new technology report, released September 29, puts the build cost of data center capacity at $5 trillion to $6.5 trillion through 2030, and Bloomberg reported that Bain thinks the industry needs $6 trillion of annual revenue by 2031, of which existing AI services could supply up to $1.8 trillion. Sequoia’s David Cahn has updated his earlier revenue-gap estimate to roughly $3 trillion of revenue needed to justify 2026’s infrastructure spending alone, by TechCrunch’s account. These are payback models and not forecasts, and they disagree by a factor of several, but even the lowest sits far above what the two leading labs report today.

I think the gap between asset lives and debt maturities matters more than the revenue gap. The buildings may last 25 years, which is now Microsoft’s accounting assumption. I’d expect the accelerators inside them to be replaced on a much shorter cycle, the leases in a structure like Hyperion run in four-year blocks, and the bonds that pay for all of it run to 2066 in Oracle’s case. Fast revenue growth does not settle who carries the asset if a customer renews on worse terms in 2030.

Where a correction would land

AI demand that falls short of the plan, or arrives late, would spread the loss across several groups, which I list below roughly in order of how directly each is exposed.

Investment-grade bondholders come first, and they have already taken a small mark-to-market loss from the repricing of those 91 bonds, though nobody has missed a payment. For the four biggest spenders the rating cushion is wide, but Oracle, at BBB-, is one downgrade away from high yield, and many investment-grade mandates have to sell a bond that falls below that line. I’d bet the Oracle CDS spread is the best real-time gauge of how credit markets feel about AI.

Private credit and the specialist clouds come next. The BIS reports that direct lending funds have roughly quadrupled their lending to AI and IT firms over five years, to about 15% of their portfolios. JPMorgan’s funding model assumed that investment-grade bonds, leveraged finance and securitizations would still leave a gap of about $1.4 trillion for private credit and other sources to fill. CoreWeave is the public example of what that exposure looks like: $35.1 billion of debt at the end of June, and $640 million of net interest expense in the quarter on $2.6 billion of revenue, so about a quarter of its revenue went to interest. Its $103.7 billion of remaining performance obligations is only as good as the labs that signed the contracts. The Federal Reserve’s May Financial Stability Report found that market contacts ranked “risks from artificial intelligence (AI), private credit, and persistent inflation” among their most frequently cited concerns, and in the data center business the first two of those are increasingly the same exposure.

The SPV investors are, on paper, the safest group. Because of the residual value guarantee, if Meta ever walks away from Hyperion the loss goes back to Meta. Moving debt off a balance sheet changes the accounting presentation and the leverage ratios that analysts quote, but the same party still pays in a bad year.

Electricity customers are the group with the least choice. PJM’s capacity auction for the delivery year that starts in June 2028 cleared at its cap of $325 per megawatt-day, for a total of $16.4 billion, and still came in 6,831 megawatts short of the reliability requirement. PJM attributes the higher requirement to forecast load growth “largely due to additional Large Loads.” That price is now fixed whether or not the data centers behind the forecast get built, and much of it will reach households through their bills, although PJM notes that self-supplied and hedged load is not exposed to the auction price. I track the auctions at /power/capacity/.

The last group is everyone, through GDP. With about a third of current growth coming from this investment, I expect that a pause in data center construction would show up in the national accounts within a quarter or two, with only part of it offset by lower imports. The Bank of England’s Financial Policy Committee noted in its September 30 record that “Equity valuations for AI companies fell sharply in July,” cited a JPMorgan estimate of about $4.1 trillion of debt-financed AI capex from 2026 to 2030, and concluded that “The risk of a sharper correction with spillovers to core markets persisted.” On October 7, the IMF’s managing director warned that market disappointment could become “a far-reaching shock”, as Reuters reported, ahead of the IMF and World Bank annual meetings.

Data in the next four weeks that could change my mind

TSMC reports third-quarter results on October 15, and I care less about the quarter, which the monthly figures already show was strong, than about what it says on capacity for 2027. The hyperscalers usually report third-quarter results in the last week of October, with the first real 2027 capex numbers; Alphabet has already said, per MLQ’s summary, that 2027 spending will rise substantially. For each release, I will check whether operating cash flow covered capex in the quarter without new stock or bonds. Census publishes construction spending at the start of each month (the August data came out on October 1), so the September data center figure arrives in early November, and the September trade release is due on November 4.

Clearly positive third-quarter free cash flow at Alphabet, Amazon and Meta, with capex still rising, would make the 2026 squeeze a timing problem and the financing story smaller than I’ve made it. Audited lab revenue on a common basis that looks like the higher of the current estimates would narrow the revenue gap faster than the payback models assume. A wider Oracle spread and a flat Census data center line, with bond issuance still growing, would mean the buildout has outrun the cash flows meant to pay for it, as infrastructure booms usually do, and the groups in the section above will find out how much of the risk they were holding.