AI's debt bill: the ad giants borrow to 2065, and others pay the price

PPC Land is now also on WhatsApp, with a new PPC Land channel carrying its stories.

On Thursday, November 12, 2026, Meta Platforms finished selling €9 billion of bonds in six slices. The longest of them will not be repaid until 2065, and it pays 6.4% a year until then. A week earlier Meta had sold $15 billion of dollar notes at coupons of up to 7.5%, on paper that runs to 2056. A company that built one of the most profitable businesses ever assembled by selling advertising, and that had never issued a bond until August 2022, now owes bondholders roughly $109 billion by PPC Land's count, has sold $37.5 billion of new shares and has prepared the ground for up to $60 billion more. Alphabet, the other half of digital advertising's old duopoly, has raised $84.75 billion of equity, sold a 100-year bond, stopped buying back its own stock and reported two consecutive quarters of negative free cash flow. Amazon has borrowed in dollars, Canadian dollars and Swiss francs. Advertisers are still paying these companies more every quarter, but that money no longer covers what they are spending on artificial intelligence, and the difference is being met with borrowed money and newly printed shares. That raises the question this special edition sets out to answer: is AI creating value, or is it destroying value by raising costs and squeezing revenues and margins across the economy, and who is paying for the experiment?

A PPC Land special edition on the economics and financing of artificial intelligence, published November 14, 2026. It draws on filings, ratings reports, research and news coverage from November 2024 to November 14, 2026, most of it from 2026, with each source dated in the text.

The cash machines start borrowing

For most of the past two decades, the economics of digital advertising were simple to describe and almost impossible to compete with. Google and Facebook built auctions that turned attention into cash at margins that manufacturers could only envy, and they spent that cash on more servers, more engineers and, from the late 2010s, enormous share buybacks. Debt played almost no role. Alphabet carried a token amount for tax and treasury reasons. Meta, as it later became, had none at all until it came to the bond market for the first time in August 2022.

That arrangement is over. The clearest evidence arrived this week, when PPC Land reported on November 14 that Meta borrowed €9bn in euros, at up to 6.4% until 2065. It was the company's first bond in the single currency. The notes were priced under a prospectus supplement dated November 4 and the sale was completed on November 12, a Form 8-K shows. Reuters, citing Bloomberg News, reported a record €53 billion of orders for a deal capped at €9 billion, which PPC Land calculated at about 5.9 times covered. Citigroup and Morgan Stanley ran the books.

The structure says a great deal about how long Meta expects to be paying for its data centres. More than €5 billion of the €9 billion is not due until 2038 or later.

Source: Meta prospectus supplement of November 4, 2026 and Form 8-K, as reported by PPC Land on November 14, 2026.

PPC Land put the weighted average coupon at about 5.32%, the annual interest bill at about €478.75 million and the lifetime coupons at about €7.8 billion, or roughly 87% of the principal. The 2065 notes alone will pay about €96 million a year, some €3.7 billion before the money is returned. Investors paid €8,983,762,500 for the notes, and after €19.5 million of underwriting discounts Meta kept €8,964,262,500 before expenses.

The euro deal followed a dollar sale only days earlier. Under an underwriting agreement of November 3, completed on November 5, Meta sold $15 billion of notes in five parts. Reuters, in a report carried by The Star on November 4, said Meta raised $15 billion in a five-part US bond sale to fund AI expansion, with Moody's assigning an Aa3 rating. The tranches, as PPC Land listed them, were $3.25 billion due 2031 at 5.850%, $3.25 billion due 2033 at 6.250%, $4.00 billion due 2036 at 6.550%, $1.75 billion due 2046 at 7.350% and $2.75 billion due 2056 at 7.500%. The gap between the two markets explains the decision to borrow in Europe: the euro coupons were 1.55 percentage points lower than the dollar coupons in 2031 and 1.60 points lower in 2033.

Taken together, the two November deals add about $1.5 billion a year to Meta's interest bill. That is not a sum that threatens a company of Meta's size. PPC Land calculated it at about 0.6% of annualised third-quarter advertising revenue, which reached $63.06 billion out of total revenue of $64.60 billion. What matters is less the cost of these particular bonds than the trajectory they belong to. Meta's long-term debt stood at $80.92 billion on September 30, 2026, up from $58.74 billion at the end of 2025. Notes outstanding were $84.00 billion, including $25 billion issued in the spring, before the November sales were added. The company also sold a then-record bond of up to $30 billion in October 2025, which Reuters described at the time as the largest bond offering ever.

Borrowing is only half of the financing. On October 30, S&P Global Ratings noted that Meta had already raised $37.5 billion of new equity, made up of $25.0 billion of Class A common stock and $12.5 billion of 6.50% mandatory convertible preferred shares, which trade on Nasdaq as depositary shares under the ticker METAM. A further $60 billion of non-voting Class C stock is to be sold through an at-the-market programme from January 2027. A Reuters factbox published on October 26 put the potential total at up to $95 billion, and Reuters on November 4 said the equity raise could reach $97.5 billion. The new share class exists for a reason that has nothing to do with data centres. As PPC Land reported on November 10, Meta's forecasts show it must raise hundreds of billions as soon as fall 2026, and the special committee that designed the Class C dividend did so in a way that would not dilute Mark Zuckerberg's control of about 60.6% of the vote. Debt, that filing said, had been considered and set aside because of its effect on Meta's credit ratings. The November bonds show that the setting aside was partial.

Why does a company with $63 billion of quarterly advertising revenue need any of this? Because the quarter that produced that revenue also produced negative free cash flow. Meta's operating cash flow in the third quarter was $32.79 billion and its capital spending was $34.91 billion, leaving free cash flow of minus $2.12 billion. Its guidance for 2026 capital expenditure is $135 billion to $145 billion. S&P, which moved Meta's outlook to negative on October 30 while affirming its AA- rating, forecasts more than $225 billion of spending in 2027, as PPC Land reported when it wrote that Meta faces an S&P downgrade as its 2027 AI capex forecast tops $225bn. On S&P's numbers Meta would run a cash shortfall of $71 billion after 2027 capital spending. Meta's chief financial officer, Susan Li, gave cash and marketable securities of $82.4 billion against $83.7 billion of debt at the end of September, which means that for the first time the company owed more than it held.

The most striking figures sit below the line most investors read. Meta disclosed leases that had been signed but not yet started worth about $355.73 billion, and total non-cancelable commitments of $454.52 billion. Those are obligations to pay landlords and developers for data centres over many years, and they do not appear as debt on the balance sheet until the buildings are handed over. Rating agencies count them anyway. In the S&P report covered in this newsletter on November 1, data centre leases were treated as debt-like obligations that could push Meta's leverage above the agency's 1.0 times downgrade threshold as early as 2027.

None of this means Meta is in difficulty. Its advertising business is growing at rates that most companies would envy, and the extra revenue in the third quarter came mainly from price: the average price per ad rose 14% while impressions rose 10%. But it does mean that the advertising machine which once financed everything Meta wanted to do now finances only part of it. The rest is coming from bond investors in New York and Frankfurt, and from shareholders asked to buy stock with no vote attached.

Alphabet sells a century and stops buying back its shares

If Meta's financing is the most aggressive among the advertising platforms, Alphabet's is the largest. The sequence began quietly. In November 2025 Alphabet sold $17.5 billion of bonds in the United States and €6.5 billion in Europe. In February 2026 it returned with a global raise worth $31.51 billion, including £5.5 billion in a five-part sterling deal, one tranche of which was a 100-year bond worth £1 billion, about $1.35 billion. Axios reported on February 17 that Big Tech bets big on bonds, noting that the five largest hyperscalers had issued $45 billion of US bonds in less than two months, nearly half of what they sold in the whole of 2025, and that Alphabet's century bond priced 1.2 percentage points above 10-year UK government debt. Ashok Bhatia, chief investment officer and global head of fixed income at Neuberger Berman, put the question that hangs over the whole cycle: "Are the hyperscalers the economic winners?"

A 100-year bond is a statement of confidence by the borrower and a bet on permanence by the lender. Few technology companies have attempted one. The last comparable sale by a technology company was in 1997, and the companies that lent to the railways of the 19th century on similar terms were not always repaid. For Alphabet, which had about $127 billion of cash and marketable securities at the end of the first quarter of 2026 against $81 billion of debt, the cost was modest and the signal was clear: the company expected to be borrowing for a long time.

Then came the equity. On June 3, 2026, Alphabet announced an expected $85 billion of new shares, and PPC Land reported that Alphabet raises $85 billion to bet everything on AI infrastructure. The package combined a $10 billion private placement with Berkshire Hathaway, an oversubscribed underwritten public offering and a $40 billion at-the-market programme due to begin in the third quarter. The Reuters factbox later gave the final figure as $84.75 billion, upsized from a planned $80 billion. Sundar Pichai, Alphabet's chief executive, said demand for its AI services was "meaningfully exceeding" the supply of computing capacity. The capital expenditure numbers showed what that meant in practice: about $31 billion in 2022, $91 billion in 2025 and guidance of $180 billion to $190 billion for 2026, raised again in July to $195 billion to $205 billion. Anat Ashkenazi, the chief financial officer, told investors to expect a further significant increase in 2027.

The third-quarter results, published on October 28, showed the cost. PPC Land reported that Alphabet's free cash flow dropped to negative $9.1bn despite a $26.2bn profit. Operating cash flow fell 13% to $42.29 billion. Purchases of property and equipment more than doubled, up 114% to $51.36 billion, equal to 121% of operating cash flow and about 40.3% of revenue, against 23.4% a year earlier. Free cash flow came in at minus $9.07 billion, the second negative quarter in a row.

Source: Alphabet quarterly results, compiled by PPC Land, October 29, 2026. Figures rounded.

The way Alphabet has paid for that gap would have been unthinkable two years ago. In the first nine months of 2026, by PPC Land's calculation, it raised $85.22 billion of debt and repaid $8.88 billion, sold $42.87 billion of common stock and $19.06 billion of 6.25% mandatory convertible preferred shares, and raised $3.76 billion by selling interests in consolidated entities. That external funding totalled about $142.0 billion, against $132.0 billion of spending on property and equipment. In other words, Alphabet's entire capital programme for the year so far has been financed from outside the company. Long-term debt rose to $125.10 billion on September 30, from $46.55 billion at the end of 2025. Quarterly interest expense rose from $143 million to $967 million. Alphabet bought back no shares at all in the first nine months of 2026, against $39.81 billion in the same period of 2025, and its share count rose by 206 million. A company that spent a decade shrinking its share base is now expanding it.

There are good reasons to think Alphabet can carry this. It ended September with $265.10 billion of cash and marketable securities, much of it raised from the same offerings. Its cloud division grew revenue 103% to $30.71 billion in the quarter and lifted operating income to $11.12 billion. Advertising revenue grew 10.6% to $82.0 billion. And Alphabet is not only buying computing power, it is selling it: inventory rose from $2.44 billion at the end of 2025 to $13.31 billion, as the company began reporting "product revenues primarily from the sale of TPU systems," its in-house AI chips.

The market's verdict was less generous. Alphabet's shares fell more than 9% in after-hours trading on October 28, as investinglive reported under the headline Alphabet slides over 9% after hours as search miss and $9bn cash burn overshadow Cloud. An analysis by AJ Bell, published on July 16, set the contrast starkly: analysts' consensus forecast free cash flow of $16.8 billion for Alphabet in 2027, against $151.3 billion for Apple, with Alphabet's capital spending forecast at $238.5 billion, about 20 times Apple's.

Amazon, Nvidia and a ledger that keeps lengthening

The third advertising giant has borrowed even more often. Amazon sold about $37 billion in an 11-part bond in March 2026, a record C$14 billion of Canadian-dollar notes in June, the largest corporate bond in that market's history, and planned a six-part Swiss franc deal in May, the Reuters factbox shows. In July it came back with a surprise $25 billion sale in eight parts, with maturities from three to 40 years, during its own earnings blackout period.

That July deal was the first sign that bond investors' appetite had limits. IFR, the capital markets publication, described it under the headline Too much, too soon? US$25bn Amazon return meets lacklustre response. Books of close to $50 billion amounted to less than two times cover, the weakest order book of the current AI debt wave, and Bank of America said new-issue concessions and secondary performance were the worst for a hyperscaler since October 2025. Amazon had printed the equivalent of $91 billion across currencies in four months. "There is a tipping point where if we get flooded with too much of this type of paper, the window will just close," Collin Supple, a credit analyst at PPM America, told IFR. Fortune, in a report on July 17 headlined The AI boom is increasingly built on debt, but investor demand is plunging just as hyperscalers ramp up their bond blitz, put the cover at 2.5 times, down from 3.2 times in March, with Amazon paying 18 to 21 basis points of extra yield on its longest bonds.

Amazon's debt outstanding stood at $162.60 billion against $78.21 billion of cash in the LSEG data compiled for the Reuters factbox. It is a different kind of borrower from Meta, because Amazon Web Services sells computing capacity to outside customers and its data centres earn rent directly. But advertising matters to Amazon's financing more than its accounts suggest. Its advertising services business generated $17.2 billion in the first quarter of 2026, up 24%, and passed $70 billion over twelve months, PPC Land reported on May 12, when it wrote that Big Tech pocketed $150bn in Q1 ads, and AI is changing who wins. Advertising is Amazon's highest-margin business after AWS, and the cash it throws off is part of what lenders are relying on.

Beyond the advertising companies, the list of borrowers reads like a roll call of the AI supply chain. Nvidia, whose chips are the reason for much of the spending, sold $25 billion of bonds in seven tranches in June 2026, maturing as late as 2056, its first debt sale since 2021. SpaceX, which absorbed Elon Musk's xAI and listed its shares in June after an initial public offering that raised about $86 billion, sold a $25 billion debut bond days later. Musk then said on X that SpaceX had "closed" about $40 billion of financing under an agreement with Apollo Global Management, in which an Apollo affiliate buys AI chips and servers and leases them to SpaceX. Salesforce borrowed $25 billion in March, though to fund a share buyback rather than data centres. Oracle, which is examined separately below, expected to raise $45 billion to $50 billion in 2026 through debt and stock.

Source: Reuters factbox with LSEG and SEC data, October 26, 2026. Cash figures are LSEG's and differ from the companies' wider "cash and marketable securities" measures cited elsewhere in this edition.

Adding up the totals depends on what counts as AI. Goldman Sachs Research estimated that nearly $500 billion of AI-related debt had been issued by early August 2026, and that Big Tech had sold about $200 billion of corporate bonds by then, nearly twice its total for 2025, Business Today reported on August 9 under the headline Nearly $500 billion in debt issued in 2026: Is big tech borrowing too much for AI boom? Morgan Stanley put global AI-related debt issuance at about $450 billion in early September, double the 2025 figure. Bloomberg calculations cited by Fortune showed Alphabet, Meta, Amazon and Oracle alone had issued more than $300 billion of bonds since the start of 2025. Bank of America raised its forecast for hyperscaler issuance in 2026 to $175 billion in March, and expected that annual pace to approach $300 billion in the years ahead. Moody's, cited by IFR in August, said hyperscaler debt had doubled to $460 billion since the end of 2024.

Equity has flowed just as fast. By July, IFR counted $90 billion of new stock from Alphabet, $25 billion from Oracle and $86.2 billion from SpaceX in 2026, while the private AI developers OpenAI and Anthropic had sold $122 billion and $65 billion of shares respectively. For comparison, Stefan Hepp of the University of Chicago's Booth School of Business noted in an essay published on August 7 that hyperscaler bond issuance averaged about $28 billion a year from 2020 to 2024 before rising to about $120 billion in 2025. The industry that once financed itself almost entirely from its own profits is now one of the largest users of other people's money in the world.

Advertising still pays most of the bills

It is worth being precise about how much of this the advertising business actually funds, because the answer explains both why the platforms can borrow so cheaply and why they now need to borrow at all.

Advertising is still growing fast at the largest platforms. In the second quarter of 2026, Meta's advertising revenue rose 27% to $59.36 billion, Amazon's advertising services rose 26% to $19.8 billion and Alphabet's Search and other revenue rose 17% to $63.3 billion, PPC Land noted in its August 30 report on data centre opposition. In the first quarter, Meta's ad impressions had risen 19% and its average price per ad 12%, and Microsoft's search and advertising revenue grew 12%. These are growth rates that most advertising markets have not seen in years. WARC raised its forecast for global advertising growth in 2026 to 8.9% in June, and the IAB forecast 9.5% growth in the United States in January.

Capital spending is growing much faster. Meta's capital expenditure in the second quarter was $31.08 billion, up from $17.01 billion a year earlier, an increase of about 83% against advertising growth of 27%. At Alphabet in the third quarter, capital spending rose 114% while advertising revenue rose 10.6%. Microsoft's capital spending rose 70% to $41.0 billion in the June quarter.

Sources: company results as reported by PPC Land on August 30, October 29 and November 14, 2026. Shares calculated for this edition.

Advertising revenue is not cash, of course. It has to pay for staff, content moderation, traffic acquisition and everything else before any of it is available for data centres. Meta's operating cash flow in the third quarter, $32.79 billion, was already smaller than its capital spending of $34.91 billion. Put simply, even if every other cost were covered, the cash generated by the world's second-largest advertising business no longer pays for its AI programme on its own.

The platforms are building new advertising businesses to help close the gap. Google was testing ads inside AI Overviews in 12 markets by May, PPC Land reported, and Lindsay Rowntree of ExchangeWire said search query volume ran about 20% higher because of AI Overviews. OpenAI's ChatGPT advertising passed a $1 billion annual run rate on August 31, 2026. Advertisers, in other words, are being asked to fund not only the platforms' existing AI spending but also the next generation of AI products, through new ad formats inside chatbots and AI answers.

The independents borrow to stand still

This edition set out to start with the ad tech companies, and the picture there is different in kind from the one at Meta and Alphabet. The platforms are borrowing to build. The independent ad tech companies, the ones that sit between advertisers and the open web, are mostly borrowing, refinancing or selling themselves to private equity in order to keep going while artificial intelligence changes where advertising money flows.

The Trade Desk is the clearest case. On April 14, 2026, it signed an amended and restated loan and security agreement with JPMorgan Chase as administrative agent, a Form 8-K filed with the Securities and Exchange Commission shows. The facility is a $750.0 million revolving line of credit, with a $100 million sublimit for letters of credit, a $75 million swingline and an accordion that allows a further $750 million. It matures on April 14, 2031. Interest is set at the Secured Overnight Financing Rate plus 1.125% to 1.500%, with a fee of 0.125% to 0.200% on undrawn amounts. The loans are secured by substantially all of the company's assets, a lien that falls away if The Trade Desk wins investment-grade ratings, and the company must keep funded debt below 3.50 times earnings before interest, tax, depreciation and amortisation. The filing does not say what the money is for, and a revolving facility of this kind is usually held in reserve rather than drawn to fund a building programme.

The context explains why a company that has never needed to borrow would want the reserve. Digiday, in a briefing on August 14 headlined The public ad tech era is over, noted that The Trade Desk's growth had fallen from 22% in the fourth quarter of 2024 to 3% in the second quarter of 2026, which triggered a 22% rout in its shares. In September the company laid off 15% of its staff after disappointing second-quarter results and was removed from the S&P 500, as AdExchanger summarised in its daily roundup of September 8, crediting PPC Land. On November 9, as this newsletter covered on November 10, PPC Land reported that Trade Desk revenue dropped 0.5%, its first decline since its 2016 IPO, with a $44.3 million restructuring charge tied to roughly 575 jobs. The company still generates cash. PPC Land's entry on free cash flow, published October 24, gives The Trade Desk's second-quarter operating cash flow as $153.6 million and its free cash flow as $136.0 million. And the board added $731 million to its share buyback authorisation in November. A company that is buying back its own shares while holding a $750 million credit line in reserve is not printing money through debt. It is insuring itself against a market that has stopped growing for it.

Magnite, the largest independent sell-side platform, has used the credit markets to cut costs rather than to raise money. On October 13, Latham & Watkins said it had advised Magnite on the completion of a term loan repricing. Magnite's $358 million senior secured term loan due 2031 now carries an interest margin 50 basis points lower than before, and the margin on its $175 million revolving facility due 2029 fell by 100 basis points. The savings come to about $1.8 million a year. That is the behaviour of a borrower lenders like: Magnite and PubMatic shares both rose 11% in the second quarter, the Digiday briefing noted, at a time when most of the sector fell.

AppLovin is the one ad tech company that turned AI-driven advertising into an investment-grade balance sheet. In November 2024, IFR reported that AppLovin marked its investment-grade bond debut with $3.55 billion of senior unsecured notes in four parts, priced between 85 and 137.5 basis points over Treasuries, after Fitch and S&P rated it BBB- and Moody's Ba1. The proceeds refinanced about $3.5 billion of term loans due in 2028 and 2030. Since then the company's machine learning engine has kept growing. AdExchanger's roundup of May 11, 2026, citing Eric Seufert of Mobile Dev Memo, noted that AppLovin's advertising revenue grew 59% year on year in the first quarter, and Digiday's August briefing recorded revenue growth of 53% even as the shares fell to a 52-week low. AppLovin's case shows that debt is cheap for an ad tech company when investors believe its AI produces results, and that even then the stock market can lose faith faster than the bond market.

Yahoo shows the opposite end of the spectrum. Apollo funds bought Yahoo from Verizon in 2021, largely with borrowed money. In April 2026, Pulse 2.0 reported, citing Bloomberg, that Yahoo had begun talks with lenders to refinance about $1.6 billion, with RBC leading. The deal closed on Friday, May 8. The same AdExchanger roundup of May 11 summarised Bloomberg's account: Yahoo paid $700 million in leveraged loans at 6.5 percentage points over the US benchmark and $900 million in high-yield bonds due by 2031, in what Bloomberg called "one of the highest-yielding deals of the year." Moody's warned clients that AI "will lead significant changes in the digital media landscape" and flagged Yahoo's "elevated dependence on desktop web traffic." In other words, lenders charged Yahoo a premium precisely because they expect AI to erode the business that services the debt.

Teads, which this newsletter covered on November 7 as dated context here, is further along that road. PPC Land reported that Teads revenue dropped 26% as direct response profit halved in the third quarter, with a $272.4 million goodwill impairment and a net loss of $325.5 million. Chief executive David Kostman blamed AI search summaries and algorithm changes, which he said had cut publisher page views by 20% to 60%. At September 30 the company had a stockholders' deficit of $318.2 million, debt obligations of $657.2 million and cash of $90.3 million, and it was in talks with debtholders. Most of that debt dates from 2025, when Outbrain bought Teads and took its name.

Sources: SEC Form 8-K; Latham & Watkins; AdExchanger and Pulse 2.0 citing Bloomberg; IFR; PPC Land. Dates are those of the transactions or balance sheets.

The rest of the independent sector has gone private, a process this newsletter has covered since September and which appears here only as background. Integral Ad Science agreed to a take-private by Novacap worth about $2 billion, and Nielsen agreed to buy DoubleVerify for $2.15 billion in cash on August 6. Digiday's Ad Tech Briefing of July 14, headlined Private equity's ad tech playbook returns, reported an offer for Criteo from Vista Equity Partners and Quinti Capital at a premium above 50%, valuing its equity at about $3.7 billion, and noted that take-private targets were trading at three to four times earnings before interest, tax, depreciation and amortisation, against 10 to 15 times for private peers. Private equity buyouts are financed largely with debt, so each of these deals moves an ad tech company from the public equity market onto a leveraged balance sheet. SI Global, cited by Digiday in August, estimated that almost a third of private equity investments in the sector were overdue for an exit. "When something is purely commoditized they recognize very quickly that there's not much there," Chance Johnson of Nexxen told Digiday.

There is also a quieter form of borrowing that runs through the whole of ad tech, and it has nothing to do with AI. Ad tech companies pay publishers before agencies and advertisers pay them. Digiday, in a report first published on October 28, 2025, explained that ad tech's economy depends on float and it is getting pricier to keep it moving. Data from OAREX showed that nearly six in ten invoices (58%) were paid late in the first half of 2025, with payment terms of 60, 90 or even 120 days. GumGum pays publishers in 45 to 60 days while waiting 90 to 120 days for agencies. "If nobody was floating money there would be chaos," said Adam Schenkel, an executive vice president at GumGum. Ciaran O'Kane of FirstParty Capital put the competitive effect bluntly: "The float is literally the moat and gives big ad tech a huge competitive advantage." When interest rates are high and growth is slow, the float becomes a financing cost, and the companies with the deepest balance sheets win.

That is the first answer to the question this edition began with. The ad tech companies printing money through debt are, for the most part, not the independents. The independents are refinancing, insuring themselves or selling out. The companies printing money through debt and new shares are the platforms that also own the largest ad businesses: Meta, Alphabet and Amazon. And the independents' troubles are partly a consequence of the platforms' AI spending, because AI summaries, AI-run campaigns and closed buying tools pull traffic and budgets towards those same platforms. Truecaller, the caller identification app, blamed a 44% drop in its first-quarter advertising revenue on algorithm changes by a programmatic partner that an analyst identified as Google, the May 11 AdExchanger roundup noted, citing TechCrunch.

Off the balance sheet: Hyperion, Sopaipilla and a $1.2 trillion lease pile

Not all of the borrowing shows up as debt on the borrowers' books. The largest single AI financing of 2025 was not a Meta bond at all, but a bond issued by a company called Beignet Investor LLC. It raised $27.3 billion to build Hyperion, Meta's data centre campus in Louisiana, through a joint venture in which funds managed by Blue Owl Capital hold most of the equity and Meta holds a minority stake and leases the buildings. The bonds pay a coupon of 6.581% and mature in 2049. PIMCO anchored the deal with about $18 billion and BlackRock took more than $3 billion. S&P rated the bonds A+.

The structure was designed to keep tens of billions of dollars of liabilities off Meta's balance sheet while giving bond investors a claim on a building leased to one of the world's most creditworthy companies. For a time it worked. The bonds were priced at par in October 2025 and rose to about 110 cents on the dollar. Then they fell. Protos reported on September 23, 2026 that Meta's AI bond just hit a record low as its stock soared: the bonds traded at 94.4 cents on the day Meta's shares rose 9% intraday after the launch of its Muse AI agent. At that price, PIMCO's $18 billion position was showing a paper loss of about $1 billion. By October 19, EnterpriseAM reported under the headline Investors turn on the debt behind the AI build-out that the bonds had slid from above par in early July to just under 92 cents.

The divergence between Meta's stock and Meta's off-balance-sheet bonds is one of the more revealing facts of the year. Equity investors were rewarding Meta for its AI products. Bond investors, whose upside is capped at the coupon, were demanding more yield for the risk of lending against long-lived buildings in a market flooded with similar paper.

The sequel made the point more sharply. In July, Capacity reported that Meta's $12bn El Paso bond exposes AI debt repricing. BlackRock was marketing more than $12 billion of bonds through Project Sopaipilla Holdings, an 80/20 joint venture with Meta, to fund a campus of about one gigawatt in El Paso, Texas, due online in 2028. Investors were seeking yields "north of 7%." Matthew Mish of UBS said the pace of AI-related borrowing, which UBS put at about $100 billion a quarter, "would make anyone familiar with credit cycles raise an eyebrow." When the deal priced in early August, IFR ran the headline Bond market begins to question AI boom as Beignet sequel falls flat. Sopaipilla Investor sold $12.547 billion of bonds against orders of only $20 billion, and its June 2048 bonds priced at a yield of 7.50%. By then the Hyperion bonds, which had rallied nine points soon after launch to yield 5.60%, were yielding 7.10%.

Three months later, Meta's own 30-year bond due 2056 was issued with a 7.500% coupon. The company's direct borrowing costs at the long end had climbed to the level that its off-balance-sheet vehicles had paid in August. Part of that reflects the wider rise in long-term interest rates, which the Bank of England noted in September had reached their highest levels since 2008. Part of it reflects the sheer volume of paper the market is being asked to absorb.

Sources: Protos, September 23, 2026; IFR, early August 2026; Capacity, July 24, 2026; EnterpriseAM, October 19, 2026.

The total off-balance-sheet exposure is larger than any single deal. Moody's, cited by IFR in August, calculated that the six largest hyperscalers, which it defined as Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave, had data centre leases worth $1.2 trillion, of which $820 billion had not yet reached their balance sheets because the buildings were still under construction. Nikkei Asia, cited by Business Today on August 9, put the off-balance-sheet liabilities of Alphabet, Microsoft, Amazon, Meta and Oracle at about $1.65 trillion, against about $1.35 trillion that they disclose as liabilities, with Meta alone accounting for about $420 billion. S&P, cited by Stefan Hepp, counted about $675 billion of leases signed but not commenced. Meta's own filing, as noted above, lists $355.73 billion of leases signed but not started.

These numbers measure different things and should not be added together, but they point the same way. A large part of the AI build-out has been financed by developers and investors who build the data centres and lease them back to the technology companies. The technology companies' reported debt therefore understates what they have committed to pay. Rating agencies know this, which is why S&P counts leases as debt when it judges Meta's leverage. But many readers of headline balance sheets do not.

The same logic runs through other structures. SpaceX's arrangement with Apollo, in which an Apollo affiliate buys AI chips and servers and leases them to SpaceX, keeps the hardware off SpaceX's balance sheet. Alphabet raised $3.76 billion in the first nine months of 2026 by selling interests in entities it consolidates. In each case the economic exposure stays with the technology company, but the formal liability, and the first loss if things go wrong, sits with someone else.

Private credit and the hunt for yield

That someone else is increasingly a private credit fund. Stefan Hepp's essay in the Chicago Booth Review, How worried should we be about AI debt?, published on August 7, offered one of the more careful breakdowns of who holds what.

Source: Stefan Hepp, Chicago Booth Review, August 7, 2026. Categories overlap and are not additive.

Hepp also cited the $35 billion financing vehicle that Apollo and Blackstone put together for Anthropic in June 2026. Morgan Stanley, quoted by Axios in February, estimated a $1.5 trillion financing gap between what hyperscalers plan to spend and what they can fund from cash, and has projected that private credit could supply up to $800 billion by 2028, Capacity reported. The private credit industry has spent a decade raising money from pension funds, insurers and wealthy individuals on the promise of higher yields than public bonds. AI data centres, with long leases to large tenants, look like the ideal asset. They are also untested in a downturn.

The first warning came not from a data centre but from a trading firm. PPC Land reported on August 15 that Jane Street lost $15bn as leveraged AI bets collapsed in July, drawing on a Financial Times report of August 14. The loss was disclosed during a refinancing in which about $11 billion of Jane Street's public debt moved to private lenders including PIMCO, as part of a $14.6 billion private debt deal led by JPMorgan. The same article described how Situational Awareness, the fund founded by Leopold Aschenbrenner, received margin calls on July 30 and 31 from Goldman Sachs, JPMorgan and Bank of America on a portfolio leveraged about four times, sold its public equity book to Citadel and lost about 67% in a single month. The July selloff that hurt them was triggered, Fortune reported on July 17, by a cheap model from China's Moonshot, Kimi K3, and ended with Apple overtaking Nvidia as the world's most valuable company.

The International Monetary Fund devoted part of its October 2026 Global Financial Stability Report to the risk. "Weaker or less durable investment and profitability in AI could trigger a correction in AI-related asset prices, exacerbated by rising leverage, concentrated exposures, and circular financing arrangements," its executive summary said. For private credit, it described an AI-led correction as "a test the sector has not yet faced," and called on authorities to "contain leverage in AI-related exposures where excessive." The Bank of England's Financial Policy Committee had said on September 30, as Reuters reported in a story carried by Insurance Journal, that "the likelihood that interconnected vulnerabilities in the financial system crystallize has risen," citing both the re-escalation of conflict involving Iran and AI-related debt issuance. Governor Andrew Bailey nonetheless said that "regulation is not, in my view, the right place to start."

Hepp's own stress test was more reassuring than the language of the official bodies might suggest. In a severe correction he estimated equity losses of $10 trillion to $14 trillion, but credit losses of only about $60 billion to $140 billion, with the first losses falling mostly outside the banking system. That is not a repeat of 2008, when losses sat inside highly leveraged banks. But it does mean the losses would land on the savings of people who never chose to lend to an AI data centre: members of pension plans, holders of insurance policies and investors in bond funds. One of the plaintiffs suing Oracle, discussed below, is a carpenters' pension plan in Ohio.

The neoclouds borrow against the chips themselves

If the hyperscalers sit at the safe end of the AI debt market, the specialist cloud providers known as neoclouds sit at the other. They rent out graphics processors, borrow against those processors and against their customer contracts, and depend on a small number of very large customers.

CoreWeave is the largest. Capacity reported on August 12 that CoreWeave's debt hit $35bn at June 30, 2026, up from $22.7 billion of long-term debt at the end of the first quarter. Net interest expense in the second quarter was $640 million, against $267 million a year earlier. Capital spending for the quarter was $9.4 billion and guidance for 2026 was $35 billion to $39 billion. The company's backlog of contracted revenue stood at $104 billion. Its GPU financing facilities carry loan-to-value ratios of 60% to 70%, and analysts have started to talk about a "neocloud refinancing wall" between 2026 and 2028, when much of that debt comes due. By October, EnterpriseAM reported, CoreWeave's unsecured bonds had fallen below 80 cents on the dollar, with yields approaching 15%. Nvidia, which supplies CoreWeave's chips, agreed in September 2025 to a $6.3 billion backstop for unused CoreWeave capacity through 2032, Hepp noted.

The most detailed picture of how a neocloud is financed arrived on November 14, the day of this edition, when PPC Land reported that Nscale lost $1.02bn in six months on the road to a NYSE listing, working through the Form S-1 that the British-founded company filed on September 18, 2026. Nscale's revenue in the first half of 2026 was $140.6 million, up from $10.4 million. Its cost of revenue, excluding depreciation, was $189.6 million, which means the company spent $49.0 million more running its data centres than it earned from them before counting the cost of the hardware. Depreciation and amortisation came to $174.0 million, the operating loss to $492.0 million and the net loss, after a $457.1 million fair value loss and $95.1 million of net interest, to $1,020.1 million.

What Nscale has instead of profits is contracts. Committed contracts stood at about $103.4 billion at August 31, 2026, of which two customers account for about 85%: statements of work with Microsoft worth up to about $43.8 billion through December 2033, and four agreements signed with Anthropic on August 25, 2026 worth up to about $44.6 billion. Customers have prepaid on a large scale. Deferred revenue rose from $2,038.7 million on January 1 to $6,492.6 million on June 30, and the interest expense of $95.1 million arose mainly from the financing component of those prepayments. In effect, Microsoft and Anthropic are lending Nscale money to build the capacity they will rent.

The rest of the financing comes from lenders, and the terms show how risk is priced at this end of the market.

Source: Nscale Form S-1, filed September 18, 2026, as reported by PPC Land on November 14, 2026. Earlier term loans from Sandton carried rates of 13.0% to 15.0%.

The spread between the cheapest and most expensive money is telling. A revolving facility backed by the whole company costs 1.75 points over the benchmark. A loan secured on GPUs in Texas, where Nvidia guarantees up to $860,340,000 of Nscale's lease obligations, costs 2.375 points. Unguaranteed GPU financing costs 5.00 points, and the earliest loans cost 13% to 15%. Nvidia's support lowers Nscale's cost of capital directly.

The prospectus is candid about the gaps. Nscale's forecasts of its funding needs relied on "uncommitted debt and equity financing," it said, and management concluded that deferring or cutting capital spending would ease any doubt about the company's ability to continue as a going concern. Under the Anthropic agreements, Nscale must use best efforts to obtain financing, and "as of the date of this prospectus, we have not obtained binding commitments" for those financings. Its auditors identified material weaknesses in internal controls in 2024 and 2025. "In only two and a half years, we have scaled from $100 million in total contracted value to over $103 billion," said Josh Payne, the 32-year-old founder and chief executive. "Every layer is a bottleneck."

Nscale's board also shows how closely the AI infrastructure business is tied to the advertising industry's former leaders. Sheryl Sandberg, Meta's former chief operating officer, joined as a director in February 2026, alongside Nick Clegg, Meta's former president of global affairs, and Fidji Simo, who ran the Facebook app before leading Instacart and who joined in September 2026. The company's revenue so far comes entirely from Europe: in the first half of 2026 it earned $73.0 million in Portugal, $63.7 million in Norway, $3.9 million in the United Kingdom and nothing in the United States. Its American future rests on sites still being built, above all the Monarch campus in Mason County, West Virginia, about 2,250 acres acquired in March 2026, with an initial two gigawatts planned by the first half of 2028 and about eight gigawatts by 2031. At June 30, Nscale had outstanding commitments of $24.0 billion for equipment and $3.5 billion for data centre construction.

The neocloud model rests on a timing mismatch that critics return to again and again. Capacity noted in July that data centre bonds run as long as 2049 while GPUs have a useful life of three to five years. The writer Cory Doctorow, whose argument PPC Land reported on September 27, puts it at two to three years. If the chips wear out or become obsolete before the contracts that justify them are paid, the lenders are left holding collateral worth far less than the loans.

Vendors that lend to their customers

The neoclouds' biggest supporter is also their biggest supplier, and that points to the most contested part of AI financing: the circular deals in which chipmakers and cloud providers finance the customers who buy from them.

On July 27, Axios reported that Nvidia reignited circular financing concerns as it weighed an OpenAI deal. Nvidia was considering guaranteeing about $250 billion of obligations for an OpenAI data centre in southern Ohio, and was also weighing financing OpenAI's purchase of $350 billion of Nvidia chips. Nvidia's shares fell 4.5%, and its credit default swaps, the insurance contracts investors buy against a default, posted their largest intraday increase since they began trading the previous November. The project was being developed by a SoftBank unit on a decommissioned uranium enrichment site on federal land, with a $33 billion gas plant pledged by Japan. The Wall Street Journal reported that the guarantee would let the developer "raise debt at more favorable terms than it could if OpenAI had no financial backer."

The final agreement was smaller. Fortune reported on August 18 that Nvidia's data centre deal with OpenAI came in $145 billion lower than reported. In an SEC filing, Nvidia disclosed a guarantee of up to $105 billion of "aggregate payment obligation" for an OpenAI campus in Pike County, Ohio, signed on August 17. Nvidia also invested $1.5 billion in SB Energy. OpenAI's lease runs for up to 20 years on a campus that could reach eight gigawatts, with the first 800 megawatts due in 2028 and 9.2 gigawatts of gas generation planned. Nvidia estimates that each gigawatt requires $50 billion to $60 billion of total spending. Jensen Huang, Nvidia's chief executive, described the aim as "securing long-lived infrastructure for Nvidia compute so OpenAI can deploy the most productive AI factories." Nvidia has also partnered with six financial institutions on compute financing platforms that target more than $500 billion of third-party funding, and Capacity reported that it may backstop up to $125 billion of that. Since July 1, 2026, Nvidia has also offered to rent back unused GPU capacity from cloud providers, starting with a six-year agreement covering 40,000 GPUs at Sharon AI.

In Nscale's case, Nvidia plays every role at once. It paid $399,999,964.32 for Series B preferred shares in October 2025 and $777,399,382.46 in the Series C round in March 2026. It guarantees up to $860,340,000 of the Texas lease obligations. It signed a $1.2 billion master cloud services agreement as a customer in July 2025. It is buying $1.0 billion of the convertible notes. And it supplies the chips. When a supplier is also the investor, the guarantor, the customer and the lender, its own sales figures become harder to read as evidence of independent demand.

The pattern extends beyond Nvidia. Anthropic's commitments to buy compute and infrastructure, as PPC Land reported on November 3 in its coverage of the company's amended IPO filing, total about $895.4 billion, with the Google Cloud commitment rising to about $121.3 billion. Its amended S-1 re-filed a note purchase agreement with Broadcom, its chip supplier. A MediaPost headline on July 31, 2026 reported that Google would guarantee billions of dollars of Anthropic's data centre leases; the article itself could not be retrieved for this edition. UBS analyst Stephen Ju estimated that OpenAI and Anthropic would account for 27% of Google Cloud's revenue in 2026 and more than 48% in 2027, more than $124 billion, with Anthropic alone at $76 billion, PPC Land reported on August 2 when it wrote that Google Cloud faces 48% revenue reliance on OpenAI and Anthropic in 2027. Barclays put the equivalent share for Microsoft's cloud at 13% in 2026 and 18% in 2027, and OpenAI accounted for 69% of the year-on-year growth in Microsoft's Intelligent Cloud segment in 2025. "OpenAI and Anthropic need continual flows of capital. They do not pay their bills out of existing cash flow," the critic Ed Zitron argued in the same article.

The circle has a public dimension too. In November 2025, Sarah Friar, OpenAI's chief financial officer, told a Wall Street Journal technology conference that a federal backstop "can really drop the cost of the financing, but also increase the loan to value." Asked whether she meant a government guarantee, she replied: "Exactly." Quartz reported that OpenAI walked back the remarks about government support for its AI spending spree the same evening. "OpenAI is not seeking a government backstop for our infrastructure commitments," Friar wrote. "I used the word 'backstop' and it muddied the point." At the time, OpenAI had signed about $1.1 trillion of agreements with chipmakers and cloud providers, including a $38 billion deal with Amazon announced that week. Matt Calkins, chief executive of the software company Appian, responded: "I don't think it ought to be too big to fail." The episode lasted a day. The question it raised, whether the public would ultimately stand behind the AI build-out if private financing faltered, has not gone away.

Power contracts: long-term promises of a different kind

The data centres being financed need electricity, and the contracts that supply it are another layer of long-term commitments that rarely appear in tallies of AI debt. PPC Land's Decoded entry on power purchase agreements, published on October 11, sets out the mechanics: a buyer agrees to pay for electricity, or its financial value, from a specific generating project at a price fixed years in advance, typically for 10 to 20 years.

The technology companies have signed these contracts on a scale no other buyers match. Constellation agreed on September 20, 2024 to supply Microsoft with about 835 megawatts from the Crane Clean Energy Center for 20 years from 2028, and on June 3, 2025 to supply Meta with 1,121 megawatts from its Clinton plant, also for 20 years. Meta has contracted 6.5 gigawatts of nuclear power through 2035, Santosh Janardhan of Meta has said, and reports more than 15 gigawatts of new clean energy contracted worldwide. Google's 2026 Environmental Report, published on June 30, counted more than 240 agreements for nearly 35 gigawatts of new clean energy from 2010 to 2025, more than 12 gigawatts of it signed in 2025 alone. "Corporate clean energy buyers are operating at two different speeds," said Nayel Brihi, an analyst at BloombergNEF.

The scale of demand behind those contracts is what alarms grid operators and local residents. Data centres used about 415 terawatt-hours of electricity in 2024, about 1.5% of global consumption, and the International Energy Agency projects about 945 terawatt-hours by 2030, PPC Land's entry on hyperscalers notes. Meta's Prometheus campus is designed for more than one gigawatt, roughly the energy used by 750,000 homes, and its site in Lebanon, Indiana is another gigawatt costing more than $10 billion. OpenAI's Ohio campus plans 9.2 gigawatts of gas generation. Not all of that power is clean, and not all of it is new: a Guardian analysis in September 2024 found that technology companies' 2022 emissions were about 7.62 times higher when measured by the location of the power they used rather than the contracts they had signed, and that Meta's reported 273 tonnes became 3.8 million tonnes on that basis. Location-based accounting "gives an accurate picture of the emissions associated with the energy that's actually being consumed," said Jay Dietrich of the Uptime Institute. In Europe, Directive (EU) 2024/825, which tightens the rules on environmental claims made to consumers, has applied since September 27, 2026.

For the financing question, power contracts matter in two ways. They are fixed obligations that a data centre owner must meet whether or not the building is full, which makes them economically similar to leases. And they pass part of the cost of AI to electricity customers, because new demand on a grid can raise prices for others connected to it, a cost that, as shown below, is disputed but not imaginary.

Oracle, the stress case

If the AI debt cycle has a cautionary tale so far, it is Oracle. The company closed its own advertising business in 2024 and turned itself into a seller of computing capacity, largely to OpenAI. The pivot has given Oracle one of the largest order books in technology and one of the weakest balance sheets among the big cloud providers.

PPC Land reported on July 18 that S&P cut Oracle to BBB- as its AI build-out widened its cash deficit to $42bn. The downgrade, on July 9, 2026, left Oracle one notch above the speculative grade that investors call junk. S&P's reasons were arithmetic. Oracle's guidance for capital spending in its 2027 fiscal year was $90 billion to $95 billion, against the $60 billion S&P had assumed. Free operating cash flow for the year was forecast at a deficit of nearly $42 billion, against an earlier forecast of minus $24 billion. Leverage was expected to reach the mid-4 times range. Oracle had committed to about $260 billion of additional leases for its 2027 to 2029 fiscal years. Its remaining performance obligations, the contracted revenue it has yet to deliver, stood at $638 billion, roughly half of it from OpenAI. Cloud infrastructure, which made up 27% of revenue in fiscal 2026, was expected to reach almost 60% by fiscal 2028. S&P said it had "underestimated the scale of the investments required to expand the AI business." It would cut again if leverage rose above 4.5 times or if Oracle showed no path to positive free operating cash flow by fiscal 2029.

Oracle's bondholders had already gone to court. Reuters reported on January 15, 2026, in a story carried by Insurance Journal, that a proposed class action had been filed the previous day in New York state court in Manhattan, led by the Ohio Carpenters' Pension Plan. The case concerns $18 billion of notes Oracle sold on September 25, 2025, shortly after announcing a contract with OpenAI worth about $300 billion over five years. Seven weeks later, banks arranged $38 billion of loans for data centres in Texas and Wisconsin linked to Oracle. Oracle's borrowings had reached about $108 billion by the end of November 2025. "The bond market's reaction to Oracle's additional debt was swift and bracing," the complaint said. The Reuters factbox of October 26 put Oracle's debt at $120.50 billion against $36.37 billion of cash, with $45 billion to $50 billion of new debt and stock expected in 2026. In February, Oracle sold $5 billion of mandatory convertible preferred shares and planned a further $20 billion of equity during the year.

The market has priced the risk. By July 28, the cost of insuring Oracle's debt against default through credit default swaps had risen about 70 basis points to roughly 215 basis points, the highest among the large AI borrowers, Business Today reported. In October, EnterpriseAM reported that about $18 billion of loans tied to Oracle's Project Jupiter in New Mexico had come under stress after power permitting delays, and that debt linked to an Oracle campus in Michigan had fallen to 87 cents on the dollar from par in July. Quartz, in a comparison of AI debt with the telecom bust published on May 5, 2026, calculated that Oracle would need 7.4 years of operating income to repay its debt and was not expected to return to positive free cash flow until 2029. Its conclusion was that the telecom parallel fits Oracle and the neoclouds far better than it fits the largest hyperscalers.

Oracle matters to this story for a reason beyond its own balance sheet. Its order book is, to a large degree, a bet on OpenAI's ability to keep raising money. OpenAI's ability to keep raising money depends in part on investors believing that Oracle, Microsoft, CoreWeave, Nvidia and others will deliver the computing power it has ordered. Each party's commitments are another party's collateral. That is what the IMF meant by circular financing arrangements, and it is why a problem at one link would travel.

What the bond market is saying

The bond market has not shut its doors to AI borrowers. It has started charging more and asking more questions. The change shows up in four measures.

The first is demand. Torsten Slok, chief economist at Apollo Global Management, tracks how many times over hyperscaler bond deals are covered by orders. Fortune reported on July 17 that the ratio fell from about five times in February 2026 to below two times in July. "Investors are pushing back," Bank of America strategists wrote. Slok said the data suggested "investors may need wider spreads to absorb additional hyperscaler supply." He also told IFR that "AI has been the one thing holding up both the economy and markets," which is both an argument for the spending and a description of how concentrated the risk has become.

The second is price. IFR's August report on the Sopaipilla deal quoted Scott Schulte, global co-head of investment-grade syndicate at Barclays, on how each new deal reprices the last: "There is this perpetual knock-on effect." Shamshad Ali, a macro credit strategist at Goldman Sachs, said: "We anticipate more credit curve steepening for AI-related issuers as this multiyear issuance cycle extends." In plain terms, investors want progressively more to lend for longer, which is why Meta's 2065 euro notes pay 2.1 percentage points more than its 2031 notes. SpaceX's $25 billion debut bond was showing paper losses of $1.8 billion by August.

The third is insurance. Credit default swaps on the large AI borrowers rose sharply in July, Business Today reported on August 9, using data to July 28.

Source: Business Today, August 9, 2026, citing data to July 28.

That Meta's credit default swaps only began trading in October 2025 is itself a sign of the change. A company that did not borrow had no need for a market in insurance against its default.

The fourth measure is secondary prices on the riskiest paper. EnterpriseAM's October report listed the losses: the Hyperion bonds just under 92, two bonds tied to Hut 8 sites at about 94 and 92, Blackstone's QTS Project Excalibur in Georgia below 88, the Oracle Michigan campus debt at 87 and CoreWeave's unsecured bonds below 80. Ludovic Subran, chief investment officer of Allianz, told the Financial Times: "Data center bonds are having their Darwin moment, where the weakest credits struggle to survive."

Not everyone reads these signals as alarming. Iggy Ioppe, chief investment officer at Theo, told IFR: "This is a serious late-cycle signal, even if it is not yet an acute one." Ravi Shukla of Wells Fargo said: "An abrupt funding shutout looks remote." And some investors have simply stayed away. "We've generally avoided investment-grade data centre deals," said Tony Trzcinka of Impax Asset Management.

The pipeline is not shrinking. Estimates have risen all year. In January, Adweek reported that Big Tech's AI spending spree was about to get awkward, citing a Goldman Sachs estimate that Big Tech's capital spending would reach about $527 billion in 2026, up from $465 billion at the start of the third-quarter 2025 earnings season. By October, PPC Land's Decoded entry on hyperscalers put the 2026 plans of the five largest at roughly $800 billion. Moody's expects hyperscaler capital spending to reach almost $1 trillion next year, from $785 billion in 2026, and counts $2.1 trillion of contracted backlog at Amazon, Microsoft, Alphabet and Oracle, IFR reported. Fortune noted forecasts that the five largest hyperscalers could issue $300 billion of bonds a year, against $175 billion in 2026, and JPMorgan analysts projected $375 billion of debt proceeds for SpaceX alone between 2026 and 2030. Bridgewater Associates warned in February that the AI boom had entered a "more dangerous phase."

Borrowing beyond Wall Street

The AI build-out is American, but its financing is increasingly global. Meta's first euro bond was cheaper than its dollar notes by about 1.55 to 1.60 percentage points at comparable maturities, which is why the company went to Frankfurt. Alphabet has borrowed in sterling, Swiss francs and, in May 2026, yen, and its century bond was denominated in sterling. Amazon set a record for the largest corporate bond ever sold in Canadian dollars in June and borrowed in Swiss francs. Each of these markets draws on a different pool of savings, and each extends the reach of AI-related credit risk into financial systems whose regulators did not choose it.

The physical build-out is spreading too. Amazon committed €33.7 billion to data centres in Spain in March 2026, PPC Land reported in August. Nscale's revenue so far comes entirely from Portugal, Norway and the United Kingdom. Japan has pledged a $33 billion gas plant for the OpenAI campus in Ohio. And the Bank of England's Financial Policy Committee named AI-related debt issuance as one of two reasons, alongside the re-escalation of conflict involving Iran, that the risk of vulnerabilities in the financial system crystallising had risen. Its Monetary Policy Committee raised Bank Rate to 4% on November 4, mainly because of energy prices, as this newsletter noted on November 7, which adds to the cost of every floating-rate loan in the chain.

Higher rates are the backdrop to everything in this edition. The Bank of England observed in September that bond yields were at their highest since 2008. Floating-rate loans, such as Nscale's GPU facilities, cost more whenever benchmark rates rise. Fixed-rate bonds issued earlier, such as Hyperion's, lose value when yields go up, which is part of the reason their prices have fallen. And every new issuer, from Meta in euros to Sopaipilla at 7.50%, has to pay the higher price of money that now prevails.

Where AI is creating value

The case that AI is creating value rests on three kinds of evidence: revenue at the companies selling it, falling costs for the people using it, and early signs of higher output in the wider economy. All three exist. None of them yet matches the scale of the spending.

The revenue is real and growing fast. Anthropic, the developer of the Claude models, reported revenue of $16,956 million for the third quarter of 2026, against $1,382 million a year earlier, PPC Land reported on November 3 when it wrote that Anthropic swung to a $567m Q3 operating loss on $5bn of short-term compute, drawing on the company's amended IPO filing. Revenue for the first nine months of 2026 reached $33,242 million. More than 11,900 customers were paying at least $100,000 a year, more than 1,900 at least $1 million and about 250 at least $10 million. Models released in the previous six months produced more than 80% of the quarter's revenue. Google Cloud's revenue grew 103% in the third quarter. Microsoft said in April that its AI business had passed a $37 billion annual run rate, up 123%, PPC Land reported on September 27. Microsoft's Azure cloud grew 43% in the June quarter and passed $100 billion in annual revenue, as this newsletter noted on October 25.

In advertising, the gains are visible in pricing. Meta's average price per ad rose 14% in the third quarter, which reflects both demand and the better conversion rates its AI ranking systems produce. More than 8 million advertisers were using at least one Meta AI creative tool by the first quarter of 2026, up from 4 million at the end of 2024, PPC Land reported on May 12.

Inside the auction itself, AI's value is concrete. PPC Land's entry on inference notes that Meta's Andromeda system narrows tens of millions of candidate ads to a few thousand for each impression, and credits Meta's routing model with a 1.6% rise in conversion rates. PubMatic's Decision Fabric, launched in June 2026, returns decisions in under 10 milliseconds, and a bidder built by Bedrock has run inside Index Exchange since April 2026 within a window of under five milliseconds, in auctions that Index Exchange's chief executive has said last about 100 milliseconds. These are the uses in which the cost of a single AI decision has fallen far enough to run billions of times a day, and they are where advertising's AI returns are easiest to measure. AppLovin's growth is the clearest financial evidence that they can be large.

The cost of using AI has collapsed. The same entry cites Stanford's 2025 AI Index: the cost of querying a model at the level of GPT-3.5 fell from $20 per million tokens in November 2022 to $0.07 by October 2024. Alphabet said its Gemini serving costs fell 78% during 2025, and it was processing 3.2 quadrillion tokens a month by June 2026. The Peterson Institute for International Economics, in an analysis published on February 5, noted that the cost of solving one standard benchmark task fell from $4,500 to $11.64 over 2025. For anyone using these tools, that is an enormous gain in value. It is less obviously a gain for the companies that financed the capacity.

At the level of the economy, the clearest effect so far is the spending itself. Citi estimated that AI investment had accounted for more than half of real US GDP growth in recent quarters, Fortune reported in July. The Federal Reserve Bank of Richmond, in an article published on April 14 titled Will AI investments pay off?, cited estimates that AI investment made up about 92% of GDP growth in the first half of 2025. US labour productivity rose 2.2% in 2025, and Erik Brynjolfsson of Stanford has argued that productivity growth could run at about 2.7% a year, against 1.4% over the past decade. "We're creating the potential to have massive productivity gains and a lot more wealth," he told the Richmond Fed. The Bank of England reached a similar conclusion for Britain. Its November Monetary Policy Report found that AI accounts for nearly a fifth of UK growth in 2026, mainly through firms that build, host and install AI rather than those that use it, PPC Land reported, in a finding this newsletter covered on November 7. Companies told the Bank that AI had raised their productivity by about 0.5% over three years, with almost no effect on employment. But GDP counts spending on data centres as output whether or not the data centres ever earn a return. Building a factory adds to GDP; whether the factory was worth building is a separate question.

The evidence of returns at the companies using AI is much thinner. The Richmond Fed article reported a National Bureau of Economic Research survey of about 6,000 executives in the United States, the United Kingdom, Germany and Australia. Just under 70% of their firms used AI. More than 90% reported no effect on employment, and 89% reported no effect on productivity. Daron Acemoglu of MIT estimates that AI will add less than 0.55% to total factor productivity over ten years. "It is difficult to arrive at very large macroeconomic gains," he said. In a PwC survey cited by Fortune on February 2, 56% of chief executives said they had got "nothing out of" AI so far.

Marketers report the same gap. PPC Land's August 15 report on the Jane Street losses also cited a TransUnion survey, published on August 5, in which 53% of senior US marketing leaders reported meaningful returns on AI, while 89% expected to raise their AI spending, and KPMG research showing that 49% of organisations had cut back AI agent rollouts when costs outran value. A Gartner survey of 1,300 senior marketers, summarised in AdExchanger's September 8 roundup, found that roughly 60% of companies using AI would incur unexpected charges or exceed their budgets for lack of human oversight, and that 56% had deployed AI tools without clear policies. Digiday reported on July 21, under the headline "We're starting to wonder": ad industry chases AI value as usage outpaces proof, that PMG had capped token spending at $50 a day per user and that the agency Cheil had moved half its clients to pricing based on output. "Now it's moving into what actual business impact do we have," said Caroline Giegerich, vice president of AI at the IAB.

The most aggressive version of the sceptical case came from Cory Doctorow. PPC Land reported on September 27 that Cory Doctorow puts AI revenue at $50bn against $1tn of spending. His revenue estimate is not sourced, and the Anthropic and Microsoft figures above suggest it is too low for 2026. But his underlying point is about circularity: revenue of $50 billion a year is easy to generate, he argued, when a trillion dollars is being spent to produce it. "This is greater fool investment theory," he said. The same article cited Ed Zitron's estimate that OpenAI lost $20.9 billion in 2025.

Anthropic's own numbers show both sides of the argument. Revenue grew about twelvefold in a year. But compute and infrastructure costs were $9,863 million in the third quarter, 58% of revenue, and the company swung from an operating profit of $974 million in the second quarter to an operating loss of $567 million. About $5.0 billion of that compute was bought at short notice, and the filing said it "currently costs significantly more than comparable compute procured under our multi-year commitments." The company also gave up about $600 million of revenue in the quarter to price reductions tied to new models and contractual discounts. Adjusted operating margin fell from 19% to 15%. "At our current stage, this cycle is an inherent aspect of our business," the filing said of the need to buy compute ahead of revenue.

The developers themselves are financed mostly with equity rather than debt. IFR counted $122 billion of share sales by OpenAI and $65 billion by Anthropic in 2026 by July. Anthropic's accumulated deficit of $156.2 billion at September 30 included $135.1 billion of fair-value changes on convertible instruments, derivatives and similar securities, PPC Land noted, which means most of its reported losses reflect the rising value of its own securities rather than cash spent. The flows are circular here too. Sales through cloud marketplaces, run by the same companies that supply Anthropic's computing power, made up 44% of its revenue in the first nine months of 2026, and the top five customers accounted for about 25%. Revenue retention among existing customers was about 900%, which means the average customer spending money a year earlier was spending about nine times as much. That is extraordinary growth. It is also growth that depends on customers who are themselves spending heavily on AI.

Where the value is being destroyed

Value destruction is easier to see in some places than in others, and it is not always caused by AI alone. But several groups have lost revenue, margin or market value in ways that the companies involved attribute directly to AI.

Publishers are the most visible. PPC Land reported on November 6 that People Inc. lost 26% of core web sessions as Google AI Overviews spread. Core sessions across the 19 sites owned by the publisher of PEOPLE, Allrecipes and Investopedia fell to 1.61 billion in the third quarter of 2026 from 2.17 billion. The company's earnings presentation put AI Overviews on 82% of the searches it tracks, and its chief financial officer, Tim Quinn, said on the call that the figure was "over 90%." Digital advertising revenue fell 8% to $161.9 million. People Inc. cut 195 jobs, about 6% of its workforce, on October 29. "We've lost nearly 70% of our Google search traffic over the last few years," chief executive Neil Vogel said. "We kind of made up the term Google Zero internally." The same results showed the other side of the trade: licensing and other revenue rose 49% to $52.7 million, and non-session revenue now accounts for 43% of digital revenue. Some of the value AI takes from publishers comes back through licensing deals. For People Inc., not enough of it yet.

The damage is worse for smaller sites. Axios reported on March 17 that Chartbeat data showed small publishers lost 60% of their search traffic over two years, against 47% for medium publishers and 22% for large ones. Google Search page views fell 34% from December 2024 to December 2025. Referrals from ChatGPT rose 200%, but chatbots still accounted for less than 1% of referrals. Even when a site appears inside an AI answer, few people click through. At the end of October, Search Engine Roundtable reported on a small case study by Cyrus Shepard which found that AI Overviews and AI Mode produced an estimated click-through rate of about 0.9%, a figure Shepard himself called "not too shabby," and which this newsletter covered on October 31. Google's own network business, which pays publishers to show Google-sold ads, fell 4% to $6.97 billion in the first quarter of 2026, PPC Land reported in May, and about 90% of Google's ad revenue now comes from its own properties.

Software companies have lost market value at a speed not seen since the financial crisis. The Motley Fool reported on March 27 that software stocks had entered a bear market. The iShares Expanded Tech-Software Sector ETF fell more than 14% after Anthropic released Claude Code earlier in 2026, its worst stretch since 2008. By late March, Microsoft was down 23.3% for the year, Adobe 32.2%, Salesforce 31.3% and Shopify 26.4%. A Jefferies trader coined the term "SaaSpocalypse." "AI is not being abandoned by markets. It is being priced more carefully," said Charu Chanana, chief investment strategist at Saxo Bank. The fear behind the selloff was that AI agents would let customers build or replace the software they now rent. Private credit funds that lent more than $500 billion to software companies by the end of 2025, as Hepp noted, hold that risk too.

AI is also generating new costs for advertisers. AdExchanger reported on October 20 that AI slop sites are on the rise. A study by DecryptAds found 2,116 AI-generated recipe sites monetised with programmatic advertising. One monetisation partner, HBAgency, was listed in the ads.txt files of 1,631 of them, and Ezoic monetised 407, though Ezoic said it rejects more than 98% of the sites that apply. "Inventory that relies on deceptive authorship violates Google's guidance," said Kean Graham, chief executive of MonetizeMore. Every advertising dollar that buys a page written by a machine and read by nobody is value transferred from advertisers to site operators, with ad tech taking a cut.

Households are paying in less direct ways. The Richmond Fed noted that consumer memory prices had risen more than 170% year on year, as chipmakers diverted production to the high-bandwidth memory used in AI servers. Electricity is the larger question. Fortune reported on September 2 that a Georgia Tech study found local electricity prices rose about 5% after a data centre began operating. PPC Land's August 30 report on local opposition that blocked or delayed $156bn of data centre projects in 2025 cited an Environmental Research Letters study projecting wholesale power price increases of as much as 57% by 2030 in the worst-affected regions and 6% to 29% nationally. The same article noted counter-evidence: the consultancy E3 found no clear state-level link between load growth and rate increases, and Lawrence Berkeley National Laboratory found nominal US electricity prices rose 29% from 2019 to 2025, with the PJM grid region as the main exception.

And some investors have already lost money. Hyperion bondholders, SpaceX bondholders, CoreWeave bondholders, Jane Street and Situational Awareness are the examples above. Shareholders of Alphabet and Meta, who until recently received tens of billions a year in buybacks, are now being diluted instead: Alphabet's share count rose by 206 million in nine months.

Margins: who keeps the money

The question of whether AI creates or destroys value is, in the end, a question about who keeps the gains. The evidence so far suggests a pattern familiar from earlier infrastructure booms: users capture much of the value, suppliers of scarce inputs capture much of the profit, and the investors who finance the capacity carry most of the risk.

The price war is the main reason. Axios reported on August 1 that DeepSeek's new bargain model accelerated AI's race to zero. DeepSeek's V4 Flash model produced output for about 28 cents that cost $25 on Anthropic's Opus 4.8, a 99% discount. OpenAI cut the price of GPT-5.6 Luna by 80% three weeks after launch. Sam Altman, OpenAI's chief executive, said: "We will have so much usage of our models that we do not need to be a gigantically high-margin business." The Peterson Institute noted that open-weight models were about 90% cheaper than closed ones but accounted for less than 30% of use, which suggests customers still pay a premium for the leading models. Anthropic's $600 million of quarterly revenue lost to price cuts shows that premium eroding.

Costs, meanwhile, are moving the other way. Alphabet's capital spending reached 40.3% of revenue in the third quarter, against 23.4% a year earlier, and every dollar of it will be depreciated through the income statement over the coming years. Anthropic's compute bill reached 58% of revenue. Nscale's cost of revenue exceeded its revenue before depreciation. CoreWeave's quarterly interest bill more than doubled. At People Inc., the digital unit's adjusted margin fell to about 23.6% in the first nine months, from 27.3% in 2025, by PPC Land's calculation. The Trade Desk guided to fourth-quarter revenue of at least $755 million and adjusted earnings before interest, tax, depreciation and amortisation of at least $280 million, about 11% and 30% below the prior year's $847 million and $400 million. Teads, by contrast, found growth in connected TV, where revenue rose 37% to 12% of its total, a reminder that AI's pressure falls hardest on the businesses tied to search and web page views.

The suppliers of scarce inputs are the exception. Nvidia's chips, high-bandwidth memory, power equipment, electricians and construction workers are all in short supply, and their prices reflect it. So does the cost of capital: the banks underwriting the bonds and the private credit funds lending at 7% or more are among the clearest winners of the cycle. Outside technology, Apollo found no sign of AI lifting margins at the other 493 companies in the S&P 500, which have stayed near 10%, as this newsletter covered on October 26.

Advertisers sit in the middle. They are paying more: Meta's price per ad rose 14% in the third quarter, and a separate PPC Land analysis this week found that B2B Google click prices rose 21% with budget, position and quality held constant. Some of that buys better results, as AI ranking and creative systems raise conversion rates. Some of it, inevitably, helps pay for the platforms' capital programmes. When the companies that sell advertising spend more than their operating cash flow on AI, the advertisers who supply that cash flow are financing the build-out whether they intend to or not.

How a dollar travels

The financing chains described in this edition can be hard to hold in mind at once, so it helps to follow a single dollar. Take, as an illustration, a dollar that an advertiser spends on Meta in the autumn of 2026.

The advertiser is paying more than a year ago: Meta's average price per ad rose 14% in the third quarter. Some of that buys better results, because AI systems choose which ad to show and to whom. Meta's operating cash flow no longer covers its capital spending, so the dollar goes into the AI programme alongside borrowed money: euro bonds due in 2065, dollar notes due in 2056 and, for some campuses, bonds issued by joint ventures such as Beignet Investor, bought by funds that manage pension and insurance money.

The data centre needs chips, and much of the money flows to Nvidia, whose profits are among the clearest gains of the cycle. Nvidia, in turn, uses part of its cash to invest in, lend to and guarantee companies such as Nscale and OpenAI, which buy more Nvidia chips. The data centre needs electricity, contracted for 20 years from a power plant, and its demand on the grid can lift prices for neighbouring households. It needs construction workers and electricians, who are earning more. And it is being built partly with tax breaks granted by a state that may not disclose what they cost.

Back at the other end of the chain, the AI products that the dollar helps to fund change how people find information. Search summaries answer questions that once sent readers to publishers' sites, so People Inc. loses a quarter of its sessions and cuts jobs, and the open web's ad tech companies lose volume. Their lenders, seeing that risk, charge them more, as Yahoo found in May. Meanwhile Meta's shareholders, who once received some of the dollar back through buybacks, are instead offered non-voting shares.

At several points along that path, the dollar creates real value: better advertising, cheaper intelligence for millions of users, higher wages for skilled workers and new revenue for cloud providers. At other points, it leaves behind obligations that will last decades and are spread across pension funds, bondholders, electricity customers and taxpayers. Each of them is, in effect, betting that the chain keeps turning.

Winners and losers

Any scorecard drawn up in the middle of an investment cycle will be wrong in places, and several of the companies below appear on both sides depending on the measure. But the evidence gathered for this edition allows a provisional ledger.

Sources as cited in this edition. Positions are as of November 14, 2026 and describe outcomes so far, not forecasts.

Two patterns stand out. The clearest winners are the suppliers of things that are scarce now: chips, memory, power, skilled labour and capital. The clearest losers are businesses that depended on human attention arriving through search, and investors who lent at fixed rates just before the market began demanding more. Between them sit the hyperscalers themselves, whose AI revenue is growing quickly but whose cash flows are, for the first time, not enough to fund what they are building.

The jobs ledger

The question of AI and jobs has two parts. The first is whether AI tools are replacing workers. The second, which gets less attention, is how the financing of AI is changing employment, because companies spending more than their cash flow on data centres look for savings elsewhere, and payroll is usually the largest cost they control. This newsletter examined AI and entry-level jobs in detail on September 28, so that subject appears here only as context for the financing question.

The headline numbers are smaller than the public debate suggests. Challenger, Gray & Christmas published its October 2026 job cuts report on November 5. US employers announced 43,994 job cuts in October, the lowest October total since 2023, and 617,189 in the first ten months of the year, 44% fewer than the 1,099,500 of the same period in 2025. Employers cited AI for 121,401 of the year's cuts, about 20% of the total, more than double the 54,836 for all of 2025. The monthly peak came in May, with a record 38,579 AI-attributed cuts. In October the figure fell to 1,265, the lowest of the year. Hiring plans announced in October reached 444,692, the highest for the month since at least 2009, led by Amazon's 300,000 seasonal hires. Microsoft's offer of voluntary exits to 60,000 employees was not counted. "This is a low-fire environment," said Andy Challenger of the firm.

Economists who study the labour data are cautious about the AI attribution. HRD America reported, in an article updated on November 5, that AI was blamed for 121,401 job cuts but the labor data isn't buying it. Elena Magrini of Lightcast said: "we are not seeing that in the labor data." Fortune reported on February 2 that the Yale Budget Lab found no evidence of AI displacing jobs at scale. "It just doesn't seem like there's major macroeconomic effects here," said Martha Gimbel, the lab's executive director. Oxford Economics calculated that AI-attributed cuts made up 4.5% of all announced cuts in the first eleven months of 2025.

Goldman Sachs reached a different conclusion. Fortune reported on April 6 that Goldman Sachs economists estimated AI was erasing about 16,000 net US jobs a month, the net of about 25,000 jobs a month lost to substitution and 9,000 gained through augmentation, with younger workers most affected. The New York Fed put unemployment among recent graduates at 5.6% in the fourth quarter of 2025 and their underemployment at 42.5%, the highest since the pandemic, MIT Technology Review reported in an essay published on May 26. But Torsten Slok of Apollo, quoted by Fortune on June 25, argued that "the entry-level squeeze is far more likely a product of the general low-hire, low-fire labor market." Recent graduate unemployment was unchanged from a year earlier, against 4.2% for all workers. Goldman's own work found that greater exposure to AI substitution widened the wage gap between entry-level and experienced workers, and Glassdoor reported that 53% of workplace discussions of AI on its platform were negative, against 41% a year earlier, Fortune noted in June.

The financing link is clearer in individual companies than in national statistics. Amazon has cut more than 30,000 corporate jobs since October 2025, which it attributed to bureaucracy rather than AI, Fortune noted in February. The Trade Desk cut about 575 jobs in September. People Inc. cut 195 on October 29. Cloudflare cut a fifth of its workforce and attributed the decision to agentic AI, AdExchanger's May 11 roundup noted, citing The Information. Block laid off 4,000 people in February 2026, the Richmond Fed noted. In each case, the company's own explanation differs, but the pressure on margins that AI spending and AI competition create is part of the background.

AI is also creating jobs, mostly in physical work. The Information Technology and Innovation Foundation reported on January 12 that the construction industry faces a worker shortage driven by data centres, with about 439,000 more workers needed as of November 2025 and data centre jobs paying up to 30% more than comparable work. Membership of the electricians' union local covering Washington, Maryland and Virginia had doubled since 2018 to 14,700. A study by PwC for the Data Center Coalition, an industry group, credited the sector with supporting 5.5 million US jobs and $927 billion of GDP in 2024, PPC Land noted in August, a figure that comes from an industry-funded study. Permanent jobs at individual sites are few: Google's withdrawn proposal for Franklin Township near Indianapolis promised about 50, PPC Land reported in October 2025. Fortune's September 2 report on data centre tax breaks cited a Georgia Tech study finding that local employment rose about 3.5% and wages about 5% when a data centre opened, gains its authors called "much smaller than what might be expected from a large investment."

Public expectations have moved faster than the data. Pew Research Center reported on August 18 that young adults in the US are increasingly wary of AI and concerned it will take jobs. In a survey of 3,488 adults conducted from June 22 to 28, 71% said AI would lead to fewer US jobs over the next 20 years, up from 64% in 2024, and only 5% expected more. For the first time, a majority of adults under 30, 55%, said they were more concerned than excited about AI in daily life.

Democracy: who decides, who pays, who carries the risk

The democratic questions raised by AI's financing are less about elections, which this newsletter covered in its November 4 special, than about who gets to decide how trillions of dollars are spent and who bears the consequences if the bet goes wrong.

Start with corporate governance. Meta's information statement, covered in this newsletter on November 10, showed that the company is creating non-voting Class C shares specifically so that it can raise capital without reducing Mark Zuckerberg's control of about 60.6% of the vote. Its adviser, Moelis, estimated in June that Meta could raise about $700 billion of equity before his voting power fell below 50%. A special committee wrote that "timing is important to secure capital before competitors saturate various funding options," naming Alphabet, SpaceX, Anthropic and OpenAI. The investors financing one of the largest capital programmes in corporate history will have no vote on how it is spent. That is legal and disclosed, but it means the decision to spend $135 billion to $145 billion this year, and what S&P expects will be more than $225 billion next year, rests in practice with one person.

Then there is public money. Fortune reported on September 2 that big tech is investing millions in data centres and saving on taxes while some states collect revenue losses. Thirty-six of the 50 states offer data centre tax incentives. Good Jobs First found that at least 14 states do not disclose the revenue they lose through those abatements. In Illinois, the number of projects receiving sales tax exemptions rose from six in 2020 to 27 by 2024, with exemptions running up to 20 years. Virginia hosts about 35% of the world's hyperscale facilities. Brookings researchers, quoted in the article, warned that "subsidies may matter more for precisely the facilities that generate the smallest employment benefits." In Indiana, data centres can receive tax breaks lasting up to 50 years, PPC Land reported in October 2025 in its account of how Indianapolis residents beat Google's $1 billion data centre proposal. Google withdrew its rezoning request on September 22, 2025, minutes before a vote. "We are a large community, really in the dark, and it's really concerning," said Andrew Filler, a resident.

Local resistance has become one of the few points at which the public has a direct say. Data Center Watch counted at least $156 billion of projects blocked or delayed by local opposition in 2025, PPC Land reported on August 30. Ian Whittaker of Liberty Sky Advisors, whose LinkedIn post the article was built around, opened with the line "Advertising forecasts may have a data centre problem" and described the force now at work as "the one thing forecasters have never been much good at pricing," which, he wrote, is politics.

The larger projects increasingly involve governments directly. The OpenAI campus in Pike County, Ohio sits on federal land at a former uranium enrichment site, with a $33 billion gas plant pledged by Japan, Axios reported in July. When governments provide the land, the power and the tax breaks, and private investors provide the debt, the line between public and private risk blurs. OpenAI's brief talk of a federal backstop in November 2025 was withdrawn within a day, but it showed that the idea is not unthinkable to the companies involved.

Risk, meanwhile, is spreading to people who never chose to take it. The IMF and the Bank of England have both warned about AI-related leverage, and Hepp's estimate is that most of the first losses in a downturn would fall outside the banks: on private credit funds, insurers and pension plans. The lead plaintiff in the Oracle bondholder suit is a carpenters' pension fund. PIMCO's $18 billion position in the Hyperion bonds is held on behalf of savers. Bailey's view that "regulation is not, in my view, the right place to start" leaves the oversight of that risk largely to the market.

The information ecosystem is also changing. When search traffic to publishers falls by a quarter or more, the revenue that funds reporting falls with it, and People Inc.'s 195 job cuts are one example among many. At the same time, AI is cheapening the production of content of every kind. AdExchanger's report on 2,116 AI-generated recipe sites funded by programmatic advertising is one example. The Washington Examiner reported on June 20 that AI deepfake ads had invaded midterm election campaigns, describing a six-figure buy by the Republican PAC Citizens for Sanity for a 15-second parody ad targeting James Talarico, the Democratic Senate candidate in Texas. "The midterm elections are set to be the first election cycle where Artificial Intelligence deepfakes play a significant role," the article said.

Finally, there is concentration. Meta, Alphabet, Amazon and Microsoft took more than $150 billion of advertising revenue between them in the first quarter of 2026, PPC Land reported in May, and in the United Kingdom two of every three pounds spent on advertising in 2025 went to Google, Meta and Amazon. The financing of AI reinforces that concentration in a way that is easy to miss. Meta can borrow euros at 4.3% for five years. Yahoo pays 6.5 percentage points over the benchmark, Nscale paid up to 15% for its earliest loans, and Teads is negotiating with its creditors. Access to cheap capital has become a competitive advantage in its own right, as O'Kane said of ad tech's float. In a market where the cost of entry is measured in gigawatts, the companies that can borrow most cheaply decide what gets built.

Other approaches

Not every company has answered the AI question the same way, and the alternatives deserve as much attention as the borrowing.

The most prominent abstainer is Apple. AJ Bell asked on July 16 whether Apple's cautious approach to AI could turn out to be a smart move. Apple had net cash of $54 billion at the end of its second fiscal quarter of 2026, analysts expected it to buy back about $108 billion of shares in fiscal 2027, and its forecast free cash flow for that year was nine times Alphabet's. Its shares traded at about 35 times forward earnings, the highest since November 2007, a sign that investors were paying a premium for restraint. When the July selloff hit, Apple overtook Nvidia as the world's most valuable company. The risk of Apple's approach is the mirror image of Meta's: if AI does become the platform through which people use technology, Apple will have to rent or buy capabilities its rivals built.

A second approach is to compete on efficiency rather than scale. DeepSeek's R1 model, released in January 2025, wiped more than half a trillion dollars off Nvidia's market value in a day. The Peterson Institute's Martin Chorzempa argued in February that the boom shrugged off that shock because cheaper models increased demand, a modern version of the Jevons paradox. DeepSeek's V4 Flash and Moonshot's Kimi K3 kept up the pressure this summer. If efficiency keeps winning, the value of the expensive data centres being financed today falls, but the value delivered to users rises.

A third approach is to lean on equity as well as debt. Alphabet raised about $62 billion from common and preferred shares in the first nine months of 2026, alongside $85 billion of new debt, which dilutes its shareholders but keeps its leverage lower than borrowing alone would. Oracle leaned on debt and paid for it with a downgrade. Meta is doing both, and the choice it made, to issue non-voting shares rather than more bonds, was explicitly driven by its credit ratings.

A fourth is to make customers pay in advance. Nscale's deferred revenue of $6.49 billion, and Microsoft's and Anthropic's willingness to prepay for capacity, shift part of the financing to the buyers. Anthropic's $895.4 billion of compute commitments are, from the suppliers' side, a form of customer financing that lets them borrow against signed contracts.

A fifth is to sell the picks and shovels. Google's AI chip business was nearing 100 customers in November, as this newsletter covered on November 10, and Alphabet has begun reporting revenue from selling its TPU systems. Turning a cost centre into a product spreads the expense of the chips across other companies' budgets. "This is the first time we have introduced four TPU systems in a single year," Amin Vahdat wrote in November, and Anthropic's filings estimate additional Google TPU compute at $125 billion to $150 billion for 2028 to 2031. When Google opened Pathways, the system that trains Gemini, to cloud customers in the same week, its announcement conceded the constraint that drives the whole cycle: "When every chip is busy, the next workload waits until capacity frees up."

A sixth approach is discipline on the buying side. PMG's $50-a-day token cap, Cheil's move to output-based pricing and the 49% of organisations that KPMG found had cut back AI agent rollouts all point to buyers who are starting to treat AI as a cost to be managed rather than a capability to be bought at any price. Platforms have pushed the other way: PPC News Feed reported in October that AI Max auto-upgrades were spotted in Google Ad Grants accounts, the free advertising Google gives to non-profit organisations.

A seventh is licensing. People Inc.'s licensing and other revenue rose 49% in the third quarter, to $52.7 million. Licensing does not replace lost traffic, but it is the one mechanism through which some of the value AI takes from publishers flows back.

And there is the low-margin model that Sam Altman described, in which AI becomes a utility that earns modest returns on enormous volumes. That model may be the most realistic of all. It is also the one least compatible with financing that assumes high returns on capital.

The builders' own case

The companies doing the spending make a consistent argument: demand is real, it exceeds supply, and the risk of building too little is greater than the risk of building too much. Sundar Pichai has said demand for Alphabet's AI services is "meaningfully exceeding" its capacity, and has described Google as "supply-constrained." Amin Vahdat, who runs Google's AI infrastructure, has written: "This is what we call the 'age of inference'." Josh Payne of Nscale calls infrastructure "the defining constraint of the age." Meta's special committee, as noted above, framed the financing itself as a race to secure capital before competitors did.

The argument has a logic that investors recognise from other platform markets: the company that controls the capacity sets the terms for everyone else. If Google, Meta or Amazon run short of computing power, customers go elsewhere, and in AI that may mean losing them for good. The cost of overbuilding is a period of low returns. The cost of underbuilding, in the builders' view, is irrelevance.

The ad tech companies make a version of the same argument on a smaller scale. Jeff Green of The Trade Desk conceded in November that the company was "still operating below the level of growth we expect from ourselves," while continuing to invest in AI buying tools. Adam Foroughi of AppLovin has described in-app purchasing as a "$100 billion market" and said: "It's very possible there'll be other types of ad formats." Neil Vogel of People Inc. expects digital growth to return to "mid to high single digits" in 2027, on the strength of licensing and revenue that does not depend on search.

What the builders rarely address is who carries the cost if they are wrong. A company that overbuilds with its own cash hurts its own shareholders. A company that overbuilds with borrowed money, leases and partner guarantees spreads the loss to people who had no say in the decision.

The case that the debt is affordable

The strongest counter-argument to the alarm is simple: the companies doing most of the borrowing can afford it.

Stefan Hepp put it directly: "Let's be clear: The hyperscalers aren't in financial difficulty." He added: "Don't conclude that AI investment is excessive or that a correction is inevitable." His point was not that there is no risk, but that the risk is concentrated in specific places, including neoclouds, project finance and private credit, rather than in the hyperscalers' own balance sheets. "Investors should keep an eye on how much AI-related debt has been issued," he wrote.

The numbers support him. Quartz's May comparison cited M&G's calculation that hyperscaler leverage runs at 0.4 to 0.7 times earnings, against about three times for the investment-grade market as a whole, and CreditSights' finding that the hyperscalers' liabilities are about 48% of their assets, against about 80% for the S&P 500. Alphabet ended September with $265.10 billion of cash and marketable securities. Meta's two November bond deals add about 0.6% of its annualised advertising revenue to its interest bill. Its euro deal drew a record €53 billion of orders, about 5.9 times the amount sold. Demand for the best credits remains deep.

The IMF noted that "core firms in the AI ecosystem have been profitable so far." The revenue is growing: Google Cloud doubled, Anthropic's revenue rose about twelvefold, and Microsoft's AI business passed a $37 billion run rate. Sundar Pichai has repeatedly described Google as "supply-constrained," with demand for computing capacity exceeding what it can build. Ravi Shukla of Wells Fargo considered an abrupt funding shutout "remote."

There is a historical argument too. Infrastructure booms often ruin the people who finance them while leaving behind assets that the rest of the economy uses for decades. Railways built on borrowed money in the 19th century carried goods long after their original investors were wiped out. Fibre optic cable laid in the late 1990s carried the internet traffic of the 2010s. If AI follows that pattern, the value it creates may be real and large, even if a good share of the money lent to build it is never repaid in full.

History rhymes, but does not repeat

The historical comparisons are worth taking seriously because the people running the companies take them seriously. Asked about overbuilding on Microsoft's July 29 earnings call, Satya Nadella said "all of us are reading this 1873 as the book to be read," PPC Land reported, referring to Liaquat Ahamed's history of the railway boom and the panic that followed. This newsletter covered that remark on October 25. Ahamed himself was unconvinced: "I have a feeling that he misread the lessons," he said on The Ringer podcast. On October 17, PPC Land reported that Ray Dalio says his bubble indicator puts AI level with 1929 and 2000, though Dalio did not disclose how his indicator is built.

The closest modern precedent is telecoms. Quartz's May analysis recalled that telecom companies issued more than $500 billion of bonds between 1996 and 2001. WorldCom collapsed with $41 billion of debt, and its bondholders recovered 35.7 cents on the dollar. Global Crossing failed owing $12.4 billion. Recovery rates on defaulted telecom bonds averaged just over 20%, and Moody's speculative-grade default rate peaked at 9.98% in 2001. In July 2002, Michael Powell, then chairman of the Federal Communications Commission, said the sector owed about a trillion dollars, "much of which will never be repaid."

The differences matter as much as the similarities. The telecom borrowers were mostly young companies with heavy debts and thin revenues. Today's largest AI borrowers are among the most profitable companies in the world, with leverage far below the market average. The resemblance is strongest at the edges: Oracle, with 7.4 years of operating income needed to repay its debt; CoreWeave, with bonds yielding close to 15%; Nscale, with a going-concern caveat and two customers making up 85% of its contracts. And the circularity of today's deals, with suppliers financing customers, has its own precedent in the vendor financing that telecom equipment makers extended to carriers in the late 1990s.

"History is rhyming," Chorzempa wrote. Whether the rhyme ends in 1873, 2001 or something milder depends on two things that cannot be known yet: how fast AI revenue grows relative to the cost of the capacity being built, and how long investors keep lending while they wait.

Timeline

What comes next

Several dates are already fixed. The Nvidia portion of Nscale's convertible notes was expected to close around November 16, 2026, and the notes mature on June 15, 2028 if the company has not listed by then. Meta's Class C shares are due to be distributed on January 19, 2027, after which the company can begin selling up to $60 billion of them through its at-the-market programme. S&P has said Meta's leverage could cross its downgrade threshold in 2027. The first 800 megawatts of OpenAI's Ohio campus and the El Paso campus are both due in 2028. Oracle needs to show a path to positive free operating cash flow by fiscal 2029 to avoid a further downgrade. The neocloud refinancing wall runs from 2026 to 2028. People Inc. expects its antitrust case against Google to be resolved in 2027.

The debts themselves run much longer. The Trade Desk's credit facility and Magnite's term loan mature in 2031, as do Yahoo's high-yield bonds. Sopaipilla's bonds are due in 2048 and Hyperion's in 2049. Meta's longest euro notes are due in 2065, and Alphabet's century bond in 2126. The people who will collect the last coupons on those bonds have not been born. Whether the data centres they financed are still producing value by then is the real answer to the question this edition set out to ask.

Did you know

Free cash flow, the measure that turned negative at Alphabet and Meta this year, has no single official definition. The Securities and Exchange Commission's staff says it "does not have a uniform definition" and "must not be presented on a per share basis." By the usual formula, Meta's free cash flow in the quarter to June 30, 2026 was $784 million, after operating cash flow of $31.86 billion, against $8.55 billion a year earlier.

A hyperscaler is a company that runs computing infrastructure at such scale that it can add capacity almost without limit as demand rises. IDC describes a hyperscale data centre as one with at least 5,000 servers and 10,000 square feet of floor space, and Synergy Research counted 1,136 of them at the end of 2024, with Amazon, Microsoft and Google holding 59% of the capacity.

Inference is the act of running a trained model on new input, as opposed to training it, and it is where most of the data centres now being financed will earn their money. Anthropic's Haiku 5.5, released on October 7, 2026, is priced at $0.10 per million input tokens and $0.50 per million output tokens, and Nvidia agreed on December 24, 2025 to license Groq's inference technology in a deal reported at about $20 billion.

A power purchase agreement is a contract in which a buyer agrees to purchase electricity, or its financial value, from a specific generating project at a price set years in advance. BloombergNEF counted 55.9 gigawatts of corporate clean power deals in 2025, down 10%, with four technology buyers taking 49%, and Talen agreed a 17-year arrangement worth $18 billion to supply Amazon with up to 1,920 megawatts.

By the numbers

€53bn
Orders placed for Meta's first euro bond, almost six times the amount it chose to sell.
PPC Land

85%
Share of Nscale's $103.4bn contract book that comes from just two customers, Microsoft and Anthropic.
PPC Land

60.6%
Mark Zuckerberg's share of Meta's voting power, which the new non-voting stock is designed to protect.
PPC Land

$731m
Extra buyback authority approved by The Trade Desk's board in the quarter its sales slipped.
PPC Land