Who carries the AI bill

Written on Friday 6 November and updated on Saturday morning with Friday’s closing prices. Market figures are closing levels for 5 and 6 November unless stated, and the full sources, with the dates and definitions behind each figure, are given in the Growth Under Hard Limits Substack article of the same title.

What the headlines blamed

Democrats won both houses of Congress on 3 November and took the Texas governorship, where much of the new US data-centre building is heading. Ballotpedia counted 41 local data-centre measures on ballots in 10 states, and early results showed voters backing nearly every proposal to halt or limit them. The wider sell-off came on Thursday, when the Nasdaq fell 5.5%, the S&P 500 4.4% and the Philadelphia Semiconductor Index 12%, its largest one-day fall since April 2025. Wall Street recovered some of the loss on Friday, although the semiconductor index still ended about 8% below Wednesday’s close, and the S&P 500 lost 2.7% on the week.

The election may have triggered the sell-off, but AI stocks had been falling for weeks. Goldman Sachs’s broad AI basket fell 15% in October and was more than 24% below its peak before anyone voted. Federal legislation could face a presidential veto, so most of the new policy risk sits in state and local permitting, which can delay a project and does nothing to reduce the debt already raised to build it.

Why some markets fell further than others

By Friday the selling had reached Asia, where Korea’s Kospi fell more than 8% shortly after the open, triggering a 20-minute trading halt, and closed down 6.02%. Japan’s Topix lost 5.9% and Taiwan’s Taiex 4.5% in the same session. India’s Nifty 50 went the other way, rising 2.0% on Friday to finish the week up 8.6%, and Europe’s Stoxx 600 closed down 1.0%.

How far each market fell lines up fairly well with how much technology it holds. Comparing each market’s move on the day the sell-off reached it with the technology weight of a local index fund gives a correlation of minus 0.66 across eight markets, with roughly 0.7% more decline for every extra 10 points of technology weight. Eight markets over two days can show a pattern without proving a cause, and technology weight is only a rough guide to AI exposure, since Taiwan’s and Korea’s technology sectors are mostly chips and hardware while India’s is mostly IT services. Japan is the clearest exception, falling almost as far as Korea with a technology weight of about 25%, while India rose on the largest foreign buying in nearly six years and its technology shares gained because the sector sells IT services rather than hardware.

Who has borrowed

The first difference between markets is who carries the debt, and the largest borrowers in this analysis are American. Microsoft, Alphabet, Amazon and Meta have guided to about $730bn of combined capital spending in 2026, a company-wide figure in which none of them separates out the AI share. With Oracle, the five carried about $551bn of bonds, term loans and other borrowings at their latest quarter-ends, and Goldman Sachs puts their off-balance-sheet commitments at a further $3.2tn, made up of about $1.3tn of leases that have not yet started and $1.9tn of purchase commitments. The hyperscalers can service that debt from existing profits, although the cushion is thinner than it was. Alphabet and Meta each reported negative free cash flow for the September quarter, Amazon’s quarterly figures imply the same, and Alphabet alone raised a net $76bn of debt and $43bn of equity in the first nine months of the year. The hyperscalers are now borrowing to fund the build-out rather than paying for it from cash.

The borrowing extends beyond the hyperscalers, because the frontier labs have signed commitments rather than issued bonds. Anthropic’s amended filing shows about $895bn of commitments, mostly for computing capacity, with $69.8bn due within 12 months. OpenAI said on 6 November 2025 that it was looking at commitments of about $1.4tn over the next eight years, and the two figures are defined differently, so they cannot be added together. The two labs differ in how much they borrow themselves. OpenAI reported no debt at 31 March 2026 and holds an undrawn $4bn credit line, while Anthropic has financed about $35bn of chips through a special-purpose vehicle and was reported in September to be arranging a $15bn credit line. Most of the debt behind their agreements is raised by partners, such as the $18bn of loans financing the Oracle-leased Project Jupiter campus in New Mexico, which syndicate banks quoted at 89 to 91 cents on the dollar in September, and by neoclouds, the specialist firms that buy GPUs, rent them out and borrow against them. CoreWeave reported $35.6bn of total debt at 30 June, with three customers supplying 72% of its quarterly revenue, so a slowdown in orders would reach its lenders before it reached the hyperscalers’ bondholders.

Europe finances far less of its own capacity, and the EU’s €200bn InvestAI programme is a target for the money it hopes to mobilise rather than a committed sum. European savers are exposed all the same, because the ECB found that five US hyperscalers accounted for 15% of the increase in euro-area investors’ holdings of euro-denominated corporate bonds in the year to March, and it has warned that the credit risk may be mispriced.

Who earns the revenue

The largest borrowers in this analysis are American, while the sales their spending pays for run heavily through East Asia. Taking the latest reported quarter for 23 suppliers and multiplying by four gives a rough annual rate of the revenue exposed to AI and data-centre spending. Using reported segment figures and stated estimates, US suppliers account for about $603bn and Taiwanese suppliers about $466bn, led by TSMC’s high-performance computing sales and Foxconn’s cloud and networking business, and East Asian companies earn 43% of the $1,086bn total. A wider count that treats memory, equipment and packaging makers at their whole segment revenue lifts the East Asian share to about 61% of $2,093bn, with Samsung’s chip division and SK Hynix alone reaching a ceiling of about $742bn that should be read as an upper bound.

The same concentration shows in the stock markets, where Samsung and SK Hynix make up about half of the Kospi and TSMC about 44% of the Taiex. The home markets of Taiwan and Korea would register any cut in hyperscaler orders quickly, whichever supplier felt it first.

Rates and the cost of debt

The cost of carrying that debt has risen since the summer. The Federal Reserve raised its target range to 3.75% to 4% in September and held it in October, and the Bank of England raised Bank Rate to 4% on 5 November as energy prices push UK inflation towards 5% by mid-2027. The US 10-year Treasury yield closed at 5.31% on 5 October, a 24-year high, and at 4.86% on 6 November. Higher rates weigh most where the borrowing is largest and newest, which means the United States and the project vehicles and bridge loans funding its build-out.

Why developers went looking for power

Power is now the first question for any data-centre site, and its price varies widely between regions. The IEA puts the average US wholesale power price at about $52/MWh in the first half of 2026, against about $105/MWh in the European Union, and Japanese spot prices reached about $90/MWh in the second quarter. The gap comes mostly from the price of gas in each region. US Henry Hub gas traded at $2.93 per million British thermal units on 22 July, while European TTF reached $20.91 and Asian spot LNG $22.00 as renewed fighting between the United States and Iran disrupted supply.

Cheaper gas and larger blocks of available power have drawn much of the build-out, and its debt, to the United States, and within the country to Texas and the Midwest. A low wholesale price does not make an individual project cheap, however. The PJM grid, which serves much of the eastern US, cleared its 2028/29 capacity auction at the $325/MW-day cap and still fell 6,831MW short of its reliability requirement, with large data-centre loads lifting the forecast, and rising power bills were among the reasons voters turned against data centres in this election. Europe’s higher power costs are one reason it finances so little of its own capacity, and OpenAI cited high energy costs and regulation when it paused its Stargate UK project in April. In East Asia, power costs bear more on the chip suppliers than on the data centres, because Taiwan, South Korea and Japan import most of their gas and the LNG shock feeds into the cost of running fabs and memory plants.

Finding power is only half the task, because the equipment that connects a financed data centre to the grid is taking longer to arrive. Lead times of four to five years have been quoted for some large power transformers and heavy-duty gas turbines take five to seven years from order. Medium-voltage switchgear is running at 78 to 104 weeks, and connection to the grid is a separate queue that is often longer. Currence, an energy research firm, estimated in July that 30% to 50% of the large US data-centre capacity scheduled for 2026 was unlikely to come online this year. Permits are one cause of delay among several, and the others were in place before anyone voted. Once a project draws debt, every month of delay in equipment or grid connection is another month of interest before any revenue arrives, and at today’s rates each of those months costs more, so delay and dearer money compound each other in the cost of the capital invested.

The price of a token

Financing costs rose while customers gained more ways to buy cheaper AI, and these pressures arrive together. Inference is the work of running a trained model to answer requests, and it is sold by the token, roughly a fragment of a word. On OpenRouter, a marketplace that lets developers send requests to many models through one connection, open-weight models accounted for 72% of tokens in September, although closed models still captured 73% of spending, down from 89% in January. Vercel’s AI Gateway reported a similar split, and the average price paid per token there has fallen 47% since December, while Ramp’s index of what US businesses pay shows the effective price per million tokens down 41% from its March peak.

The gateways are only part of the market, and large companies have moved more slowly. Menlo Ventures’ September survey put open-source models at 19% of enterprise token usage, up from 11% a year earlier, while their share of enterprise spending fell from 6% to 4%. The same survey found frontier models more than doubling their share of enterprise tokens, from 18% to 42%, while smaller closed models fell from 29% to 13%, so the squeeze is sharpest on the smaller closed models and enterprise buyers still pay for frontier capability on demanding work.

Anthropic’s third-quarter filing shows that pressure arriving in its accounts: revenue rose to about $17bn as usage grew, yet the filing says price reductions and contractual discounts offset approximately $600m of that increase, and compute and infrastructure expense reached $9.9bn, or 58% of revenue, against 48% in the second quarter. The company sold more AI and collected less per token, while the bill for capacity committed months earlier kept rising. Whether the labs are cutting to defend share is my reading rather than anything they have disclosed, and each new generation of chips lowers the cost of serving a token, which helps the labs’ margins and also lowers the price at which rivals can sell. The financial question is whether volume growth and efficiency gains can cover lower prices, committed capacity and financing costs for borrowers whose debt assumes rapid revenue growth.

Where the strain could travel  

Refinancing is the channel most likely to carry trouble beyond the borrowers, because the same lenders and bond investors may be asked to roll over AI project debt and other private debt, including commercial property loans, at the same time. Morgan Stanley forecasts that private credit will fund about $700bn of data-centre capital spending between 2026 and 2028, and bonds backed by data-centre leases and mortgages have grown from $4bn outstanding in 2020 to about $61bn. Trepp put the delinquency rate on commercial mortgage-backed securities at 7.87% in October, and at 12.33% for office loans, and the default rate in Fitch’s monitored US private credit portfolio reached a record 6.3% in the year to August. In its October stability report the IMF found that banks holding 22% of global banking assets would breach capital limits in its adverse scenario, a broad stress test not specific to AI, and noted that their links to non-bank lenders are deepening.

What boards should watch

The midterms may slow some projects, yet many organisations hold more exposure to an AI slowdown than their technology budgets suggest. Their credit portfolios, insurers and pension funds hold it indirectly, through hyperscaler bonds, private credit funds, data-centre securitisations and Asian chip equities. Boards should ask their treasury and pension advisers where the same borrowers and lenders appear across portfolios and supplier contracts, and then test what happens if new capacity opens a year late while refinancing costs stay high, and how much worse that becomes if inference prices also fall faster than expected.

My own view is that AI will deliver extraordinary benefits, and I doubt they will accrue solely to closed, centralised models. Distillation spreads intelligence quickly from frontier systems into open models, and open models run on distributed infrastructure, which also eases the power limits of giant data centres. What worries me is the debt underneath the build-out, and the question of who is left holding it if the revenue arrives later, and at lower prices, than the borrowing assumed.

illuminem Voices is a democratic space presenting the opinions of leading Sustainability Thought Leaders, their views do not necessarily represent those of illuminem.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.