AI, Work and the Economy

Latest/Story

Lending for data centres

Fifteen items, from 26 Sep 2026 to 18 Nov 2026, oldest first.

See also: Point 4 of our view: Scarce capital, power and sites · Point 5 of our view: A downturn about the size of 2001's

Policy

Japan's FSA (25 Sep) steps up review of banks' and life insurers' AI data-centre loans, mainly to US projects

Japan's Financial Services Agency is intensifying its review of lending to AI data centres as major banks and life insurers raise exposure, focusing on risk management for US projects; it says it is not trying to choke off funding. It follows Oracle's force-majeure notice on the Jupiter campus (24 Sep), which Reuters reports has unsettled AI infrastructure lenders.

Why it matters. Japan's bank regulator is checking how exposed its big banks and insurers are to loans for US AI data centres, a sign lenders' risk from the AI building boom is getting official attention.

Next. Whether other supervisors, including the ECB, follow with findings or limits on data-centre lending.

Sources: japantimes.co.jp, usnews.com

PolicyReportUpdate

Riksbank's Thedéen (26 Sep) warns loan-financed AI build-out can't be paid from earnings; expects tighter fiscal policy

In Ekot's Saturday interview (published 26-28 Sep), Riksbank governor Erik Thedéen said the data-centre expansion cannot be financed from earnings alone, that some AI firms could face chaotic falls, that Sweden's labour market remains weak and inflation lingers, and that he expects the incoming government to tighten fiscal policy. Lagarde told the European Parliament on 28 Sep that a sharp reassessment of AI firms' prospects and debt could trigger market corrections.

Why it matters. Sweden's central bank chief and the ECB's president both warned this week that the borrowing behind the AI data-centre boom could end badly, while Sweden's jobs market stays weak and a tighter budget is expected.

Next. The new Swedish government's budget, and the Riksbank's November rate decision.

Sources: placera.se, ecb.europa.eu

Economy and marketsPolicyUpdate

Bank of England FPC (30 Sep): AI debt issuance ~$450bn by early Sep, over twice 2025's; sharper correction risk

The Financial Policy Committee said AI-related debt issuance reached about $450bn by early September, more than double all of 2025 (Morgan Stanley), cited JP Morgan's estimate of $4.1tn of debt-financed AI capex in 2026-30, and warned that a correction worse than July's could spill into sovereign debt markets. It flagged leverage, opacity and 'circular arrangements'.

Why it matters. Britain's central bank says borrowing to build AI data centres has more than doubled this year and that a sharper market fall could spread to government bond markets.

Next. The IMF's Global Financial Stability Report in mid-October will give its own reading of AI-related debt.

Sources: tribuneindia.com, news.google.com

AI and technologyReportUpdate

AI chips go off balance sheet: Broadcom to lend Anthropic up to $42bn (filing, 1 Oct); Amazon seeks $8bn chip vehicle

Anthropic's IPO filing shows Broadcom will lend it up to $42bn, convertible into Anthropic shares, toward a $125.2bn five-year lease of chip capacity; the FT reports Amazon is sounding investors on moving about $8bn of Nvidia Grace Blackwell chips into a debt-funded special-purpose vehicle (1-2 Oct, not agreed).

Why it matters. The AI build-out is increasingly paid for with borrowed money arranged by chip sellers and moved off company books, which spreads the risk to lenders if AI revenue disappoints.

Next. Anthropic's listing and the marketing of the Broadcom-arranged debt in the coming weeks.

Sources: semafor.com, bloomberg.com

AI and technologyReportUpdate

NBER (w35865, Van Nieuwerburgh): US AI build-out of 188 GW by 2032 needs ~$9tn, 3.2% of GDP a year

'Financing the AI Buildout' costs a 1 GW campus at about $41bn and finds that as hyperscalers exhaust internal cash, funding shifts to leases, joint ventures, project debt, private credit, securitisation and special-purpose vehicles, with leverage at asset level, hidden contingent obligations, tenant concentration and obsolescence risk.

Why it matters. A new paper puts the US AI data-centre build-out at about $9 trillion by 2032 and shows it is increasingly funded with borrowed money held outside the big tech firms' balance sheets, which is where losses would surface if it turns.

Next. Big tech's third-quarter results in late October, with their spending plans and how they fund them.

Sources: nber.org

Economy and marketsNew dataUpdate

SpaceX seeks ~$40bn ($10bn bank loans, $30bn bonds), led by Apollo, to buy Nvidia chips (FT, 7 Oct)

SpaceX is in talks with Apollo, Pimco and banks on about $40bn of financing for Nvidia GPUs, roughly $10bn of bank loans and $30bn of investment-grade debt, with a close expected in 2027 (FT, CNBC, 7 Oct). It would about double SpaceX's debt of $39.7bn at June; its Q2 capex was $18.4bn against $2.9bn a year earlier.

Why it matters. One company is seeking about $40 billion of new borrowing, most of it in bonds, to buy AI chips, at a time when long-term interest rates are near their highest in more than twenty years. The more of the AI build-out rests on debt, the more any fall in AI demand or in the chips' value might pass to lenders and, through them, to the banks that fund them.

Next. Pricing of the bonds and the lenders' terms, and US credit spreads in the weeks ahead.

Sources: finance.yahoo.com, cnbc.com

Economy and marketsMarket moveUpdate

JPMorgan markets $5bn loan for Volta's Norway AI campus at ~11%, a record yield for such deals; commitments due 14 Oct

Volta Infrastructure, founded this year, seeks a $5bn leveraged loan split $3.7bn/$1.3bn at 6.25-6.5 points over the benchmark and 97-98 cents, about 11% all-in, for GPUs and a data-centre complex in Norway run with Bitdeer, with Anthropic reported as the six-year anchor customer.

Why it matters. A company founded this year is borrowing $5bn at about 11% to fill a Norwegian data centre with AI chips, the highest rate yet paid for such a loan. Lenders demanding this much suggests rising doubt about these projects; if similar deals struggle to sell, the build-out's financing could tighten quickly.

Next. Whether the loan is fully placed after commitments close on 14 October.

Sources: thestar.com.my · Point 4 of our view · Context: not counted as evidence

Economy and marketsNew dataUpdate

AI-related debt issuance $23bn in Sep, under half of Aug and a fifth of June's $113bn record (FT/Morgan Stanley, 12 Oct)

Global AI-related bond and private-placement issuance fell to about $23bn in September, less than half of August's total and far below June's record $113bn, on Morgan Stanley data reported by the FT. About $18bn of loans tied to Oracle's Project Jupiter data centre in New Mexico were recently quoted privately near 85 cents on the dollar.

Why it matters. Borrowing by AI data-centre builders, which reached a record $113bn in June, fell to $23bn in September, and loans for one large Oracle project trade well below face value. If lenders stay away while long-term rates are high, some planned projects could stall, and losses on loans already made could spread to the banks and funds that hold them.

Next. October's issuance figures, and whether the large deals planned by SpaceX and Broadcom are completed.

Sources: privateequitywire.co.uk · Point 5 of our view · Context: not counted as evidence

Economy and marketsMarket moveUpdate

JPMorgan and Goldman underwrite only $1.3bn of Volta's $5bn AI loan; the $3.7bn GPU loan is not underwritten (12 Oct)

Bloomberg reports the two banks committed to the $1.3bn term loan but left the $3.7bn loan for about 36,000 Nvidia chips to be sold without a bank underwriting guarantee, shielding themselves from losses; the package is marketed at an all-in yield of about 11.2-11.75%, with commitments due 14 Oct.

Why it matters. Two of the largest US banks agreed to guarantee only the smaller part of a $5bn loan for a Norwegian AI data centre, leaving the $3.7bn borrowed against AI chips for other investors to take. If banks keep limiting their exposure this way, financing for the AI build-out could become harder and dearer to find, and losses would fall on funds rather than banks.

Next. Whether the loan is fully sold after commitments close on 14 October.

Sources: bloomberg.com, thenextweb.com · Point 5 of our view · Context: not counted as evidence

Economy and marketsMarket moveUpdate

Bloomberg (28 Oct): at least 18 data-centre loans and junk bonds yield double digits; top-rated data-centre bonds ~7% from under 5%

Bloomberg reports that buyers of AI infrastructure debt are nearing their limits: at least 18 data-centre financings in the junk-bond and leveraged-loan markets carry double-digit yields, the highest-rated data-centre bonds yield close to 7% against under 5% in January (Morgan Stanley), and the Bloomberg investment-grade technology index yields about 6%, 1.4 points above a year ago. Oaktree calls itself 'very, very cautious', Carlyle has passed on deals, and a Willis Towers Watson survey finds insurers near their data-centre limits.

Why it matters. Lenders and bond buyers are asking for much higher returns to fund AI data centres, and several large investors say they are pulling back. If this caution spreads, the build-out might slow for lack of finance rather than of demand, which is how past investment booms have turned.

Next. The Federal Reserve's survey of bank lending officers in early November will show whether banks are tightening too.

Sources: ca.headtopics.com, bloomberg.com · Point 5 of our view · Context: not counted as evidence

Economy and marketsNew dataUpdate

Fed stability report (13 Nov): AI-related debt issuance over $680bn by October 2026, from $130bn in all of 2023

The Federal Reserve's November Financial Stability Report puts debt issued for advanced computing at more than $680bn by October 2026, about 31% of investment-grade non-financial issuance (under 6% in 2023), with more than $1.1tn outstanding in public markets. The largest tech firms' leases not yet started exceed $1tn and their guarantees and take-or-pay contracts $2tn, undiscounted; the largest US bank holding companies have close to $1tn of exposure to tech firms, data-centre developers, their special-purpose vehicles, private credit and utilities.

Why it matters. The Federal Reserve finds that borrowing to build AI computing has grown about fivefold since 2023 and now makes up close to a third of high-grade corporate bond sales, with a further $3 trillion or so of leases and guarantees outside the balance sheets. If AI revenue falls short of what this debt assumes, losses could reach pension funds, insurers, private lenders and banks well beyond the firms themselves.

Next. Nvidia reports its quarterly results in the coming week; the Fed's survey of bank lending officers follows in early February.

Sources: federalreserve.gov, cnbc.com · Point 4 of our view

Economy and marketsNew dataUpdate

NBER (w36032, Van Nieuwerburgh & Su, Nov 2026): debt funds ~60% of the AI compute build-out; $16.8tn at risk in a bad state

A new NBER paper maps the financial network behind AI compute. Equity cushions protect creditors against obligations outstanding in mid-2026, but debt funds nearly 60% of the planned expansion, non-recourse project lenders stay exposed, and in a low-productivity state shareholders and creditors lose $16.8 trillion, with 40% loss-given-default on debt; payment links between key firms account for about 30% of private losses.

Why it matters. Economists at Columbia map who owes what in the AI build-out and find that debt pays for nearly 60% of it. Should AI's returns disappoint, they estimate losses of $16.8 trillion for shareholders and lenders, and much of the harm would travel through payment links between a few firms, which is how a setback in technology might become a credit problem.

Next. Nvidia's quarterly results this week, and the builders' next debt sales, show whether lenders are still willing.

Sources: nber.org · Point 4 of our view

Economy and marketsReportUpdate

Fed Governor Barr (16 Nov): top AI firms' debt to pass $800bn by year-end; off-balance-sheet commitments over $3tn

In a speech in Milan, Fed Governor Michael Barr said on-balance-sheet debt of the seven largest AI firms is set to exceed $800bn by end-2026 and their leases and other commitments exceed $3tn, likening it to the late-1990s telecom boom; a sharp reversal could bring 'self-reinforcing fire sales'. He added that AI agents could speed deposit outflows into runs that 'unfold in minutes', and that weaker firms dependent on outside funding face a rising cost of capital.

Why it matters. A Federal Reserve governor puts the largest AI firms' debt above $800 billion this year, with over $3 trillion more in leases and commitments, and warns that a sharp reversal could force fire sales. He also says AI agents moving money could turn bank runs into events of minutes. If both hold, a turn in AI investment might travel through credit markets faster than past busts.

Next. The Fed's December policy meeting and any guidance on agent access to bank accounts.

Sources: federalreserve.gov, news.google.com · Point 4 of our view · Context: not counted as evidence

Economy and marketsReportUpdate

ECB stability review (18 Nov): most US hyperscalers' free cash flow negative in 2026; off-balance-sheet commitments over $4tn

A box in the ECB's November Financial Stability Review finds AI capex exceeded operating cash flow at most US hyperscalers in 2026, their off-balance-sheet commitments passed $4 trillion at Q2 2026, well above their on-balance-sheet liabilities, and their bond spreads have widened against equally rated firms. It projects big-tech capex of about $5.5 trillion cumulatively by 2030 and draws parallels with the dot-com and subprime booms.

Why it matters. The European Central Bank finds that most of the largest US cloud builders spent more than their operating cash this year and hold over $4 trillion of commitments off their balance sheets. If the build-out keeps shifting to debt, a disappointment in AI revenue could perhaps travel to lenders and bond investors in the way the mortgage boom did, rather than staying in share prices as the dot-com bust largely did.

Next. The Federal Reserve's next survey of bank lending standards, in early 2027, and the cloud builders' fourth-quarter filings.

Sources: ecb.europa.eu, ecb.europa.eu · Point 4 of our view

Economy and marketsReportUpdate

Goldman (FT, 17 Nov) lifts 2027 AI project-finance forecast to $545bn from $300bn; chip loans $320bn vs $48bn in 2026

Goldman Sachs' credit strategists now expect $545bn of AI project finance in 2027, about $225bn for data-centre shells and $320bn for chips, up from a $300bn forecast; chip financing so far in 2026 is $48bn. A large share of the chip debt may come without corporate guarantees.

Why it matters. Goldman Sachs now expects about $545 billion of project loans for AI in 2027, with chip purchases financed by debt rising almost sevenfold. Chips lose value within a few years, so a borrowing boom secured on them could perhaps leave lenders with large losses if AI revenue falls short of what the loans assume.

Next. Whether lenders take the first large chip-backed loans at stable prices over the coming weeks.

Sources: cryptobriefing.com, news.futunn.com · Point 4 of our view