With apologies to Sergio Leone and Clint Eastwood.
Figures current to 7 August 2026, including the July employment report released that morning.
This paper is general information only. It is not financial or investment advice. It does not take account of any person’s objectives, financial situation or needs, and should not be relied on in making an investment decision. The author is not a licensed financial adviser. Figures are drawn from public sources, are current only to the date above, and may contain errors.
Introduction
As we approach the frothy end of the business cycle, there are growing voices predicting a crash, some within months, others within one to three years. Some see a shallow recession like the dot-com crash of 2000, while others see the deep trauma of another 2008 Great Financial Crisis. The purpose of this paper is to allay fears of an immediate catastrophic crash. The US economy is currently healthy, and the danger signs of 2007-08 do not appear to be present.
That last claim carries one significant qualification, and it is why this paper has three parts rather than two. The risk that most resembles 2008 has not disappeared. It has moved out of the banks and into private credit, where nobody can price it.
This paper looks only at the United States, because that is where the AI buildout and its risks sit. The five hyperscalers, the chip designers, the bond market funding the whole thing and the private credit funds and insurers increasingly financing it are American, and the electricity constraint binds in Virginia, Ohio and Texas. The frontier labs are mostly but not all American. For everyone else, including Australia, the exposure is second-hand and runs through four channels: equity markets, where the top ten US names are 41% of the S&P 500 and sit inside almost every global index fund and superannuation balanced option; commodity and capital goods demand; the price of AI services themselves; and the overflow of the buildout itself.
In short, the argument runs as follows. The strength of the US economy makes a deep, GFC-style recession unlikely (the good). The arithmetic of the AI buildout makes disappointing returns on much of that capital investment likely (the bad). And the private credit funds and insurers increasingly funding the buildout are the ones who will absorb those disappointing returns, and who will at some point stop financing the next project (the unknowable). That withdrawal of finance, rather than any single spectacular failure, is the most plausible route from an investment disappointment to a slower economy, and at some point, more likely than not, to a shallow recession. President Trump's trade and immigration policy and the war in Iran bear on the same question and are outside this paper’s scope.
The question, then, is not simply whether AI is overbuilt. It is whether the excess works itself out gradually, through poor returns and slower investment, or whether the forces are large enough to produce a deep recession and a seismic correction in markets.
We begin with the bull case on the US economy.
I. The Good: The bull case on the economy
Growth and productivity are both real. Gross domestic product (GDP) grew 2.1% in 2025, private domestic demand ran +3.9% annualised in Q2 2026, and productivity has grown 2.1% annualised since Q4 2019 against 1.5% in the previous cycle. Caveat: some of this may be composition rather than genuine gain. If the workers leaving the labour force under current immigration policy have below-average output per hour, measured productivity rises without anything improving. Participation has fallen 0.7 points since January, to 61.4%. No primary source I could find separates the two effects.
Earnings growth is broad, not just mega-cap. Q2 2026 came in at +47.4%, still +28.8% excluding Alphabet and Amazon, the seventh consecutive double-digit quarter, on a record 16.7% net margin.
Forward valuations are not 2000. The forward price-to-earnings ratio (P/E) is 19.6, below its own five-year average. March 2000 was nearer 25x.
Households are in the best shape in a generation. Debt-to-GDP sits at 25-year lows, debt service at 11.16% of disposable income against 15.85% at the Q4 2007 peak, net worth at \$183tn, and credit card delinquency is falling. Caveat: in a K-shaped economy aggregates hide the rich getting richer and the poor getting poorer. The top 10% of households hold about 87% of all shares, so the net worth figure describes a minority. Real hourly pay fell 3.1% annualised in Q2 2026 and labour’s share of output in the nonfarm business sector, on the BLS measure, is 52.9%, the lowest in a series that begins in the first quarter of 1947. The saving rate is 2.7%, a thin buffer. Car loan arrears sit at the 90th percentile of their historical range, which is where strain in lower-income households appears first. The balance sheet strength is real, but it is concentrated in the households least likely to cut spending when markets fall, which is also why the effect in point 14 is small.
Corporate leverage outside AI is low. Business debt-to-GDP is back at early-2000s levels and non-AI borrowing is running below its fifteen-year average.
Banks are heavily capitalised and barely exposed. All 32 passed the 2026 stress test absorbing \$708bn of modelled losses. Lending to AI-adjacent borrowers is 9% of tier 1 capital, the core shareholder funds regulators require banks to hold, and lending against data centre property was \$14.9bn.
Credit markets show no broad stress. Lenders are charging very little extra to lend to riskier borrowers. The premium over government debt is 78 basis points for investment grade and 275 for high yield, both near the lowest levels since 1997. Widening premiums are the classic early warning of credit trouble, and they are absent.
Unemployment is low, but the labour force is shrinking. The jobless rate is 4.1% and the long-term unemployed share improved to 25.5%. Participation, however, has fallen 0.7 points since January to 61.4%, and the foreign-born labour force is down about a million from its March 2025 peak. So the rate is low partly because the denominator is smaller, not because the market is strong.
Layoffs are historically low even as hiring stalls. The layoff rate is 1.1% and announced cuts are down 41% year to date. Firms are not shedding staff; they have stopped adding them. Note: the July report, released on 7 August, showed payrolls falling 23,000 against an expected gain of about 80,000, with May and June revised down by a combined 103,000. Unemployment nevertheless fell to 4.1%, because participation fell too. See II below.
AI capex is still majority equity-funded, though less than it looks. Debt covers 32% of capex, up from 9% two years ago, and that figure excludes about \$821bn of leases not yet started and any debt held in special purpose vehicles, which are separate project companies set up to hold a development and its borrowings outside the parent’s headline debt figure.
AI revenue is real and growing fast. Anthropic disclosed a \$47bn run-rate, meaning current monthly revenue multiplied out to a full year, Microsoft \$37bn, AWS passed \$25bn, and Google Cloud grew 82% on a \$514bn backlog.
AI models are still getting better, and faster than before. Epoch AI, an independent research group, tracks the performance of the leading models on a combined index of benchmark tests. That index has risen about 15.5 points a year since early 2024, against roughly 8 a year before, so the rate of improvement roughly doubled. Whatever is wrong with the investment case, the technology itself is not stalling.
Data centres are tight, not glutted. Vacancy is 1.6% and 74% of capacity under construction is pre-leased.
A falling share market does not by itself cause a recession. When shares rise people spend a little more, and when they fall they spend a little less. The response is small: about one cent of changed spending for every dollar of share wealth, against roughly five cents for housing. So a 30% fall in the Magnificent Seven, the seven giant US technology stocks that dominate the index, would mechanically knock roughly 10% off the S&P 500 if everything else were unchanged, costing somewhere between 0.2% and 0.5% of GDP in lost consumption. Painful for portfolios, not on its own a recession.
The relevant precedent is mild. The 2001 recession ran eight months and cost 0.6% of GDP, despite telecom losing \$700bn of value and 22% of its jobs.
Non-AI supports exist. Record oil and gas output, a deficit running 5.8% of GDP, and retail sales up 6.4% year-over-year.
Inflation is the one real pressure point. It is not disastrous, but it is uncomfortable, with core inflation on the Fed’s preferred measure, personal consumption expenditures, at 3.3% against a 2% target. The overshoot is concentrated in energy; core goods are running at 0.8%, and the Fed's own projections have it falling to 2.3% next year. So it constrains policy now rather than permanently.
II. The Bad: The bear case on investment returns
A. The revenue arithmetic
The spending is running far ahead of the revenue. For every \$1 an end customer spends on AI at the moment, the AI complex is spending about \$5. That \$1 is net of double counting, which is substantial, because the labs’ revenue flows through the same cloud companies that sell them their computing power. David Cahn of Sequoia puts the revenue now needed to justify the spending near \$3tn a year. Bain’s model, which generously assumes the savings AI generates are reinvested, still leaves an \$800bn shortfall in 2030.
Labour substitution is nowhere near paying for it. About 113,000 AI-attributed US job cuts year to date, 0.07% of employment, implying roughly \$20bn of wage saving against \$700bn of capex.
Neither adoption nor pricing is closing the gap. Firm-level adoption has sat flat at 17% to 20% for six months. McKinsey finds only 6% of firms reporting an earnings impact above 5%, Deloitte only 25% with pilots actually in production, and PwC finds 74% of the measured value captured by 20% of firms. No study shows population-level revenue growth attributable to AI. On the consumer side, roughly 5% convert at \$240 a year and advertising runs a 1.3% click-through rate against Google Search’s 29.2%.
The value is captured upstream, not by anyone selling AI. Nvidia runs gross margins near 75%. The labs run 33% to 50%. Cursor, the fastest-growing application business in the sector, reached a \$2bn annualised revenue run-rate and only slight gross margin profitability. Revenue can therefore grow enormously without ever justifying the capital, because most of what customers pay flows through to whoever sold the computing power.
B. Moats and Chinese pricing
There is no moat. A moat is whatever stops customers leaving for a rival. Here there is very little. Providers publish compatible application programming interfaces (APIs), the standard way one piece of software talks to another, so swapping model is a settings change rather than a rebuild. Nor does inference, the business of actually running a model to answer a request, have network effects. A model does not get better for one customer because other customers use it.
Open-weight models reach frontier parity on a short lag. Open-weight models are those whose parameters are published, so anyone can run them on their own hardware without paying the developer. Industry estimates put it at three to six months, and the effect is visible in the price series rather than the benchmark tables: Epoch measures the cost of a fixed capability level, priced per million tokens, the chunks of text a model reads and writes, falling anywhere from 9x to 900x a year, with GPT-4-level performance on GPQA falling 40x a year. DeepSeek now leads all providers on OpenRouter by token count. The lag figure is a practitioner estimate rather than a measured series, but the direction is not in dispute.
Chinese models deliver most of the product at a fraction of the price. DeepSeek V4 Flash costs \$0.28 per million output tokens against Claude Opus at \$25. Claims of near-parity on coding depend heavily on which benchmark and agent harness is used, but the price gap is not in dispute.
Buyers are choosing adequate over best. The highest-volume Chinese models on OpenRouter sit outside the intelligence top ten, and US token share fell from about 70% to about 30% in a year.
The price war is live. OpenAI cut Luna 80% three weeks after launch, vendors are giving away tokens to buy share, and Telnyx went from \$100,000 a day on Anthropic to \$100 per agent per day on open-weight models.
The economics are utility-like, not software-like. Lab gross margins run 33% to 50% against 75% to 80% for software as a service (SaaS), and they must spend 45 to 57 cents of capital for every dollar of revenue, which is what a power station looks like, not a software company.
C. Depreciation and accounting
Six-year lives are defensible physically and flattering economically. The hourly rental price of an H100, the apex Nvidia chip of 2023, is down about 70% from its peak while its book value has fallen about half that, because it is written down evenly over six years regardless. The chip still works. It just earns far less, and only earnings repay the borrowing.
Capital spending is quickly becoming operating cost. The shell, the power and the cooling last twenty-five years. The chips, memory and boards inside do not, and they are about 70% of the capital cost: Epoch puts servers at \$5.0bn of a one-gigawatt build against \$2.1bn for shell, substations and cabling. So a large share of this spending is not a one-off build but the first instalment of a bill that arrives every few years for as long as the business exists. That changes what has to be earned back. It is not \$700bn once, it is something close to \$700bn a cycle, plus a return, plus interest on whatever was borrowed. It also means the depreciation charge is less a statement about accounting honesty than a forecast of cash that will have to be spent again. Trailing depreciation and amortisation (D&A), the charge that spreads an asset’s cost over its useful life, is running about \$149bn against \$434bn of capital spending, and that gap closes over 2027 to 2029 whatever demand does. A company that stops expanding does not stop spending. It stops growing while still writing the cheques.
The commitments are bigger than the spending, and cannot be unwound. Microsoft has signed \$329bn of data centre leases that have not yet started, up from \$92.7bn a year earlier. Across the five, obligations of this kind total about \$1.67tn and appear in no debt figure. Because developers are building against those signatures, they are locked in to spend even if the bet goes south. Capital spending will keep running well after the case for it weakens.
D. Financing
Free cash flow has gone. On each company’s own definition: Alphabet’s Q2 2026 free cash flow was minus \$5.9bn, its first negative quarter since the 2004 listing; Amazon’s trailing twelve-month figure was minus \$7.6bn against plus \$18.2bn a year earlier; Meta’s Q2 was \$784m, down 91%; and Oracle’s fiscal 2026, ended 31 May, was minus \$23.7bn. The periods and lease treatments differ, so read these as four separate signals rather than one aggregate.
Debt went structural in four quarters. From 9% to 32% of capex, with \$570bn of AI-related issuance projected for 2026.
Bond demand is already turning. In February these bond sales attracted five dollars of orders for every dollar on offer. By July it was under two. Some deals were pulled entirely, and nearly 80% of data centre bonds sold since early 2025 now change hands below their issue price.
Oracle is the visible stress case. It was cut to BBB-, the lowest rung still counted as investment grade, one notch above junk. The cost of insuring its debt against default reached about 200 basis points, the highest since 2008. Oracle is the largest non-bank borrower in the US investment grade bond index at \$117bn, and many funds are barred from holding junk, so a further downgrade would force them to sell whatever the price.
A further downgrade would be a trigger, not a failure. Oracle would survive junk status comfortably as a business, on $67bn of revenue growing 17% and $32bn of operating cash flow. What it could not do is fund $90bn of annual capital spending without cheap and continuous market access, since it needs to raise roughly $40bn next year to do so. A downgrade would therefore force a spending cut rather than a default. That matters because none of the five wants to be first to retreat, so a forced cut by one supplies the cover for the rest. It would also be the only dated, public event anywhere in this chain.
E. Counterparty risk, meaning who fails to pay whom
OpenAI is the keystone. About \$24bn of annualised revenue against \$750bn of planned compute and over \$1tn of commitments. Its own figure moved from \$1.4tn to \$600bn to \$750bn inside six months.
Circularity: the seller funds the buyer. Nvidia’s customers often cannot afford Nvidia’s chips, so Nvidia takes equity in them or guarantees their borrowing, and they spend it on Nvidia chips. It owns about 13% of the neocloud CoreWeave and backstops \$6.3bn of its capacity. It put \$30bn into OpenAI, which contracted \$300bn with Oracle, which borrowed to build, and much of that money returns to Nvidia at list price. AMD handed OpenAI a warrant over about 10% of its own shares in exchange for a 6GW order. The exposure has since shifted from equity into far larger guarantees, including a reported \$250bn backstop for an Ohio site against an existing guarantee book of \$3.5bn. The revenue is real in accounting terms. What you cannot see from outside is how much of it would exist without the seller’s money behind it.
This is what happened in 2000. A supplier lends a customer the money to buy the supplier’s product, booking a sale today against repayment that depends on the customer surviving. Lucent committed \$8.1bn, roughly 24% of its revenue, and took heavy provisions as the loans soured. Nortel extended \$3.1bn. Forty-seven of the telecom start-ups they funded went bankrupt between 2000 and 2003, and money available in January 2001 was gone by April. The difference now is that the financing sits in equity stakes and contingent guarantees rather than on the income statement, so it is far less visible. Against that, Nvidia shows none of the distress signature: gross margins near 75% and days sales outstanding falling from 51 to 45.
Failure would not look like bankruptcy. Neither lab carries meaningful debt, and both filed confidentially for listings in June. OpenAI holds about \$73bn of cash against a \$27bn annual burn, and the firms who would be creditors are the same ones whose demand forecasts depend on survival. Nobody wants a filing; they want a renegotiation. Expect a down round, a painful listing, or contracts quietly stretched, surfacing as a shrinking backlog in a footnote.
A company failing is not the economy failing. Firms fail constantly without macroeconomic consequence. What would matter is capital spending turning down across the whole complex at once. Guidance is a poor way to see that coming, since definitions differ, off-balance-sheet vehicles sit outside it, and nobody wants to be first to announce a retreat. Actual cash spending and Nvidia’s data centre revenue are harder to dress up.
F. Physical constraints
Power binds, not chips. Nadella says Microsoft is not chip-constrained, and \$80bn of Azure orders sit unfulfilled for want of power.
Equipment lead times are hard walls. Transformers run 128 weeks, generator step-ups to four years, turbines are sold out to 2030, and interconnection takes a median 61 months. Electricians take four years to train and their wages are rising 9.9% against 2.4% economy-wide. The scramble has already reached consumers: memory prices rose about 60% in a quarter, with Gartner projecting 130% by year end and personal computer shipments down 10.4%.
A note on Australia. These constraints are why the buildout is looking offshore, toward jurisdictions with spare grid and faster permitting. The Australian Financial Review reported on 8 July 2026 that Anthropic has issued a confidential request for proposals seeking at least 1.4 gigawatts (GW) of Australian capacity, worth up to US\$15bn, with at least 1GW operational by end-2027. The RFP went to CDC Data Centres, AirTrunk, NextDC, Iren and Stack, and may be split across four or five contracts. It is non-binding, no site has been named, and the report placed a final decision at least six weeks away. Treat it as a live proposal rather than a commitment. Australia is a beneficiary of the American bottleneck rather than of independent Australian demand, which makes the activity contingent on the same capital cycle. If the buildout stalls in the United States, projects that were only ever the overflow are the first to be cancelled.
G. Stranding risk, meaning kit that loses its value before it is paid off
Efficiency outruns hardware economics two to one. Algorithmic efficiency improves about 3x a year against 1.37x for hardware performance per dollar. For a fixed workload that strands compute faster than hardware economics improve, so demand must expand rapidly merely to hold utilisation constant. Whether it does is the whole bull case.
Fibre is the precedent, and the warning. Capacity per unit of installed capital rose about 128x in six years. You need no demand disappointment to get a credit event, only a deflation surprise.
H. Macro fragilities
The market is concentrated and the comfortable multiple is margin-dependent. The top ten companies are 41% of index weight against 32% to 34% of earnings, and the cyclically adjusted price-to-earnings ratio (CAPE) at 42.1 sits within 5% of the December 1999 record. The reassuring forward multiple of 19.6 rests entirely on record margins: at the 11% to 12% normal before 2020, the same share prices would represent about 30 times earnings.
Capex is the transmission channel, not the stock market. Bridgewater puts AI capex at about 140 basis points of US growth, so if it merely stops growing rather than contracting, that 1.4 points goes. Add the equity wealth effect at 0.2 to 0.5 points and, taken mechanically and before offsets or second-round effects, the two channels amount to 1.6 to 1.9 points off a trend near 2%. Neither requires a financial crisis.
The labour flow has stalled, and a downturn would show in output before jobs. Payrolls fell 23,000 in July against an expected gain of about 80,000, with May and June revised down by a combined 103,000 and the twelve-month average now 34,000. Manufacturing construction is down 22% year-over-year, so the non-AI capital spending leg has already rolled over. Data centres are extraordinarily capital-intensive per job, with Meta’s \$1.5bn Texas site supporting about 100 permanent positions against 1,620 for a comparable battery plant, so a reversal subtracts output without mass layoffs and will be slow to recognise.
III. The Unknowable: Private credit, and why nobody can tell you
The claim in I, that the danger signs of 2007-08 are absent, rests on bank balance sheets, and bank balance sheets are genuinely clean. That is not the same as saying the risk is gone. It has been moved to institutions that do not mark to market, meaning they do not restate their holdings at what those holdings would actually fetch today, and which do not report weekly or face stress tests. Whether that is prudent dispersal or a hidden concentration is the one question in this paper that cannot currently be answered.
Nobody can price it. More precisely, there is no continuous market-clearing price. Private credit loans do not generally trade continuously, so value comes from managers, third-party valuation agents and models struck quarterly. S&P notes that two managers applying the same discounted cash flow framework can produce different marks on the same credit because of differing benchmark and credit assumptions. Both are compliant.
The one price signal available is unflattering. Listed business development companies (BDCs), the publicly traded vehicles that hold private credit, have traded at discounts to their stated net asset value (NAV) through 2026, with the larger names around 10% below and Blue Owl’s vehicle nearer 25%. Saba Capital has separately bid for private credit fund stakes at discounts of 20% to 35%, which is a tender price rather than a trading level. Persistent discounts are consistent with public investors valuing the loans below the managers’ marks, though fees, liquidity and expected future losses also contribute.
The exposure is measured, and it is concentrated in software. The Bank for International Settlements (BIS) finds that loans to software-as-a-service firms were about 19% of all direct lending at end-2025, and that business development companies alone have lent around \$115bn to software firms, roughly a fifth of their total lending and over 80% of their technology portfolios. The International Monetary Fund (IMF) finds some funds with software above half their net asset value, and software is the one sector where the share of borrowers with interest cover below one has clearly worsened. Note this is exposure to AI-disrupted software rather than to AI infrastructure itself, which is the sharper point: private credit’s largest sector concentration is the sector most exposed to being automated.
Banks are distributing the risk rather than holding it, and it is getting harder. JPMorgan and MUFG did place almost all of Oracle’s \$38bn data centre loan, but it took roughly eight months and more than two dozen banks and investors to absorb, with under \$1bn still outstanding in April 2026. The deal succeeded; the effort required is what market participants read as the warning. Banks are also using contracts that pass credit risk to private credit funds and insurers while the loan stays on their own books. CoreWeave’s \$8.5bn facility was the first GPU-backed financing rated investment grade, at A3, which makes it eligible for insurers and pension funds whose mandates require investment-grade paper.
The holders are insurers and pension funds. Privately placed bonds, many of which do not trade in a liquid public market, are 23.4% of the bonds US insurers are permitted to count towards their capital, up from 18.3% in 2021, and private-equity-backed insurers control about \$900bn of US insurance liabilities against \$67bn in 2012.
Losses will surface slowly rather than acutely. There are no depositors who can withdraw, no borrowing that must be renewed overnight, and no requirement to restate values daily. This is a grind over years rather than a single weekend of collapse.
Investors are already trying to leave. Blue Owl received redemption requests of roughly 22% and 41% of assets across two funds in 2026, against a contractual quarterly limit of 5% built into the vehicles by design. That limit did what it was written to do, so this was not an emergency gate. Only 0.2% of the relevant book was flagged as unlikely to be repaid, meaning investors were leaving rather than borrowers failing. The IMF calculates that liquidity buffers could be exhausted in five to seven quarters under moderate stress.
The channel that matters is the funding stop. Private credit is expected to supply \$800bn of a \$1.5tn financing gap to 2028. Impaired holders stop writing new business, and the AI buildout loses its marginal lender at precisely the moment hyperscaler free cash flow has turned negative and bond order coverage has halved.
Insurers can carry impaired assets at full value for years. US insurers report on statutory accounting rather than mark to market. A performing bond sits at amortised cost and is written down only once it is deemed other than temporarily impaired. That judgement can be made before an actual payment default, but it is a judgement, and it lags a market price. An insurer can hold paper worth 70 cents, carry it at 100, and be entirely compliant while the coupon keeps arriving. Payment-in-kind toggles, which let a loan capitalise its interest instead of paying cash, extend that further. The rot can run deep before anything is found.
When it is found, recognition is forced by ratings rather than prices. Capital requirements run off the credit rating, so downgrades raise the capital charge and compress the ratio. This is why the National Association of Insurance Commissioners (NAIC) introduced a ratings discretion amendment in January 2026, letting its valuation office override a rating that differs by three or more notches from the NAIC’s own assessment. That is a regulator stating in writing that it does not trust the ratings insurers rely on to value this paper, and taking back the one lever that forces recognition. Treasury Secretary Bessent met state insurance commissioners in May specifically on private credit.
The failure mode is runoff, not collapse. Life insurers have no run mechanism worth the name, because liabilities are long-dated and annuities carry surrender charges. Faced with a capital shortfall they raise equity, hand a slab of their policies to a Bermudan reinsurer where the capital rules are lighter, put the business into runoff, meaning they stop selling new policies and simply pay out the old ones over decades, or sell to a consolidator. Selling the assets comes last, because that crystallises the loss. Policyholders are largely protected by state guaranty associations funded by levies on surviving insurers, so the cost is mutualised across the industry and surfaces as reduced capacity over years.
Institutional funding agreements are the one place a run can start. Executive Life in 1991 and General American in 1999 both failed through short-dated institutional liabilities rather than retail policies. If something breaks acutely this time, that is where to look. The private-equity-backed annuity writers are the concentration to watch, having grown fastest, held the most illiquid assets, and leaned hardest on offshore reinsurance.
The chain from a data centre to a retiree is shorter than it looks. A campus in Ohio raises construction debt. The banks fail to syndicate it and transfer the risk to a private credit fund. The fund places it with a private-equity-backed annuity writer, which holds pension obligations transferred out of a corporate defined-benefit scheme. The member of that scheme made no choice, selected no fund and was never asked. They are now an unsecured creditor of an entity holding data centre paper they have never heard of. If it fails, state guaranty associations cover annuity benefits only to a present-value cap of roughly \$250,000 per person. The cap is a floor of guaranteed recovery rather than a ceiling on total recovery, since amounts above it remain claims on the failed insurer’s estate. Still, those just above the cap are the ones exposed to loss: they saved enough to have something at risk and not enough to absorb losing it. Life insurer failures are rare, so this is a tail risk, an unlikely but severe outcome, rather than a forecast. It is also the reason regulators are moving while the numbers still look manageable.
Acting early tightens the funding, which is the point and the problem. Banks withdrew from lending to heavily indebted companies after 2008 because capital rules made holding it expensive, and private credit grew from almost nothing to \$1.4tn to fill the space. If regulators now raise the cost of insurers holding this paper, whether by reclassifying it as riskier, demanding more capital against it, or questioning the values put on it, the second lender retreats as the first one did. There is no third in line. Prudential caution and the funding stop in point 8 are the same event viewed from different ends, which is why the response to the unknowable is likely to arrive as a bad quarter for the bad rather than as a crisis of its own.
In short. The problem is not that private credit is large. It is that nobody can price it, that the people holding it are the same people expected to fund the next leg of the buildout, and that the only forced-selling trigger sits with insurance regulators rather than with markets. If this becomes the story it will arrive quietly, and by the time it is visible in anyone’s accounts it will already be several years old.
Weighing the three parts
The three parts of this paper answer different questions, which is why they can all be true at once. Part II (the bad) is probability: will the capital investment earn its return? Part I (the good) is severity: what happens to the economy if it does not? Part III (the unknowable) is transmission: how does disappointment in the first become damage in the second? Read that way, the Bad is more likely right than optimists concede, and the Good more likely to hold than pessimists concede.
On probability, the weight sits with the Bad. A five-to-one gap between spending and revenue can be closed by growth. The same gap in a market with no switching costs, open-weight parity within months and a competitor pricing at a 99% discount cannot, because the growth arrives at collapsing prices. A large share of this capital will probably earn poor returns.
Return expectations are running well ahead of what is likely to be delivered, which will disappoint in the short to medium term but probably not the long term. The technology looks like a slow build as use cases are found and exploited, financed on a far shorter schedule than that implies.
The recession case does not really rest on the arithmetic. It rests on correlation. Five firms are building the same thing on the same assumptions, none wants to be first to cut, and their suppliers and lenders are exposed to the same bet. That is why the spending keeps rising past the point the returns justify it, and why there is a real risk it stops together rather than tapering. A sector that large moving in one direction at one time is what turns a disappointing investment into a downturn, and the correction overshoots because capital spending is the fastest thing in an economy to cancel.
None of this is exotic. Recessions happen from time to time, usually at the end of a long expansion, and usually because something that attracted too much capital stops attracting it. The railways did it in the 1840s, electric utilities in the 1920s, and fibre in 2000. In each case the technology mattered enormously and the investors were not the ones who got paid. 2001 is the shape: eight months, 0.6% of GDP, with the sector that drew in the money taking most of the pain. What is unusual here is the scale of the single sector involved and the opacity of who has funded it, not the fact of a downturn.
On severity, the weight sits with the Good. Surviving is not the same as avoiding: a shallow recession at some point is the reasonable expectation rather than the bad case, since the direct arithmetic is roughly 1.6 to 1.9 percentage points off a trend near 2%. But that points to a 2001-style downturn rather than a 2008-style collapse. Households, banks and non-AI companies are not carrying the leverage that made 2008 systemic, the capital at risk is mostly equity rather than deposits, and the spending response to falling share prices is small.
Two things prevent settling comfortably there. Concentration: the ten largest companies are now 41% of the S&P 500, so an equity correction would fall disproportionately on the very companies driving the investment boom. And private credit has no 2001 equivalent, because it did not exist at this scale.
Two things push the other way. Capital spending is sticky: no hyperscaler has yet signalled retreat, and the commitments run through multi-year contracts and long leases. And the power constraint, filed here under the Bad, paradoxically prevents a glut. You cannot overbuild what you cannot connect to the grid, which is a real difference from the fibre build, where cable went into the ground as fast as the money arrived.
So the resolution is this. The Good decisively answers the question “is this 2008”, and the answer is no. The Bad decisively answers the question “will the AI capital investment earn its cost”, and the answer is largely no. Both are true because they answer different questions.
The point most likely to be underweighted is timing. The revenue gap has no deadline. Nothing forces it to resolve in any particular quarter, which is why the buildout can persist far longer than the pessimists expect, even years. Private credit can remain hopeful without ever being marked to market. What has a deadline is the tolerance for poor returns and the willingness to lend against them.
No comments:
Post a Comment