Too Soon

The Clause

On 28 October a clause expires in a loan agreement that very few people have read.

It sits in an amendment CoreWeave filed with the Securities and Exchange Commission on the last day of 2025, governing one of its delayed-draw term loans (DDTL), the facilities through which it borrows against chips and customer contracts as it deploys them. The amendment deferred the first test of the loan's debt service coverage ratio, the measure of whether cash flow covers interest and principal, to 31 October 2027. And until 28 October 2026, it allowed CoreWeave's shareholders to cure any covenant shortfall with fresh equity as many times as they liked. After that date, cures are limited.¹

Nothing dramatic need happen on 28 October. CoreWeave has not breached a covenant, and nothing in its filings suggests an imminent one. But the clause is a small, precise picture of the problem this part of the series is about. Lenders financing the AI build have written its timing risk into their documents. They have granted grace, deferred tests and allowed shareholders to top up, all of which assume that the revenue is merely late, not missing. The grace period runs on a calendar. The power station does not.

Part I, Too Slow, set out the physical problem: the plans call for far more capacity than measured demand requires, every gigawatt is built from scarce and concentrated inputs, and the critical path now runs in years. This part follows the bill. What happens to the cost when the timetable slips? What does that do to a single building's economics? How do three real projects distribute the same shock? And who, in the end, holds the debt when the revenue misses the connection?

The short answer is that the debt service starts on time, and the revenue does not. That is the second leg of the Catch: too soon.

The Ratchet

Inflation in a construction project behaves differently from inflation in a shop. In a shop, a higher price is paid once. In a project, a higher price usually arrives with a delay, and the delay costs money too.

Call it the cost-to-completion ratchet. A ratchet turns one way. Each click holds, and nothing slips back. A price paid, a month lost and a chip ageing in its crate are not refunded when conditions ease. This ratchet has three teeth: prices, time and the chips. The industry's own cost guide names the first two almost in passing: project costs, Cushman & Wakefield writes, are being driven up by "higher equipment pricing, increased carrying costs and rising demand".²

Prices

The inputs that matter most to a data centre have risen far faster than inflation. Since December 2021, according to the producer price indices Cushman & Wakefield tracks, switchgear prices are up 60%, generators and motors 33%, cooling equipment 32% and transformers 29%. Copper is up 46%. General construction materials, by comparison, are up 7%.²

Labour has followed. Skilled construction labour costs are up 13% since 2021, and the building cost index that reflects skilled trades is up 22%.² Data-centre construction costs per megawatt are up 21% since late 2024 alone, against general inflation of about 5% over the same period.² Even the land is ratcheting, because the scarce thing is the power; powered land in primary US markets averaged $584,000 per megawatt this year, 35% above its five-year average and 51% more than a year earlier.²

Time

A price rise is a nuisance. A price rise that comes with a delay is a different animal, because capital is never free while it waits.

Every month a finished building cannot be energised, interest accrues on the money already drawn, the contractors and the security guards are paid, and the chips already delivered depreciate without earning. The industry knows this. Cushman's cost breakdown sets aside 16% of a facility's budget for contingency and a further 3% for escalation before construction begins.² Nearly a fifth of the budget is provision for surprises, and the surprises listed in that guide (long lead times, price escalation during procurement, labour shortages, utility-driven changes and shifting rules on permits) are precisely the ones now occurring.

The Chips

The third tooth is peculiar to AI. The most expensive component in the building has the shortest commercial life. A chip delivered to a site that is waiting for power loses value while it waits, because the next generation arrives on schedule even when the electricity does not. A project delayed by a year may open with hardware a full generation behind the one its customers now want.

Put the three together and each click makes the next more likely. Prices rise because capacity is scarce. Scarcity causes delay. Delay adds carrying cost and erodes the value of the chips. The project ends up costing more for the same capacity, delivered later, with older equipment.

Two Inflations

The ratchet has a macroeconomic cousin. Artificial intelligence is now deflationary and inflationary at the same time, in different places.

The price of intelligence is falling fast, as Part III shows. The price of the physical things needed to produce it is rising. September's inflation data made the split unusually visible. Consumer prices rose 0.6% in the month and 3.7% on a year earlier. Producer prices rose 1.3% in the month and 6.3% on the year, with final-demand energy up 10.8% in a single month.³

The route from an oil field to a data centre runs through a refinery. The producer price index for crude petroleum rose 28.3% in September and for diesel 21.0%.³ Diesel moves the steel, concrete and switchgear to the site, powers the cranes and earthmovers, and fuels the backup generators. A disruption to refining reaches the data centre at several points before it is finished. That does not mean the same cost is counted several times. It means the shock touches several genuine costs at once, and where capacity is tight, it lengthens schedules as well as raising prices.

Fairness requires the other side. September's surge was driven by energy; the final-demand construction index was unchanged, and producer prices for commercial electric power fell 0.6%.³ The large hyperscalers can contract and hedge ahead. Not every project faces the headline number.

But put the two inflations together and the combination is awkward. Rising input costs and wider credit spreads both cut the return on the next project. If enough marginal projects are deferred, investment slows and growth weakens, and government bonds may rally on the prospect. The cost pressures, though, do not vanish with the rally; contracts have been repriced, equipment is still scarce, schedules are still slipping. Weaker growth and persistent cost inflation can coexist, and a Treasury rally in those conditions does not mean financing has become easier for the companies that need it most. The bond market section below returns to this.

Inflation Buys No Gigawatts

To size the effect, HindeSight modelling applies three forward scenarios to the US and Canadian development pipeline in Cushman & Wakefield's 2026 cost guide: about $2.3 trillion,* of which about $492 billion is under construction and $1.8 trillion planned or precommitted.⁴ A pipeline is not the same as capital already committed or financed. Each scenario applies a further, weighted rise in component costs to the share of spending not yet fixed by contract: about 2% in a controlled case, 5% in a central case and 10% in an adverse case, with delays of nothing, six months and a year.

The central case adds about $71 billion to the pipeline's cost. The adverse case adds about $193 billion. The controlled case still adds about $20 billion.

*Cushman & Wakefield's press release of 3 September 2026 describes the $2.3 trillion as a global pipeline. This letter follows the cost guide itself, the source document, which presents the figure alongside its US and Canadian market analysis (pp.29-31).

Not one of those dollars buys an extra megawatt. A larger bill for the same capacity is the ratchet's signature, and it is worth stating plainly, because it will be misread in the headlines: inflation buys no gigawatts. When announced AI spending rises next year, some of the increase will be more capacity. Some will simply be the same capacity costing more. Amazon has already shown how. Its chief financial officer attributed most of this year’s increase in its capital spending, from about $200 billion to about $220 billion, to the rising cost of memory, “with no change to the build plan”. Morgan Stanley draws the conclusion: headline capex growth “is therefore partly a memory-price variable”.⁵

One Hall, Modelled

Aggregate pipelines hide as much as they show. So our modelling also takes a single, ordinary building, a 100-megawatt AI hall, built and equipped.

The Base

The base uses Cushman's published average cost for a greenfield facility, $17.6 million per megawatt, all-in except for the chips, and adds $20 million per megawatt for the accelerator servers.² That puts the hall at $3.76 billion, or about $37.6 billion per gigawatt, inside the range Part I arrived at.

The choice of the average, rather than Cushman's higher figure for AI-grade facilities, is deliberate. If the results bite on the average, they cannot be dismissed as stretched. Using the AI-grade facility cost and dearer accelerators, the same hall costs up to $5.13 billion.

What It Must Earn

Before any delay or overrun, the hall must earn about $848 million a year, about $8.5 million per megawatt. That requires a billed price of about $883 per kilowatt per month.

The composition matters more than the total. Of that $848 million, about $486 million is needed simply to recover the accelerators over their six-year life. The building, which costs almost as much, needs about $207 million. The chips, in other words, account for well over half of the revenue the hall must earn, and they must earn it quickly, before the next generation makes them cheaper to replace than to run.

What Delay and Inflation Do

The central case adds about $120 million of capital and $14 million a year to the revenue the hall must earn. The adverse case adds about $500 million of capital and $81 million a year. Neither adds a megawatt.

Those numbers are not catastrophic for one building. That is the point. The ratchet does not usually destroy a project. It erodes the margin a leveraged project was financed on. A hall built with two-thirds borrowed money that needs 10% more revenue than planned, a year later than planned, does not fail. It simply moves from covering its debt service comfortably to covering it narrowly. Minsky had a name for that movement, and we will come to it.

Three Projects, One Shock

A model is a test. Real projects are where the test is marked. HindeSight modelling applies the same inflation and delay scenario to three very different projects, each chosen because it represents a different way of financing the build.⁶

The central scenario applies about 5% further construction inflation to the spending still exposed to repricing, with hypothetical delays of six months at Jupiter and three months at the others. These are sensitivities on our assumptions, not reported overruns. Jupiter's and Galaxy's construction costs are derived (Jupiter's from its loans at an assumed 70% loan-to-cost; Galaxy's from a cost per megawatt), because neither has published a budget.

Jupiter: The Completion Problem

Jupiter produces the largest modelled drag, about $960 million in the central case: $433 million of construction escalation and $527 million of financing carry. The carry is the larger half, which is the ratchet's second tooth, time, at work. The figure is highly sensitive to assumptions, because the actual budget, draw schedule and allocation of overruns are not public. But the direction is not in doubt. A project of this size, financed with bank loans, whose power supply still depends on a pipeline permit and a regulator's timetable, carries its delay risk on the people who lent to build it. Oracle's force majeure notice, if accepted, could defer its rent for up to three years after it begins.⁷ The developer's loans would not be deferred with it.

Galaxy Helios: The Credit Problem

Galaxy is the more interesting case, because it has done almost everything right.

The first phase of its Helios campus in Dickens County, Texas, delivered 133 megawatts to CoreWeave on schedule, and rent began in the second quarter.⁸ The second phase, 260 megawatts, is contracted to CoreWeave under 15-year leases and financed by $3.507 billion of secured notes at a fixed 9.875%, amortising 4% a year, with an uncapped completion guarantee from a Galaxy affiliate.⁹ Delivery is expected from the second quarter of 2027.

Yet between 28 September and 16 October those notes fell from 95.25 to 91.75, and their spread over Treasuries widened from 611 to 729 basis points.¹⁰ That is a deterioration of 118 basis points in under three weeks, on a project that is on time, fully contracted and guaranteed.

The coupon does not change; the notes are fixed rate. What changed is the market's view of the tenant. An illustrative test in our modelling allocates Galaxy's reported campus revenue to Phase II in proportion to its megawatts and finds coverage of the coupon and scheduled amortisation of about 1.10 times. A hypothetical 12% shortfall in collected rent would take it below one.⁶ That is not a covenant ratio, and no one is forecasting a shortfall. But it shows how thin the cushion is when the tenant is itself a borrower, and why the market reprices the building when it reprices the tenant.

Hyperion: The Risk Transfer

Hyperion is the counter-example, and the most instructive of the three.

Meta's campus in Richland Parish, Louisiana, is financed by $27.3 billion of bonds issued by Beignet, a vehicle owned largely by Blue Owl funds, with Meta as tenant and 20% partner. S&P rates the bonds A+, not because the project is safe in isolation (its modelled debt service coverage is a thin 1.12 times) but because Meta's contractual commitments, including a residual value guarantee, transfer most of the construction, delay and residual risk to Meta itself.¹¹

So the $235 million central stress at Hyperion is not a $235 million loss for its bondholders. It is a cost Meta has largely agreed to bear. The bondholders' exposure is to Meta's credit, the legal durability of its promises, and the price of a long bond in a rising-spread market. Even so, the bonds traded around 92 this month.¹² The market did not wait to find out which risk it held.

What Three Projects Show

Put the three side by side and a conclusion emerges that is more precise than "too many data centres". The weakest project is not the one with the largest cost increase. It is the one least able to absorb the increase without disrupting its financing, its contracts or its returns.

Jupiter's risk is physical and sits with its lenders. Galaxy's is financial and sits with its tenant. Hyperion's is contractual and sits, for now, with one of the strongest companies in the world. The same shock, three different addresses. That is why adding up every lease, loan and bond in the AI build as if each were an independent exposure is misleading. The task is to find out where the cash finally comes from, and who legally bears the loss when it does not.

Follow the Cash

So follow it.

Eight Names

Ares, one of the largest private-credit managers, recently documented more than a hundred digital-infrastructure financings from the past twelve months, issued across seven different debt markets: corporate bonds, construction loans, private placements, chip-backed loans, high yield, chip special-purpose vehicles and securitisations.¹³ Almost every one was engineered to achieve an investment-grade rating. And almost all of the risk, Ares found, converges on eight names: Meta, Oracle, Microsoft, Amazon, Google, Nvidia and, looking through, OpenAI and Anthropic.

Its matrix of 26 of those financings, about $573 billion, shows the concentration. Google and Broadcom touch 32% of the total; Meta, Oracle and Anthropic about 22% each; Nvidia 21%.^{13 14} The names are not independent of each other, either. Nvidia has committed capital to OpenAI, which buys compute from Oracle, which buys chips from Nvidia. Several names sit on both sides of the same structure: Meta pays the rent that services the Hyperion bonds and guarantees the value behind them.

One Name on Every Line

If one company stands at more interchanges than any other, it is CoreWeave. Across HindeSight's own tally of AI financings it appears in 14 of them. It issues four of the 24 high-yield data-centre bonds in Morgan Stanley's list and is the tenant on seven of the other 20.¹⁵ It borrows through unsecured notes, convertibles and a ladder of chip-backed loans. It rents from Applied Digital, Core Scientific, Galaxy and others, who have borrowed against its leases. Nvidia stands behind it with a commitment to buy up to $6.3 billion of its unsold capacity through April 2032.¹⁶

At the end of June, CoreWeave carried about $35.6 billion of debt principal and $16.3 billion of operating leases, with a further $35.5 billion of lease payments on facilities not yet delivered. About $4.4 billion of principal falls due in the rest of 2026 and $6.2 billion in 2027.¹⁷ Its quarterly interest bill more than doubled in the second quarter to about $640 million.¹⁸

CoreWeave sits on both sides of the lease. As a tenant, it rents buildings from Applied Digital, Core Scientific, Galaxy, Blue Owl's project at Lancaster, Pennsylvania, and PowerHouse's Digital Drive near Richmond, Virginia, whose lenders are lending against its rent. As a landlord, it rents its chips to Meta, Anthropic and Jane Street, among others; Microsoft was the largest, at 62% of revenue in 2024. In the second quarter of 2026 its three largest customers accounted for 36%, 26% and 10%, though the filing does not name them.¹⁹ The strong customers let it borrow cheaply. The weaker ones, and shorter contracts, make it borrow dearly. Each of its landlords' bonds rests on CoreWeave, and CoreWeave rests on its customers.

Adjusted earnings covered the quarter's interest more than twice. But those earnings are counted before the chips wear out, and once their depreciation is charged the business made an operating loss.¹⁷ Since June it has borrowed more, including a $2.6 billion chip-backed loan in August.²¹

The Leveraged Hyperscaler, and a Different Model

Two more names complete the picture, because they show how much the same demand can be financed in different ways.

Oracle is the hyperscaler that has chosen to borrow. It is the tenant at Jupiter in New Mexico, at Abilene, Texas, and in Vantage's $38 billion Project Jacquard, campuses in Shackelford County, Texas, and Port Washington, Wisconsin, and its backlog of contracted revenue, $664 billion at the end of August, is dominated by OpenAI. Only 13% of it is expected to be recognised in the next twelve months, and about half after three years.²² Its spending now runs ahead of its cash. In the quarter to the end of August, capital expenditure of $28.5 billion exceeded operating cash flow of $23.1 billion, leaving free cash flow of minus $5.4 billion, and the operating cash flow itself included $11.4 billion of customer prepayments.²³ It raised $43 billion of debt and $5 billion of equity in its 2026 financial year, and a further $20 billion by selling shares in the quarter just ended.²³ S&P cut its rating to BBB-, the lowest rung of investment grade, with a stable outlook, in July.^{24 25} Its five-year credit default swap widened from about 227 to 288 basis points between late September and mid-October.²⁶ Oracle is not CoreWeave. It is a large, profitable company. But it is the hyperscaler whose balance sheet looks most like a neocloud's.

Nebius shows a different route. Its agreement with Meta, described as worth up to $27 billion, combines about $12 billion of dedicated capacity over five years with up to $15 billion of further capacity that Nebius intends to sell to others, with Meta committed to take the remainder.²⁷ At the end of June it held about $6 billion of deferred revenue, cash received for services not yet delivered, and in the first half the increase in those advances contributed about $4.4 billion of its $4.5 billion of operating cash flow, against $8.1 billion of capital spending.^{28 29} In July it borrowed $775 million against deployed chips and an investment-grade customer's contract at SOFR plus 250 basis points, with a covenant requiring cash flow of at least 1.15 times debt service.²⁸

Customer prepayments are an excellent way to fund a build, because the money arrives before the construction is finished. They are not the same thing as profit. The cash has already been received; the computing must still be delivered, at a cost that justifies the investment. Advances flatter operating cash flow in exactly the years the capital is being spent.

The Laboratories

Behind the neoclouds stand the model laboratories, and their economics are the final link. Anthropic's draft prospectus, as reported by Reuters, shows 2025 revenue of $4.6 billion, an operating loss of about $8.1 billion, and $518 billion of compute commitments, about 80% of them non-cancellable. About a quarter of its revenue came from two customers.³⁰ The same filing says demand is limited by the availability of compute.

That is the Catch-22 in one document. The customer says it cannot get enough capacity. It has committed to more than half a trillion dollars of it. And its revenue, growing at an extraordinary pace, is still a small fraction of the commitments.

Where the Money Goes

The final destination of much of this paper is not a bank. It is an insurance company.

United States insurers hold about $9 trillion of invested assets. AI and technology now make up more than a quarter of the investment-grade credit market, and digital infrastructure 42% of investment-grade private placements this year.¹³ Record annuity sales are pushing new money into insurers that must be invested in investment-grade debt, at precisely the moment the AI build needs it. CoreWeave's co-founder Brannin McBee, who runs its financing, told the Financial Times this month that the pool of insurance capital is "unlimited", and that the company wants investment-grade ratings on more of its chip-backed debt to reach it. Loans backed by contracts with AI laboratories, he conceded, are "not ready" yet.¹⁸

“

The pool of insurance capital is “unlimited”. Loans backed by contracts with AI laboratories are “not ready”.

Brannin McBee, Co-founder, CoreWeave · Financial Times, October 2026

Ares put the mismatch more sharply than anyone. Insurers are looking for 50 to 100 basis points of extra spread; venture capital is looking for 50 to 100 times its money. Much of what backs these investments, Ares wrote, "in economic substance, is venture".¹³ The insurance industry declines to underwrite certain catastrophic risks as liabilities, it noted, yet "swallows whole as an asset" the same exposure when it arrives as an investment-grade bond.¹³

Ares, it should be said, is itself a large private-credit manager with insurers among its clients, so its warning is also a positioning statement. That does not make it wrong.

What Kind of Promise

Not every obligation in this series is debt, and they cannot simply be added together. They differ in who must pay, when, and whether they appear on the balance sheet at all. The same promise often appears twice; one company's purchase commitment is another's backlog.

Read across a row and the figures can be compared. Read down a column and they cannot be added. The first two rows are debt in anyone's sense. The next four are promises that become debt only when something happens: a building is delivered, a chip ships, a tenant fails. The last two are the other side of somebody else's promise. That is why a headline total of "AI debt" usually adds things that do not belong together, and why this series speaks of obligations and commitments instead.

The Construction Layer

The most important number in this series may be one nobody has audited. In July 2025 Morgan Stanley forecast about $2.9 trillion of global data-centre spending from 2025 to 2028, excluding power, and it was still describing the same framework this March. It expects the hyperscalers' own cash flow to cover about $1.4 trillion, leaving what it calls a $1.5 trillion financing gap. Corporate bonds fill about $200 billion of it and securitisation about $150 billion; private equity, venture and sovereign money another $350 billion. Private credit, led by asset-based finance, fills the rest: about $800 billion, more than half of the gap.³¹ That makes private credit about 70% of all the borrowing, more than a quarter of the entire build, and four times the corporate bond market's share. The hyperscalers' own cash covers less than half of the total.

The Estimates Keep Rising

Morgan Stanley's July 2025 framework is now the low end of the estimates, and every revision has been upwards.³²

The horizons differ, so the rows cannot be added or compared directly. The direction is the point. Even Morgan Stanley’s own count has moved: its October estimate, which adds leases and covers a different scope, comes to about $4.4 trillion over just three years, against the $2.9 trillion over four on which its financing framework rests.⁵ Dell'Oro, an independent market research and analysis firm based in Silicon Valley and founded in 1995, attributes its revision to higher hyperscaler spending guidance, and Goldman's rose by $800 billion after a single round of earnings. These are not independent measurements of demand. They are built largely from the same company plans that Part I called paper gigawatts, so when the plans grow, the estimates grow with them, and the larger estimates are then cited to justify the next round of financing. That is Soros's reflexivity again, in the forecasters' own spreadsheets, and it matches what Part III finds in the laboratories' revenue forecasts.

Set against those estimates, the debt that can be identified by name is still the smaller part. HindeSight's own tally, which counts only financings large enough to be reported, finds more than $370 billion of project, special-purpose, securitised and neocloud debt committed or arranged, including $60 billion still only proposed.³³ That is a floor, not a total. It misses bank loans and private deals that are never announced, and most of the build outside the United States. Around it sits the hyperscalers' own borrowing: about $290 billion of bonds issued by them and their suppliers in the past year, within some $770 billion of debt and lease obligations already on their balance sheets.^{33 34} Together, roughly $1.1 trillion of debt is already in place.

The point is what comes next. Morgan Stanley's low estimate alone needs about $950 billion of private credit and securitised debt by 2028; the project debt identified so far is well under half of that, and the higher 2026 estimates need far more. Most of the borrowing is still to come. It will have to be raised in markets that, as this part has shown, are already charging more each time they are asked.

It is arriving fast. UBS counted $125 billion of AI data-centre and project financings in the first part of 2025, against $15 billion a year earlier, and estimated that private credit's AI lending had nearly doubled in twelve months.³⁵ One tally of lender disclosures puts private credit funds' data-centre loans at about $135 billion, up from about $40 billion in 2023.³⁶ Eagle Point's $1.3 billion loan at Hubbard, Texas, is a typical piece.³⁷

What matters is where it sits. The financing has settled into layers. Private credit lends during construction, when the risk is greatest. Insurers buy the debt once a project is built and let. Securitisation turns the leases into bonds, and equity sits on top. The layer that carries the construction risk is the one funded least permanently. Some of its money comes from insurers and pension funds that can wait. Some comes from funds sold to individual investors on the promise of regular redemptions, like the Blue Owl technology fund whose investors have asked to redeem about 40% of its shares in every quarter this year and have received about 13% of what they asked for.³⁸

That is the shape of 2007. Then, the long-dated risk of mortgages was carried by vehicles that relied on short-term money, structured investment vehicles funded with commercial paper and hedge funds borrowing against what they held, and the funding broke before the mortgages did. Here, the long-dated risk of an unfinished data centre sits with lenders whose own investors can ask to leave every quarter. The building takes years. The redemption queue forms every three months.

Who Stands Behind the Private Credit

August's letter showed where much of this risk lands: in life insurers owned by the same private-equity firms that originate the loans. Apollo owns Athene; KKR owns Global Atlantic; Brookfield and Blackstone have built insurance arms of their own. Insurers with disclosed private-equity ownership already hold more than $700 billion of assets, and a broader count puts the figure near $1.2 trillion.³⁹ These insurers sell annuities, long-dated promises to savers, and need assets that yield more than government bonds while still counting as investment grade for their regulators.

They enter the build at three points. The largest is private credit itself: the manager originates the loan and places much of it on its own insurer's balance sheet, structured to carry an investment-grade rating. Morgan Stanley's own research calls this kind of lending "the new frontier for insurance balance sheets".³¹ Ares found that almost every AI financing it studied had been engineered to achieve an investment-grade rating, and digital infrastructure now makes up 42% of investment-grade private placements, a market insurers dominate.¹³ They are also among the largest buyers of the senior tranches of data-centre securitisations, and of investment-grade project bonds such as Meta's Hyperion paper once a campus is built and let.

This is the clearest form of the 2007 analogue. Then, subprime mortgages were turned into AAA tranches that insurers and pension funds could hold. Now, construction-stage AI risk, unfinished buildings, depreciating chips and three-year customer contracts, is turned into investment-grade private paper that annuity insurers can hold. The difference is that the originator, the manager and the buyer are often one group. That keeps the paper off public markets, with no daily price, and concentrates the risk inside the firms that also run the private-credit funds. Two pressure points follow. Much of the paper carries private ratings, so a wave of downgrades would raise insurers' capital charges all at once. And annuities are stickier than a fund's quarterly redemptions, but they can be surrendered; if rates rose sharply and savers moved, insurers would need cash from assets that have no daily market.

Who Owns What, and How Much

The thing I remember most about 2007 and 2008 is not the size of the losses. It is the shock of discovering who owned what, and how much. Mortgages written in California turned up in German regional banks such as IKB and Sachsen LB, in the investment funds of small Norwegian towns, and in the London derivatives unit of AIG, then one of the world's largest insurers, which had sold protection on a large part of the market.⁴⁰ Nobody had a clue where it had all been placed. Most holders found out at the same time as everyone else.

The AI build is being financed in a way that makes the same surprise likely. Morgan Stanley's $800 billion was private credit's share of a $2.9 trillion build. If that share held as the estimates rise, private credit's part of Goldman Sachs's $5.3 trillion would be about $1.5 trillion, and of Gartner's roughly $7 trillion close to $2 trillion.⁴¹ That is an illustration, not a forecast, but the direction is plain: the more the plans grow, the more of this paper ends up with private credit funds and the insurers behind them. CoreWeave's co-founder in charge of financing describes the pool of insurance money available to buy it as "unlimited".

And the paper is hard to follow once it is sold. As August's letter set out, much of it passes from the manager that originates it to an insurer the same group owns, and from there, often, to an affiliated reinsurer in Bermuda, where reserves ceded by US life insurers have grown from about $205 billion in 2014 to roughly $928 billion.⁴² Each step is legal and disclosed somewhere. None is consolidated anywhere. Private ratings, no daily price and offshore reinsurance mean that the question that mattered most in 2008, who holds it, is the hardest one to answer today.

That is not a prediction of losses. It is a warning about surprise. If the build stalls, the losses will be discovered rather than announced, and they will turn up in places whose owners did not know they were exposed.

Where the Hyperscalers Come Unstuck

The money, followed far enough, arrives at five companies with some of the strongest balance sheets in the world: Google, Amazon, Meta, Microsoft and Oracle. They are spending at a pace few people inside them claim to fully understand. On a podcast released on 12 October, Mark Zuckerberg described Meta's decision in late 2025 to "write this check for like, I don't know, like $100 billion or $200 billion" before knowing whether its models would work. His chief AI officer could not promise they would, and his chief financial officer warned repeatedly of the cost of failure. Meta, he said, was "building out this compute for inference ahead of knowing that the product was going to be ready."⁴³

The cascade described above does not end these companies. But it reaches them, and it does not reach them all in the same way. For the four that fund themselves largely from cash, the danger is not default. It is being forced to cut spending, write assets down and see earnings and valuations fall together. For the one that borrows, it is credit.

Five questions separate them.

• How is the spending funded? From operating cash flow, or from debt?
• How dependent is it on a few leveraged customers? The laboratories and neoclouds are the weakest links in the chain.
• What contingent promises has it made? Residual value guarantees, lease backstops and commitments to buy capacity are invisible until they are called.
• Can it resell spare capacity? A company with a large external cloud can rent out what one customer no longer wants. One without cannot.
• Is its core business exposed if AI squeezes the economy? Advertising and retail are more sensitive to the consumer than enterprise software.

Oracle is the one that could come unstuck in credit terms. It has chosen to borrow, its spending runs ahead of its cash, its rating sits at the lowest rung of investment grade, and its backlog is dominated by a single customer, OpenAI. It is the tenant at Jupiter, Abilene and Vantage's Project Jacquard. If OpenAI's revenue falls short of its commitments, Oracle's backlog weakens, it defers or renegotiates leases, and the developers' loans and bonds feel it. A downgrade below investment grade would turn one of the largest corporate borrowers into a fallen angel, forcing some of its holders, insurers among them, to sell.

Meta funds itself from a very profitable advertising business, but it has no external cloud to rent its capacity to, which makes it the company least able to resell what it has built. It has guaranteed the residual value of the Hyperion campus in Louisiana and the Sopaipilla campus in El Paso, Texas, so if those buildings lose value, Meta pays. At the end of June it reported $349 billion of non-cancellable contractual commitments and, separately, $279 billion of leases not yet commenced, with about $68 billion more agreed in July, none of them on its balance sheet as debt.⁴⁴ And its core revenue depends on advertisers, who cut first in a downturn. What upends Meta is an advertising slowdown meeting fixed commitments, or its guarantees being called.

Microsoft is OpenAI's largest backer and has signed capacity contracts with several neoclouds, among them Nebius, Nscale, Lambda and IREN. Those are precisely the weakest links. At the end of its fiscal year in June, Microsoft reported $744 billion of contractual obligations, $444 billion of them leases and $194 billion purchase commitments.⁴⁵ If a laboratory or a neocloud fails, the risk returns to Microsoft as commitments it no longer needs. Its balance sheet can absorb that. Its earnings and valuation would notice.

Amazon has the best ability of all to resell capacity, because its cloud is the largest. But it has borrowed more than its peers: $77 billion of bonds in three offerings over eight months, the most of any hyperscaler in Ares's census.¹³ At the end of June it also reported $130 billion of unconditional purchase obligations and $137 billion of leases signed but not yet commenced.⁴⁶ Its largest AI customer, Anthropic, is also its investment, and is spreading its computing across Google's chips and Microsoft's data centres. In its June quarter, Amazon booked a $53.4 billion non-operating pre-tax gain on its investments, mainly Anthropic, while its free cash flow for the twelve months to June was minus $7.6 billion.⁴⁷ Gains of that kind flatter reported earnings, and would reverse if Anthropic's shares came to market at a lower price. Its own chips turn over quickly; the August letter noted Trainium 2 accelerators being replaced within about twenty months against a five-year book life. And its retail and advertising businesses are the most consumer-facing in the group.

Google (its parent company is Alphabet) remains the strongest of the five, but no longer the simplest. It designs its own chips. It ended June with $242.5 billion of cash and marketable securities against $98.2 billion of long-term debt, and earned $40.8 billion of operating income in the quarter.⁴⁸ What has changed is that its spending has overtaken its cash. In the second quarter it spent $44.9 billion on capital projects against $39.1 billion of operating cash flow, so free cash flow turned negative, and it expects to spend $195 billion to $205 billion this year and significantly more in 2027.⁴⁸ Its reported earnings flatter the picture: about $99 billion of the quarter's $112 billion of net income came from unrealised gains on shareholdings, not from the business.⁴⁸ Google is the clearest case of a wider turn. Deutsche Bank’s research institute puts it plainly: the hyperscalers are reaching the point “where capex is no longer covered by operating cashflow and where net debt exceeds” their earnings before interest, tax, depreciation and amortisation, and financing for AI has so far been conditional “not on revenue but on belief in a future payoff.”⁴⁹

It also carries the largest commitments of the five: $811 billion of purchase commitments and other contractual obligations at the end of June, $200.7 billion of them due in the short term, roughly a year of its capital spending.⁵⁰ And its exposure is the least visible. It has guaranteed the leases of Fluidstack, TeraWulf, Cipher and Hut 8, which serve Anthropic; if those tenants fail, Google pays the rent. It is reported to sit at the centre of roughly $200 billion of interlocking arrangements supporting Anthropic's computing, and to have backstopped ten data-centre projects totalling 2.4 gigawatts, leaving it liable for up to $44 billion if the leases fail, against about $815 million recognised on its balance sheet.⁵¹ In Texas, a package of about $16 billion financing the Hubbard campus leased to Anthropic is reported to rest on Google's guarantee of the senior debt.³⁷ And the concentration runs both ways: The Information reported in May that Anthropic's commitment to Google, about $200 billion over five years, was more than 40% of Google's cloud backlog at the time. Neither company has confirmed the figure, and later reporting on Anthropic's prospectus cited a smaller one.⁵²

So the question for Google is not solvency. A $44 billion call on its guarantees would be under a fifth of its cash and securities. The question is concentration. Much of its exposure runs to one customer, Anthropic, and it sits largely in guarantees and contracts rather than on its balance sheet. Google is the hyperscaler best able to absorb a loss. It is also the one whose loss would be hardest to see coming.

The pattern is the same as everywhere else in this series. Risk does not collect at the end of the chain. It collects where a strong balance sheet has promised to stand behind a weak one. The cash-rich hyperscalers are unlikely to fail. But each has written promises that convert someone else's shortfall into its own cost, and they can all be called in the same downturn.

Pricing the Tenant

The bond market has begun to make the distinction this series has been drawing. It is no longer pricing data centres. It is pricing the tenants. Even the industry's property adviser has drawn the distinction. Investors, Cushman & Wakefield writes, have moved from a "build at any cost" mindset to one focused on proving returns, and "newer neocloud platforms, particularly those that offer inferencing as a service with relatively thin margins, face a more challenging outlook".⁵³

Between 28 September and 16 October, the 24 high-yield data-centre bonds in Morgan Stanley's list moved very differently depending on who pays the rent.¹⁰ CoreWeave's own unsecured bonds widened by an average of 234 basis points. Bonds where CoreWeave is the tenant widened by 77. Bonds leased to SoftBank and Jane Street widened by 127, though that average rests on two bonds and one of them did almost all the work. Where Google stands behind the lease, the widening was 37 basis points; with Nvidia or Oracle as tenant, 31; with Amazon or Meta, 23.

The derivatives market said the same thing more loudly. CoreWeave's five-year credit default swap reached a record of about 1,195 basis points intraday on 15 October, from about 850 on 24 September.⁵⁴ Its unsecured 2032 notes traded below 80, and its 9.75% notes due 2031 yielded close to 16%.⁵⁴ On the same day, in Stockholm, Jensen Huang told Bloomberg that "CoreWeave's strength is improving every day". "I'm optimistic," he said. "Their cost of capital is actually decreasing." Nvidia, he added, had no need to backstop it.⁵⁵ The swaps, on the quoting source's assumptions, implied a probability of about 64% that CoreWeave would default within five years.⁵⁶ Huang's words were meant as reassurance. They can also be read as a boundary. Nvidia already supplies CoreWeave's chips, owns part of its equity, added $2 billion more in January and has agreed to buy its unsold capacity until 2032.⁵⁷ When it made that January investment, Huang called it "a small percentage of the amount of money that they ultimately have to go raise".⁵⁸ Telling the market that no backstop is needed is also telling it that no more money is coming from Nvidia. Oracle's five-year swap widened from about 227 to 288 basis points over the same weeks.²⁶ Investment-grade paper was not spared: Blackstone's QTS bonds for Project Excalibur in Fayetteville, Georgia, leased to Microsoft, traded below 88, and Meta's Hyperion bonds around 92.¹²

The cost of new money points the same way. CoreWeave's chip-backed loans now form a visible ladder: the Secured Overnight Financing Rate (SOFR) plus 225 basis points for a facility backed by Meta's contract and rated A3; plus 450 for one backed by non-investment-grade customers; plus 550 for its most recent, where the loan runs about five years against customer contracts averaging about three.²¹ The spread rises as the customer's credit falls and as the gap between loan and contract lengthens.

Inside a Delayed-Draw Loan

Most of CoreWeave's chip debt takes one form, and it repays a closer look. A delayed-draw term loan, or DDTL, is a commitment rather than a cheque. The lenders agree to provide up to a fixed amount; the borrower draws it in stages as chips are delivered and customer contracts start, pays interest only on what it has drawn, and pays a smaller fee on the rest. Each draw depends on conditions: equipment delivered, contracts signed, covenants met. A company can therefore announce billions of committed financing without being able to draw all of it, or any of it at once. Committed financing is not the same thing as available cash.

The delayed draw describes when the money arrives, not what secures it. A DDTL can be unsecured, resting on the borrower's general credit, its covenants and perhaps a guarantee. Or it can be secured, with a package that may reach well beyond the chips: data-centre equipment, receivables, bank accounts, the shares of subsidiaries and rights under customer contracts. CoreWeave's are secured on the borrowing subsidiary's assets and a pledge of its equity, and guaranteed by the parent. Security, though, is only as good as what it secures. A loan secured on GPUs is secured on an asset that loses value every year it is used, and faster with each new generation. A secured DDTL can still carry substantial refinancing and recovery risk.

The most recent, DDTL 5.5, shows the anatomy.⁵⁹

Six things decide whether a loan like this is safe. The commitment: whether every tranche is committed, or some can be withdrawn or are offered only at the lenders' discretion, which decides how much money is genuinely there to finish a build. The draw conditions: if chips or power arrive late, the money cannot be drawn, but the costs keep coming. The draw period: whatever is not drawn by the deadline lapses. The cost: the margin floats over SOFR, so the burden moves with rates. The ranking and the collateral: who is paid first, and what the chips will be worth when they are. And the customers: the contracts end, on average, about two years before the loan does.

That last gap is where the risk sits, and it is worth being exact about who holds it. First the lenders. JPMorgan and MUFG arranged the loan and placed it with lenders at an original issue discount, reportedly about 96 to 97 cents on the dollar; the other lenders are not named.⁵⁹ They hold the chips and the contracts. But this is not a stand-alone project loan: CoreWeave, Inc. guarantees it. If the chips, re-let in 2029, cannot cover the debt, the lenders turn to CoreWeave, and the shortfall falls on those behind them: its unsecured bondholders, whose notes trade below 80, and its shareholders. Each new secured, guaranteed facility moves the unsecured bondholders further back in the queue, which is part of why their bonds are priced as they are. And the covenants can bring the reckoning forward. A coverage test, cross-defaults and defaults tied to its material contracts mean that a stumble by a large customer could trigger the loan years before 2031.

The Same Ladder, for the Laboratories

The same ladder is now visible in the financings built for the laboratories themselves. In June, Apollo and Blackstone closed a $35 billion vehicle that buys Google's tensor processing units and leases them to Anthropic. Its $24 billion senior tranche pays 5.75%, because Broadcom guarantees the chips' residual value behind the senior notes; the $4.5 billion junior tranche, with no such guarantee, pays 8.5%.⁶⁰ Four months later, the junior tranche of a proposed $60 billion successor, Project North Star, financing about 1.5 gigawatts of the same chips, is reported to offer Blackstone more than 10%.⁶¹ Same customer, same chips, same sponsors; the price of the unguaranteed money has risen by more than a point and a half in four months. Where a supplier stands behind the debt, the cost stays low. Where only the laboratory's lease payments do, the market is charging more each time it is asked. And behind both, the supplier stands ready to fund the customer. Broadcom's latest quarterly filing discloses that Anthropic may, under certain circumstances and if needed, issue Broadcom convertible notes of up to $42 billion, tied to its five-year leases of Broadcom's chips; none had been issued by early August. Separately, Broadcom backstops those leases, with a maximum potential liability of $29 billion once they are fully deployed.⁶² A supplier that guarantees the lease and stands ready to fund the lessee is underwriting its own sales.

And the equity window is closing. In the space of a week, Firmus cancelled a $5 billion float in Sydney, AirTrunk postponed a $1.5 billion listing in Singapore, and SB Energy delayed its New York offering.⁶³ Hock Tan, whose company sits at the centre of the supply chain, told a conference in Singapore that the biggest constraint on AI may now be neither chips nor power but capital.⁶⁴

Morgan Stanley’s credit strategists expect the hyperscalers’ and neoclouds’ combined free cash flow to reach its most negative in 2027, the year of the largest spending. Their distinction is the one this part has been drawing: between “aggregate capital availability and the ability of individual borrowers to access that capital on workable terms”. The strongest can borrow almost regardless of price. In their illustration, $120 billion of new debt at 8.5% would take Oracle’s and SpaceX’s combined interest cover from 4.0 times to 1.8 times. The neoclouds, they add, are “almost fully relying on debt issuance”.⁶⁵

None of this is a default. All of it is the market charging a different price for the same technological demand, depending on who stands behind it. That is what one would expect if creditors had begun, at last, to look through the building to the tenant, and through the tenant to the cash.

Dominoes

Hyman Minsky divided borrowers into three kinds. Hedge borrowers' cash flows cover both interest and principal. Speculative borrowers' cash flows cover the interest, but principal must be refinanced. Ponzi borrowers must borrow to pay the interest itself. His insight was that stability breeds instability: a long period of easy finance moves borrowers steadily from the first kind to the third, until a shock reveals how far they have travelled.

The AI build gives that migration a physical mechanism. A project financed as a hedge, with contracted revenue comfortably covering its debt, becomes speculative when its equipment arrives late, because costs run before revenue starts. If the delay persists, or a tenant weakens, it must borrow to meet its interest. Each of the three legs of the Catch is a route down the same ladder.

George Soros added the reflexive turn. Prices do not merely reflect fundamentals; they change them. As spreads widen on the weakest names, lenders tighten terms for the next project. The next project slips. Suppliers demand more upfront. The marks fall on the structured paper insurers hold, and insurers become more cautious buyers. The hyperscalers, solvent and profitable, slow their next round of commitments, and the suppliers' order books follow. Contagion travels through expectations, orders and positioning long before it needs a default.

Not LTCM. Still 2007, With a Stronger Floor

The comparison that comes to mind is Long-Term Capital Management. Its Nobel-laureate partner Myron Scholes, as Roger Lowenstein tells it, pitched the fund as earning a tiny spread on thousands of trades, as if it were vacuuming up nickels that others could not see.⁶⁶ LTCM was not wrong on its models. It ran out of time and funding before its trades could converge: right on the asset, wrong on the clock.

The difference matters. LTCM concentrated its leverage in one fund, and fourteen banks could rescue it over a weekend. Here the leverage is spread across special-purpose vehicles, project bonds, private placements and insurance balance sheets. That might look safer. It is not. Because so many of those financings rest on the same few names, a break at one interchange does not stay small. It propagates through every structure that depends on the same counterparty. Dispersion hides the risk. It does not reduce it.

September's series, The Credit Line, measured the AI credit complex against 2007 and put it at roughly June: the warning issued, a liquidity mismatch visible in Blue Owl's funds, banks warehousing loans they could not sell, but no specialist lender failed and the best securitised markets still open. A month on, the comparison looks closer, not further away, because the resemblance is not only in the stress. It is in the purpose of the machine. It is also in the voices.

In July 2007, a month before the credit markets froze, Citigroup's chief executive, Chuck Prince, told the Financial Times that "as long as the music is playing, you've got to get up and dance. We're still dancing."⁶⁷ It is probably the most quoted line of the last crisis, and I make no apology for quoting it again. It reads as that crisis's epitaph, the moment the people inside a cycle told everyone outside it exactly where they stood, and it may yet serve for the next one. This cycle may already have its own lines. In August, CoreWeave's chief executive, Michael Intrator, said the company "enters the second half of the year, with more momentum than at any point in our history".⁶⁸ This month its co-founder Brannin McBee, who runs its financing, told the FT that the pool of insurance money available to fund it is "unlimited".¹⁸ Days later, with its bonds below 80 cents, Intrator told staff in a memo that the market was acting irrationally.⁶⁹ And from inside the best-funded laboratory of all, Google DeepMind’s chief strategy officer, Jasjeet Sekhon, called recursive self-improvement a “key part of the investment thesis” in “the biggest scientific bet civilisation has ever made”, then added: “We have this danger we could hit an AI air pocket such that the expenditures happen but the revenues don’t show up.”⁷⁰ None of this is dishonest. Prince was right that Citigroup had to keep lending while its rivals did, and Intrator may be right about demand. What makes such lines epochal is not that they are false. It is the certainty they carry, voiced at the moment the funding is most stretched, and how differently they read a year later.

“

CoreWeave enters the second half of the year, with more momentum than at any point in our history.

Michael Intrator, Chief Executive, CoreWeave · Second-quarter earnings call, 11 August 2026

“

We have this danger we could hit an AI air pocket such that the expenditures happen but the revenues don’t show up.

Jasjeet Sekhon, Chief Strategy Officer, Google DeepMind · August 2026, as reported by The Information

In 2007 the purpose of the mortgage machine was not to lend. It was to sell houses. Builders and specialist lenders such as Household and New Century originated the loans. The government-sponsored enterprises and Wall Street bought them, pooled them and sliced them into an alphabet of asset-backed securities, sold to insurers and pension funds that needed yield to pay their annuities and pensions. The borrowers ran from prime to subprime, on fixed rates and floating. Every slice paid a fee, and every fee depended on the next house being sold. The industry called it originate-to-distribute.

The AI build has rebuilt the machine, with chips in place of houses. The makers of the product now arrange the finance that lets their customers buy it. Nvidia has agreed to buy CoreWeave's unsold capacity until 2032, bought its shares, and offered conditional support of up to $105 billion for OpenAI's leases in Ohio.^{16 71} Broadcom stands behind the residual value of the chips in Anthropic's $30 billion of senior financing, backstops its leases up to $29 billion, and may take up to $42 billion of its convertible notes.⁶² Google is reported to have backstopped ten data-centre projects for up to $44 billion.⁵¹ Apollo, Blackstone, Blue Owl and the banks package the rest into special-purpose vehicles, project bonds, delayed-draw term loans, commercial mortgage-backed securities and take-or-pay leases, secured and unsecured, fixed and floating, and sell them to insurers and pension funds that need the yield; digital infrastructure now makes up 42% of investment-grade private placements.¹³ The tenants run from hyperscalers rated AA or better to neoclouds that lose money. Every slice pays a fee, and every fee depends on the next rack being sold.

Call it Originate-to-Ship (or, given what it finances, Originate-to-Chip). And as in 2007, every acronym rests on the same collateral: a building on a piece of land, worth what its tenant can pay. The alphabet changes the address of the risk, not the risk.

What differs is the floor. A project bond backed by Meta is not a mortgage-backed security stuffed with loans to borrowers who could never repay, and the strongest names in this chain earn extraordinary profits. The lesson from 2007 is the sequence: warnings appear first in particular borrowers, markets then discriminate between strong and weak structures, and funding stress finally spreads through the institutions that connect them.

A month on, the sequence has moved. Investment-grade project paper has repriced; bonds for a campus leased to Microsoft trade below 88, and Meta's Hyperion bonds around 92. The unguaranteed money for the same laboratory and the same chips has risen from 8.5% to more than 10% in four months.

The canaries are no longer confined to one mine. Jupiter's loans are still quoted at 89 to 91 cents. September's trigger for the next stage was a second station lighting up, and it has, though not on the line we were watching. No second project loan has yet been quoted below 95. CoreWeave itself has lit up instead.

Its unsecured bonds have traded below 80, and the swaps market prices a probability of default of about 64% within five years. It is also the tenant behind Lancaster, Galaxy's Helios and Richmond's Digital Drive. In 2007 terms, that is a specialist borrower in distress, priced rather than realised. The project loans may follow the tenant.

The strain shows in the funding, too. Earlier in the year a $38 billion Oracle package in Texas and Wisconsin took months to sell down, and lenders were still trying to offload the last pieces.⁷² Investors in Blue Owl's technology lending fund asked to redeem about 39% of its shares in the third quarter, as they have asked for roughly 40% in every quarter this year; the fund repurchases 5%, so about 13% of what was requested will be paid.³⁸

And at the edges of the build. The listing window that Pricing the Tenant described has narrowed further: Holtec has withdrawn its listing, and Nscale, which sought about $30 billion, is being told by investors that $16 billion to $19 billion would find more favour.⁷³ Texas's audit of data-centre connections will not report until December.⁷⁴ A hyperscaler has walked away from a neocloud: Microsoft has cancelled its deal for Nscale's Essex site, which cannot get grid power until the 2030s.⁷⁵

What has not happened is the August moment. No default, missed coupon or covenant breach has been reported, no fund has been frozen outright and no funding line has been pulled.

So there are two clocks here too. On the credit clock, the complex has moved from June to about July; high-grade paper is repricing, but funding has not seized. On the equity clock it looks more like October. In October 2007 the stock market made its high on narrowing leadership while credit was already breaking. This month the semiconductor index fell about 16% from its peak while the S&P 500 fell 2%. My judgement is that the equity clock is running ahead of the credit clock, as it did then. The question is not whether AI is subprime. It is whether weakness in the marginal financing can change the behaviour of much stronger participants, even while they remain solvent.

The Response

My own view, and it is a judgement rather than a measurement, is that a disorderly break in AI credit would not stay in AI credit, and that the authorities understand this. Treasury has expanded its buybacks of longer-dated debt. Changes to bank leverage rules have increased the capacity of the largest banks to hold and intermediate Treasuries. The Federal Reserve maintains ample reserves.⁷⁶ None of this is an announced policy to cap yields, and talk of a new Fed-Treasury accord remains talk until the authorities say otherwise. But the direction of travel is towards keeping long-term borrowing costs contained, which is another way of saying financial repression. The macro letter, coming later this week, takes that thought further.

There is an irony here worth noting. Pension and insurance money, steered by regulation towards investment-grade paper, is now financing risk that is, in substance, venture. That looks very much like the conscripted capital HindeSight has written about for years, in a new uniform.

The strongest counter deserves its place. The International Monetary Fund's October Global Financial Stability Report names an AI-led market correction as one of three salient risks and flags leverage, concentration and circular financing in the sector, but finds the global banking system resilient in aggregate under its severe stress test.⁷⁷ Even there, the tail is growing; under the IMF's adverse scenario, banks holding 22% of global banking assets would be weak, with capital ratios below 7% after the stress, up from 18% a year earlier. The largest hyperscalers can absorb large losses. And contagion through expectations, however plausible, is not yet visible in defaults.

The Bond Market, Briefly

One market relationship deserves watching above the others.

In September, chip stocks rallied even as the 10-year Treasury yield climbed from 4.79% on 1 September to 5.29% at the month's end, peaking at 5.31% on 5 October.⁷⁸ Day to day, they still behaved as rate-sensitive assets should; on days yields rose, chips tended to fall. In October that link broke. The yield fell to 5.09% by 16 October, even after hotter-than-expected September inflation briefly pushed it up, and the semiconductor index fell about 15% anyway. Lower rates no longer rescued the chips. In credit, the divergence was plainer still: as the sovereign's cost of borrowing fell, spreads on CoreWeave's bonds widened by about 300 basis points. The sovereign's cost of borrowing fell while the private borrower's rose.

That divergence has several possible explanations, and they are not mutually exclusive. The bond market may be reassessing growth and the investment cycle. It may be unwinding stretched positions. Or it may be anticipating changes to Treasury issuance and market plumbing. Each implies something different for equities, credit and the dollar, and the macro letter takes up the policy question.

For this series, one point is enough. A Treasury rally does not mean capital is getting cheaper for the businesses that need it. If Treasuries keep rallying while AI credit keeps widening, the market is telling us that private investment conditions are deteriorating, whatever the yield on the risk-free bond.

The Bill Arrives First

Strip away the detail and Part II comes down to sequence.

The bill arrives first. Construction costs rise because the inputs are scarce. Delay adds carrying cost and ages the chips. Interest starts on schedule. The revenue starts when the power does, and the power, as Part I showed, runs on a timetable measured in years.

Each layer of the financing has passed that timing risk to the next. The customer signs a contract that starts on delivery. The neocloud borrows against the contract. The developer borrows against the lease. The bank sells the loan to an insurer. Everyone has protected themselves against the risk except the person who ends up holding it, and the people holding most of it are those most dependent on a few names, CoreWeave chief among them.

The market has begun to notice. It is charging different prices for the same buildings depending on who pays the rent, and it is closing the door on new equity. On 28 October, a clause expires. Whatever happens that day, the clause is a reminder that lenders have written grace into these structures on the assumption that the revenue is late, not lost.

Part III, Too Cheap, asks the last question. Suppose the buildings are finished, the power arrives and the customers come. Who gets paid when intelligence keeps getting cheaper?

Ben Davies
Co-founder · Hinde Capital · London, 19 October 2026