Borrowed Time

Ask most market participants to narrate the mortgage crisis and the account arrives in a straight line. Underwriting standards were abandoned. Loans were written that should not have been written. House prices fell. Borrowers found themselves owing more than the collateral was worth, stopped paying, and the securities assembled on top of those loans were destroyed.

Prices fell, therefore borrowers defaulted. The account is compact and intuitive. It is also out of sequence, and the same sequencing error is now being applied with equal confidence to artificial intelligence.

What powered the subprime complex was never the price of houses. It was the rate at which that price was changing. The signature instrument of the period, the 2/28 and 3/27 hybrid adjustable-rate mortgage, was not designed to be repaid on the terms printed on the note. It was designed to be refinanced before those terms took effect.

The borrower accepted a suppressed introductory rate for two or three years. Both parties understood, without either quite saying it, that the loan would not survive to its reset. Appreciation would manufacture equity where the borrower had contributed none; that equity would collateralise a new loan; the new loan would carry a new teaser rate; the clock would restart. The mechanism was a treadmill and appreciation was the motor. Roughly four in five of the subprime hybrids originated in 2003 had been refinanced out of existence before the end of 2006.

The timing is the point. National home-price appreciation did not collapse in 2006. It slowed. Year-over-year gains that had printed in the mid to high teens through 2004 and into early 2005 began bleeding lower, still positive, still green on every chart, simply less green than before. Nominal prices stood at record highs and were still climbing. With prices at their peak and rising, subprime delinquencies inflected upward.

Read that through the documents rather than the price series and the sequence becomes obvious. Nothing in any of those loan files had been breached. Rates had not reset. Payments were current until they were not. Every written term was performing. What had failed was a condition that appeared in none of the documents: the appreciation rate required to make the refinancing work.

This is why the standard account is, in Didier Sornette’s characterisation, mechanically correct and causally backwards. It treats the fall in house prices as exogenous, as though the decline arrived from outside the system. It did not. The slowdown was produced by the structure itself. The structure required prices to keep rising at a rate sufficient to refinance its way past its own reset calendar, and no series sustains that indefinitely.

When the required rate stopped being delivered, growth followed it down through zero, negative equity spread inward from the margin, and the defaults widely attributed to falling prices had in fact begun roughly a year earlier, while prices were still rising.

Every financed structure carries two sets of conditions, and only one of them is in the file.

The first set is documented. Leverage tests, debt-service coverage ratios, minimum liquidity, change-of-control provisions, cross-default triggers. These are drafted, negotiated, priced, monitored quarterly and reported. When one is breached something happens. A waiver is sought, pricing steps up, an acceleration right vests, a rating is placed on watch. The documented covenant exists in order to be observed.

The second set is written nowhere. It is the rate of growth the structure requires in order to keep performing, and it exists whether or not anyone reduces it to language. We refer to it throughout this note as the implied covenant.

It is present wherever financing embeds a growth assumption. A reset that presumes a refinancing has an implied covenant: the collateral must appreciate fast enough to support the new loan. A capacity commitment sized to an extrapolation has one: demand must arrive on the extrapolated schedule. A borrower servicing obligations out of successive equity raises has one: each raise must price above the last by enough to cover the interval’s burn.

The implied covenant has four properties that make it the most dangerous term in any credit agreement.

It is satisfied by sustained growth, not by a high level. A company can print record revenue and be in breach. Growth of 40% is a triumphant press release and a covenant failure if the capacity contracted for assumed 70%. The headline and the requirement diverge, and only the requirement determines whether the financing performs.

It is never monitored, because it was never drafted. No covenant compliance certificate reports it. No agent bank tests it. There is no quarterly calculation and no threshold to fail. It can be breached for a year without a single party to the transaction being formally notified, because there is no mechanism by which notification would occur.

It is breached before any documented term. This follows from the first two properties. Documented covenants are calibrated to current cash flows and current asset values, both of which are lagging quantities. The implied covenant is calibrated to a forward growth rate, which deteriorates first. In the last cycle the gap ran to roughly a year.

Compliance with every documented term is not evidence of health. During the interval between the silent breach and the visible one, full compliance is precisely what one would expect to observe. The reported data look sound because the reported data are measuring the wrong conditions.

We refer to that interval as borrowed time. Inside it, revenues print records, growth remains positive, ratios are met, management is congratulated and paper trades at spread. The structure is already impaired. It has simply not been informed, because nothing in its documentation is capable of informing it.

How dangerous borrowed time proves to be depends on what is riding on top of the structure. An equity multiple is roughly linear in expectations. It can deflate gradually and reflate later, which is broadly what happened to the survivors of 2000. A leveraged credit structure is negatively convex: it earns a fixed coupon while conditions hold and absorbs open-ended loss when they do not, and its documented covenants are step functions rather than smooth curves. The move from compliance to breach is not a slope. It is a cliff, and it arrives long after the implied covenant has already gone.

The practical instruction that follows is simple to state and almost never carried out. For any structure, ask what rate of growth it requires rather than what rate it is currently delivering, and monitor the difference. We refer to that difference as growth headroom. Where headroom is positive, documented covenants are the binding constraint and conventional credit analysis applies. Where headroom has gone negative, the documented covenants are measuring a structure that has already failed.

The remainder of this note applies that single test across the AI financing complex.

Open any AI note, bull or bear, and observe what it argues about. It argues about levels (billions of revenue, gigawatts of capacity, the size of the addressable market) and about growth rates (whether growth is 200% or 150%, whether enterprise adoption is inflecting, whether tokens per minute are rising). The bears contend the levels cannot be sustained. The bulls contend the growth justifies them.

Neither camp is asking what growth rate the financing requires. Yet the architecture assembled over the past twenty-four months is, in substance, a set of positions on that requirement being met, presented as positions on the level.

This matters because the analogy in general use is the wrong one. Asking whether AI is a bubble like dot-com produces dot-com answers: perhaps the leaders survive, perhaps multiples compress, perhaps the market absorbs a 50% drawdown and recovers. That framing is a category error.

The year 2000 was an equity-multiple event. Excessive optimism was priced into companies with thin balance sheets and almost no debt. The losses were substantial but they were equity losses, and equity is patient capital that can be marked down and held. Equity carries no covenants at all, documented or implied. The AI build-out is financed differently. It is funded not principally by issuing expensive stock but by contracting future cash flows and borrowing against hardware: take-or-pay capacity agreements, GPU-collateralised term loans, off-balance-sheet vehicles, and asset-backed notes placed with insurers and banks. Every one of those instruments carries both sets of conditions. That is not the architecture of 2000. It is the architecture of 2008.

Beneath the 2000-versus-2008 question sits the misclassification doing the real analytical damage. The market is pricing AI as a technology cycle. Its actual anatomy is that of a credit-driven real estate cycle, which is precisely why the 2008 mechanics transfer, and the two break for different reasons.

Technology cycles are driven by innovation and adoption. Their principal risks are obsolescence and competition. They live or die on whether the product is wanted, and they can de-rate slowly as the future is repriced.

Real estate cycles are mechanical. Leverage, hard assets, occupancy. Debt-financed construction at scale, long leases described in other language, and construction lags long enough to guarantee that supply arrives after demand has already turned.

Every feature of the AI build reproduces property development in different vocabulary. A data centre is a structure on entitled land, financed with debt secured against that structure, let to tenants on contracted terms. Translate the language and the take-or-pay becomes a lease, the remaining performance obligation becomes a rent roll, and the frontier lab becomes an anchor tenant. This is not a software business that happens to own servers. It is a landlord that happens to compute.

Real estate cycles break the same way each time. Not because demand collapses, which it rarely does, but because the rate of demand growth falls below the rate the completed supply was financed against. The implied covenant again, in the one asset class where a century of literature has already documented it.

Credit, for its part, does not de-rate gently. It refinances or it seizes. Within the complex the mechanics diverge by borrower. Pure-play neoclouds carry non-recourse debt written against specific tenant contracts; when a tenant cannot pay, they do not compress, they stop. Hyperscalers fund from corporate bonds and operating cash flow, so a tenant default produces impairment and margin compression rather than seizure. Oracle occupies the space between: corporate-funded, but concentrated to a degree that removes most of the protection. As of July, that distinction is no longer theoretical.

Set the compute-scarcity narrative aside and examine what these agreements are.

A frontier lab signs a contract committing it to pay a counterparty tens of billions of dollars, across several years, for computation it has not yet consumed. The counterparty (Oracle, CoreWeave, a hyperscaler) records that commitment as backlog and borrows against it, raising debt to pour concrete and rack silicon.

Reduced to its skeleton, the arrangement is a loan. Capital is advanced in kind, in the form of a building full of chips delivered up front, against the borrower’s commitment to pay it back, with interest and amortisation embedded in the contracted rate. The data centre is the collateral. The contracted payments are the debt service. The structure performs only so long as the borrower can keep funding those payments which, for a company without profits, means only so long as it can keep raising capital.

This is the substance of our argument that the correct analogue is 2008 rather than 2000. AI capital expenditure is not a capital budget. It is a loan book. Leases in form, debt in substance. The capex is funded principal; the contracted backlog disclosed in the filings is the receivable.

We therefore read record capex disclosures differently from the way they are generally reported. When the financial press treats them as evidence of confidence and proof of demand, we read them as origination volume. Each gigawatt of committed build is credit extended to whichever tenant signed the take-or-pay beneath it, and the quality of that credit is precisely the quality of that tenant. What is being reported as revenue growth is more accurately described as underwriting growth.

Every one of those extensions carries an implied covenant. It is the tenant’s ability to keep funding contracted payments out of financing, which reduces to the tenant’s ability to keep raising at ascending valuations. Nobody has written that condition into any credit agreement in this complex. It nonetheless governs the entire book, and Section V is an attempt to measure it.

Every credit cycle produces a representative borrower: the one that could always refinance and could never repay. In this cycle we identify that borrower as OpenAI.

OpenAI has committed to purchase compute on a scale without corporate precedent, through multi-year take-or-pay capacity agreements whose aggregate obligations exceed $600bn through 2030. Against those obligations sits an operating business that does not earn a profit. The company reported a net loss of roughly $38.5bn for 2025 on approximately $13.1bn of revenue, and is now running near $2bn of revenue per month. Cash burn is projected at roughly $27bn in 2026 and toward $63bn in 2027 under the renegotiated Microsoft terms. Revenue is real, large, and growing quickly. It does not cover the company’s own burn and it is nowhere near covering the contracted payments.

Those payments are therefore not serviced out of earnings. They are serviced out of financing, and financing for a borrower in this position is available on a single condition: that each round prices above the last.

For OpenAI the up-round is not a measure of progress. It is the funding event itself, the mechanism by which the previous period’s commitments are paid and the next period’s become signable. The step-up is the operating cash flow.

Which gives us a rare opportunity. Most implied covenants must be inferred. This one can be computed, because the borrower’s entire liquidity function is a single observable ratio.

The implied covenant is the required step-up. A company funded by its own appreciation is solvent in proportion to how fast its mark is rising, not how high it stands. The burn is a schedule, and the schedule is accelerating; it is indifferent to the size of the last round. Each successive phase of compute expansion requires a larger absolute injection than the one before. The condition the structure requires is therefore a multiple, not a valuation, and the historical record tells us what that multiple has been.

Measured as a level, the sequence is the most remarkable appreciation in the history of private markets: roughly $86bn in early 2024, then approximately $157bn in October 2024, $300bn in March 2025, $500bn at the October 2025 employee tender, and $852bn at the close of the Series F on 31 March 2026.

Measured as the ratio the structure actually runs on, the same sequence inverts. Round-over-round step-ups ran 1.83x, 1.91x, 1.67x and 1.70x. Those four observations are the implied covenant, revealed by performance rather than by drafting. The structure has required something in the region of 1.7x, sustained, to carry each interval’s burn into the next.

The reported public-offering target above $1trn implies roughly 1.23x.

That is the breach. It is not a decline, a collapse, or a repricing of the theme. The valuation would still be a record. Every documented term everywhere in the complex remains in compliance. What has failed is a condition that appears in none of them, and it has failed by a wide margin at the one mark the structure cannot negotiate.

Private marks are lumpy by construction, being negotiated, episodic, and set by a small number of insiders, so no individual step is decisive. Note also that the mark is set by the parties who need it to rise. The Series F opened on 27 February 2026 at a $730bn pre-money valuation and closed on 31 March with $122bn of committed capital at $852bn post-money, anchored by Amazon at $50bn with Nvidia and SoftBank at $30bn each. These are the same counterparties whose compute the proceeds are contracted to purchase, and the Amazon commitment arrived alongside an agreement for OpenAI to spend approximately $100bn on AWS over eight years. The mark rises because capital came in; more capital comes in because the mark rose; the capital returns to the investor as contracted spend. Appraiser, lender and vendor are substantially the same parties. That is exactly why the private sequence held at 1.7x for as long as it did, and exactly why the public mark is the one that tests it.

The erosion in required-versus-delivered is not arbitrary either. Two structural forces are attacking precisely the revenue growth the next mark requires. The first is token efficiency: the industry’s own central optimisation project, which routes trivial queries to cheaper models and trims reasoning tokens, has eroded the tokens-per-task tailwind that had been padding revenue. The second is open-weight competition from China. On OpenRouter, Chinese-origin models held roughly 61% of token volume among the top ten models through the first half of 2026, and combined Chinese-provider traffic reached approximately half of all platform tokens, against under 2% in late 2024. The US big three fell from roughly 70% of platform token share in June 2025 to roughly 30% a year later.

One qualification, routinely omitted. Token share is not revenue share. Open-weight models capture volume because they are 60% to 90% cheaper; the premium closed models retain a disproportionate share of spend on high-value agentic and reasoning workloads. The commoditisation is real but concentrated in the low-value middle. What it attacks is not the revenue base directly. It is the growth rate of that base, and the growth rate is what the covenant is written on.

Here the bulls make their strongest objection. The labs burn cash, but so did Amazon, and so did most great compounders during their infrastructure phase. Burn is investment. Today’s borrower is tomorrow’s cash generator.

We consider this half right, and the error is not in the analogy. It is in treating the frontier labs as a single credit. There are at least two, they carry different implied covenants, and over the past two quarters they have separated to a degree we think is the most important development in the complex.

Anthropic. The company closed a $65bn Series H on 28 May 2026 at a $965bn post-money valuation, above OpenAI, having carried a $380bn valuation as recently as February. Run-rate revenue crossed $47bn in early May, up from roughly $9bn at the end of 2025 and $14bn in February. Revenue is predominantly enterprise, contracted and recurring. The company filed confidentially for a listing on 1 June, is reported to be tracking an October window, and reporting indicates it will arrive with a profitable quarter, which OpenAI cannot show.

Its liabilities are also wrapped, which changes the covenant in a way that deserves precision. In June, Apollo and Blackstone closed a $35bn private credit facility through a special-purpose vehicle that purchases Google-designed, Broadcom-manufactured TPUs and leases them to Anthropic, deployed at Fluidstack-operated sites. Google backstops lease shortfalls across five US data centres. Broadcom provides a deficiency guarantee covering $30bn of senior notes, $6bn of Class A1 sold to banks and $24bn of Class A2 sold to institutions, lifting those tranches close to Broadcom’s own investment-grade level. The hardware sits in the SPV and off Anthropic’s balance sheet.

What the wrap does is transfer the implied covenant from the tenant to the guarantors. The condition is no longer that Anthropic keep growing fast enough to service the lease. It is that Google and Broadcom remain willing and able to stand behind it. That is a materially better covenant, held by materially better credits.

OpenAI. The unwrapped borrower, and on our work the weakest and most volatile credit in the ecosystem. Revenue is roughly half Anthropic’s on a run-rate basis and mixed toward consumer. Burn is projected near $27bn in 2026 and $63bn in 2027, against a revenue base of roughly $25bn. We see no path to positive cash flow this decade. Its implied covenant remains where it started: with the borrower, on the borrower’s own ability to keep clearing a rising mark.

The critical distinction is look-through. Underwriting Anthropic means ultimately underwriting Google and Broadcom. Underwriting OpenAI provides nothing to look through to.

Two things cut against our own framing, and both are worth stating plainly.

The first concerns the wrap. Broadcom’s guarantee is not free. S&P downgraded Broadcom on the leverage created by extending it. This is the monoline dynamic of the last cycle in contemporary dress: the guarantor’s own rating is the resource being consumed to manufacture credit quality elsewhere, and it is finite. The same platform is targeted at more than 20GW of capacity for multiple labs including OpenAI through 2028. Investors underwriting the wrap should underwrite the wrapper, and should note that the wrapper now carries its own implied covenant regarding how many such guarantees it can extend before its rating becomes the binding constraint.

The second concerns our thesis directly. Anthropic’s step-up from $380bn to $965bn in three months is roughly 2.5x, which is not merely positive but well above any plausible requirement, at the same moment OpenAI’s compressed toward 1.23x. This is a direct counter-example to any reading of our argument as a claim about the sector. We do not make that claim. The failure we identify is borrower-specific, and the divergence between the two labs is evidence that private capital is now discriminating on credit quality rather than repricing the theme. That is the more useful reading, and it is the more dangerous one for the counterparties concentrated in the weaker name.

Moving from the borrowers to the lenders, the credit extended takes a specific reported form: remaining performance obligations, or RPO, being contracted revenue under signed agreement. Backlog. The market reads backlog as visibility, as demand locked in and pulled forward.

The aggregate is now roughly $2.3trn across the four large platforms. Oracle disclosed $638bn in its most recent quarter, up 363% year over year. Microsoft’s RPO reached $678bn, up 84%. Google Cloud backlog grew to approximately $514bn from roughly $460bn in the prior quarter. AWS disclosed $496bn, up more than 100%. Each quarter’s increase is presented as confirmation that demand is real and the build justified.

An RPO is not an asset in any liquid sense. It is a forward promise of future payment in exchange for future compute, and a multi-year promise is worth precisely the creditworthiness of whoever made it. Where the counterparty is investment grade and cash generative, backlog is what it claims to be: high-quality visibility, merely deferred. Where the counterparty is a pre-profit company that can pay only by continuously refinancing its own equity valuation, backlog inherits that counterparty’s implied covenant in full. It is a subprime commitment, used to justify substantial un-depreciated capital expenditure, and reported to shareholders as structural strength.

Two disclosures make the point better than any inference we could offer.

The first is Microsoft’s own. The company reported that commercial RPO grew 26% excluding OpenAI, in line with historical seasonality, against 99% including it. The organic book and the concentrated book are different businesses carrying different conditions, and management has effectively said so in the filing.

The second is Oracle’s, delivered by its rating agency rather than its management. S&P cut Oracle to BBB- on 9 July 2026, one notch above speculative grade, citing as a critical vulnerability that OpenAI accounts for roughly half of the $638bn backlog and that Oracle would be left holding data centre leases it may struggle to exit or re-lease should OpenAI face financing difficulty. That is our Section IV argument, reproduced almost exactly, by the agency that rates the paper. It is also the moment at which an implied covenant became a documented one: the condition moved out of nobody’s file and into a published rating action.

In economic substance, the platform has extended a concentrated and largely unsecured credit facility to tenants with no independent operating income. Should those tenants fail, the backlog does not convert into revenue. It converts into non-cash impairment, leaving the corporate balance sheet holding the fixed costs of customised, rapidly depreciating assets.

Take-or-pay contracts are the foundation of the structure, the collateral that makes the borrowing possible. They are not a passive constraint sitting outside the system. The competitive equilibrium produces them, and produces them at progressively larger scale. Justifying the next gigawatt of spend requires the next gigawatt of backlog, so each platform pushes its tenants to commit further forward. The $2.3trn is not a pre-existing ceiling on the arms race. It is the arms race’s paper trail. The contracts are the collateral and the race is the demand for more collateral.

Once signed, that collateral must be converted into physical infrastructure before any cash arrives. That conversion drives the capex machine, and the machine now consumes cash faster than the backlog can validate it.

The four largest platforms are guiding to roughly $725bn of capital expenditure in 2026, up approximately 77% from about $410bn in 2025, with analysts modelling above $1trn in 2027. Bank of America’s more expansive definition puts 2026 above $860bn and 2027 near $1.2trn. Both series describe the same shape: growth that remains extraordinary in level while the increment slows. On the four-platform series, growth runs roughly 51%, 81%, 77% and 52% across 2023 to 2027. On Bank of America’s, the 2027 growth rate falls to 38% from 80%.

A recursion sits beneath these figures. Platform capex is not primarily a response to AI demand. To a substantial degree it is the demand. A large share of the labs’ revenue is platform spending recycled through cloud credits, compute commitments and equity-funded consumption. Nvidia’s revenue is platform capex. Neocloud revenue is platform capex with leverage applied. Net out the spending and the demand it manufactures, and organic capex-independent demand is a fraction of the headline.

Which means the spending equilibrium itself carries an implied covenant, and it is the most consequential one in this note.

The condition was market approval, and it has failed

The capex arms race is a Nash equilibrium, but a conditional one. It holds only while the market treats each incremental dollar of spending as a call option on growth. In that regime the dominant strategy for every operator is to spend, because the alternative is to be the participant who blinked. Mutual escalation is stable precisely because it is rewarded.

The implied covenant of that equilibrium is therefore not financial at all. It is behavioural: the market must continue to reward the spender. Nothing in any credit agreement references it. Every platform’s capital plan depends on it.

It was breached on 22 July 2026.

Alphabet reported a strong quarter, with cloud revenue up 82% to $24.8bn, and raised 2026 capex guidance to a range of $195bn to $205bn from $180bn to $190bn, warning that 2027 would increase significantly. The stock fell 7.1% the following session. Microsoft fell 2.2% and Meta 3.4% in sympathy, and the Nasdaq fell 2.2%. In April, Meta raised its own guidance and fell 6% after hours. The commentary was unambiguous about what had changed. One investor characterised the shift as the market having moved from more being better to less being better. Another described the quarter as a transition from AI hype to monetisation discipline.

Be clear on what this is and is not. It is not a collapse in the sector, a repricing of AI, or a verdict on demand. It is the removal of the specific payoff that made mutual escalation stable. Spending is no longer applauded. It is being underwritten, line by line, against a return the platforms have not yet demonstrated.

What has not yet occurred is the second half: a platform announcing a genuine reduction and being re-rated for it. When that happens the payoff matrix rewrites completely, and because this is a coordination equilibrium the first mover rewarded for cutting supplies every other operator with both cover and incentive. Discipline would then cascade as quickly as the spending it replaces.

What the repricing does to the contracts

The severity lies in what it does to the paper. In the boom regime a signed take-or-pay is an asset to every party that touches it: forward demand for the platform, bankable backlog for the neocloud, collateral for the lender. In the repriced regime the identical contract becomes a liability for all three simultaneously.

The lab cannot fund the payments it locked in. The platform holds a receivable from a visibly distressed counterparty. The neocloud services debt against facilities financed on a contract now worth less than the debt secured against it.

This is negative convexity wired into the demand side: the same instrument is an asset while conditions hold and a liability once they do not, and the transition between those states is a repricing of belief rather than a change in the hardware. Nothing physical must break. The market need only change its assessment, and over the past six weeks it has begun to.

Every coordination cascade requires a first mover. Our original framing here has been partially overtaken by events and requires revision.

We argued that the first to cut would combine the best information, the credibility to reframe a reduction as discipline, and sufficient balance-sheet capacity to be rewarded rather than punished. We nominated Meta on the grounds that Zuckerberg holds dual-class control, has run this playbook before, and lacks a public cloud into which to monetise the capacity. We are downgrading that call materially.

Meta has not blinked and is building the monetisation channel whose absence underpinned our thesis. On the second-quarter call Zuckerberg stated an intention to grow a large business serving large customers, which is a compute-rental line by another name. The company narrowed 2026 guidance to $130bn to $145bn by raising the low end, and it funded that by cutting roughly 8,000 positions, about 10% of headcount, explicitly framed as a trade against the infrastructure budget. Free cash flow fell 91% and is at its lowest since late 2022. Meta cut people to protect compute, which is the opposite of the move we anticipated.

Alphabet now looks closer to the constraint than we allowed. We wrote that Google would not blink because it fabricates its own accelerators at a structural cost advantage and holds a large cloud backlog. The cost advantage is real and the backlog has grown to roughly $514bn. But long-term debt rose 111% to $98bn in the first half of 2026, the company raised $80bn of equity in June, and it turned free-cash-flow negative in the second quarter for the first time since listing, at negative $5.9bn. It is also the name the market has already punished for spending. Structural advantage is not the same as balance-sheet indifference.

Oracle cannot cut, and its stress has already surfaced exactly as we expected, as a credit event rather than a strategic decision. Fiscal 2026 capital expenditure rose 162% to $55.7bn against roughly $67bn of revenue. Free cash flow was negative $23.7bn. Total debt stands near $167bn, with roughly $260bn of data centre leases signed. S&P expects leverage in the mid-4x area in fiscal 2027 against a 4x threshold, and the stock has fallen roughly 60% to 65% from its peak. Fiscal 2027 capex is guided to $90bn to $95bn against S&P’s prior forecast of $60bn. Oracle is one notch from losing investment grade while guiding spending higher.

Amazon faces the same direction of travel, with long-term debt up 81% to $119bn in the first quarter and capex guided near $200bn.

Our revised view is that the first genuine reduction is more likely to be forced than chosen, and more likely to come from the name with the least tolerance for further leverage than from the name with the most freedom to act. The marginal gigawatt remains the most discretionary line on a leveraged, hard-asset, refinance-dependent balance sheet.

Assume the sequence runs. OpenAI proves to be the weak credit, belief reprices, the capital window narrows, and a platform cuts the marginal gigawatt to protect its rating. The question is who holds the exposure.

OpenAI is the largest customer, directly or one counterparty removed, of very nearly every name selling into the AI build. It accounts for roughly half of Oracle’s $638bn backlog. It is the dominant tenant behind SoftBank’s Stargate commitments, and SoftBank’s total exposure is projected near $65bn. Reporting indicates it is a substantial share of Microsoft’s book, and Microsoft has disclosed the organic-versus-including-OpenAI split precisely because the difference is material.

This is the structure that made 2008’s senior tranches lethal. Thousands of individual mortgages, geographically dispersed and apparently independent, each with its own documented covenants, until it emerged that all of them shared a single undocumented one. Here the shared condition is not national home prices. It is whether one borrower can keep clearing a rising mark. Chip stocks sold off on 26 June when reporting indicated the listing might slip to 2027, which is the shared condition revealing itself in miniature.

That correlated exposure is now being distributed rather than warehoused. CoreWeave closed its DDTL 5.0 facility on 15 May 2026, at $3.1bn, issued through a bankruptcy-remote financing subsidiary and priced at SOFR plus 450 with a maturity near five and a half years. It carries Ba2 from Moody’s and BB+ from Fitch, and the company describes it as the first publicly syndicated HPC infrastructure-backed financing vehicle, expressly built to enable secondary market trading. The disclosed underlying capacity serves two large non-investment-grade customers, whom CoreWeave has not named. Ignore the confident attribution circulating in the market. The identities are inferred, not disclosed.

Against our own thesis: the transaction was meaningfully oversubscribed and priced 50 basis points tighter than initial discussions. Credit appetite for this paper was strong three months ago. That is a fact the bear case has to accommodate rather than ignore.

The more significant development came on 11 August 2026, one day before this note, and it is the reason the framework in Section II is more than a metaphor. CoreWeave closed a $2.6bn DDTL 5.5 facility. Unlike prior facilities, which were backed by customer contracts extending through the maturity of the debt, DDTL 5.5 carries an approximate five-year maturity while its underlying customer contracts average approximately three years.

Read that as a covenant statement. The debt now outlives the cash flows contracted to service it. The gap must be closed by re-contracting the capacity at the end of year three, in whatever market exists at that point, at whatever price then clears. The condition required for repayment is therefore that demand at year three supports a rate covering debt service on a five-year obligation.

That is the 2/28 structure exactly. It is also, unusually, an implied covenant that has been disclosed. The issuer has published the mismatch, the agencies have rated the paper knowing it, and the facility has been syndicated into the broad credit complex on those terms. This template gets tested first and hardest in any slowdown, and the year-three re-contracting outcomes are the cleanest test of everything in this note.

Following the refinancing chain to its terminus leads to the public market.

Private capital is deep but bounded. SoftBank, the sovereign funds, the platforms and the megafunds can each absorb a round or two. Burn measured in tens of billions annually and compounding eventually exhausts them, and at that point the only pool large enough to sustain the treadmill is the one in which index funds and the retail bid reside.

In this structure the listing is not an exit. It is the refinancing of last resort, the deepest teaser into which the edifice expects to roll once private rounds can no longer carry the burn. It also carries a constraint the private rounds did not. Reaching the public pool requires an S-1, and the S-1 discloses precisely the fragility that made the refinance necessary. The document that unlocks the capital is the document that prices the risk. More precisely: the S-1 is the instrument by which an implied covenant becomes public, and the borrower whose implied covenant is under strain has the strongest reason to defer filing it in effective form.

The evidence that this constraint binds is now on the record. OpenAI filed confidentially on 8 June 2026, one week after Anthropic, and simultaneously cautioned that it had not decided on timing and that a listing may be a while. On 25 June, reporting indicated the company was leaning toward 2027, with the chief financial officer advocating that timeline internally on the basis of cash burn, the compute commitments, and public-company readiness. Sam Altman is reported to regard any valuation below $1trn as a non-starter. Bankers presented the choice plainly: reduce the expectation and list sooner, or defer and protect the number.

The arithmetic the deferral obscures is the covenant computed in Section V. Moving from $852bn to a target above $1trn is a step-up of roughly 1.23x against a demonstrated requirement near 1.7x. It must accomplish two incompatible objectives simultaneously: clear at a level the public market will pay, and raise enough to cover a burn running toward $63bn in 2027 alone. The price that clears the market does not retire the burn. The price that retires the burn does not clear the market.

Two contextual observations sharpen this. The first is that the public market has just demonstrated its willingness to reprice a marquee listing: SpaceX listed in June 2026 and gave back roughly a third of its initial gains within days. The second is that SoftBank, whose $40bn bridge facility matures in March 2027 and was structured on the assumption that a listing would help repay it, has been visibly working to reduce its dependence on that event.

Which brings us to a correction we owe our own earlier work. We previously wrote that SoftBank was struggling to raise a margin loan against its OpenAI stake, and characterised this as the co-signer lacking a co-signer. That is no longer accurate. Having been cut to $6bn over collateral-valuation concerns, the facility was restored to $10bn once SoftBank added a corporate guarantee giving lenders recourse to the group balance sheet, and it closed on 6 August 2026 with Goldman Sachs, JPMorgan, Mizuho, Apollo and Sumitomo Mitsui as mandated lead arrangers.

Take the correct inference here rather than the convenient one. The loan cleared, which is evidence of available credit. But it cleared only after the collateral was supplemented, having been rejected on the collateral alone. Lenders were willing to lend against SoftBank. They were not willing to lend against the OpenAI stake by itself. That is the market declining to underwrite the borrower’s implied covenant, expressed through a different instrument, and arriving at the same answer as the deferred listing.

The trigger is narrow and specific. The next mark fails to clear at the required step-up. Not a collapse, merely a shortfall against a condition nobody has written down. This is the 2006 dynamic replayed, with the requirement failing while the level is still climbing.

From that point the sequence runs in order. We note in brackets where a stage appears to us already in progress.

1. The terminal refinance prices short. The step from $852bn toward $1trn either does not clear or clears at a level insufficient to retire the burn.In progress. The deferral is the observable form of this.
2. The borrower conserves cash by reducing compute commitments. A reduction against a take-or-pay obligation converts an implied covenant into a documented one, constituting a breach at whichever provider’s debt is collateralised by that commitment.
3. The breach lands first on the neoclouds. CoreWeave, Lambda and Crusoe exist economically as the spread between borrowed money and resold compute. A neocloud is less a business than a spread trade with no balance sheet to warehouse the risk; when the spread inverts it is insolvent by definition rather than by decision. Oracle takes the next impact: corporate-funded but concentrated, bleeding rather than seizing. A platform can absorb a missed payment out of Search, Windows or Retail. A neocloud has no second cash flow.Partially in progress. Oracle’s downgrade and negative free cash flow are the first-order version of this stage, arriving ahead of any borrower default.
4. Credit freezes across the complex. RPO is repriced from forward demand to counterparty exposure, GPU-backed notes cannot roll, and the originate-to-distribute machine seizes.Not yet. Appetite remained strong at the May and August syndications, and this is the stage we watch most closely.
5. Equity is decimated. Negative convexity operates in reverse: capex is repriced from option to cost and multiples compress across the chain.Partially in progress, and name-specific rather than general. Oracle has fallen roughly 60% to 65% from its peak; the platforms have de-rated modestly on capex guidance.
6. The best-capitalised survive and acquire. Stranded facilities change hands at distressed valuations, and the leases that must endure are backstopped by whoever retains the capacity.

One concession on timing, and it is not a small one. We do not know when. Borrowed time, being the interval between the silent breach and the documented one, can extend well beyond the patience of any short position, and three developments can extend it further: a larger-than-expected private round, a sovereign or strategic backstop that postpones the terminal refinance, and the platforms’ continued capacity to lever up against the very backlog described above.

The last is the most powerful near-term stabiliser, because the platforms possess real balance sheets, real cash flows and real access to debt markets. It is not unlimited. Alphabet’s long-term debt has more than doubled in six months, Amazon’s is up 81%, Oracle sits one notch above speculative grade, and the guarantor underwriting the largest private-credit wrap in the complex has itself been downgraded for doing so. Documented covenants exist across all of these, and unlike implied ones they are monitored.

We grant the bull case in full. Demand is real. Backlogs are expanding and reported growth rates remain extraordinary. Inference is in its infancy. The risk of underbuilding a generational platform is acute, supply is committed years forward, and credible analysts model a path toward $1.2trn of annual capital expenditure. We take the case seriously. It does not rescue the structure.

Note the form of those claims. Backlogs are a level. Inference ramping is a level. Supply growth is a rate. None of them addresses the gap between the rate delivered and the rate required. Our thesis does not require demand to fail. It requires growth to fall below the requirement, and a structure this levered breaks on that shortfall alone. Housing demand was real in 2006. The financing failed on the shortfall, not on the level.

Four rebuttals warrant direct treatment. The second and fourth are the strongest arguments against us.

The fortress balance sheet. Platforms generate substantial cash flow, so a tenant impairment is absorbable. This misidentifies the wound. Impairment is an accounting event; margin collapse is a structural one. AI capacity carries a large fixed-cost base of depreciation, power and interest that does not flex when a tenant stops paying. Utilisation falls and operating expense does not. The operating leverage that supercharged margins on the way up destroys them on the way down. The fortress framing is also already dated: Alphabet and Oracle both turned free-cash-flow negative in their most recent reported periods, and Meta’s fell 91%.

The credit is discriminating, not deteriorating. This is the strongest objection and it is our own evidence. Anthropic re-rated 2.5x in three months, is listing with a profitable quarter, and its paper is wrapped by two investment-grade guarantors. CoreWeave’s May syndication was oversubscribed and priced tighter. SoftBank’s margin loan closed at full size. If capital were repricing the theme, none of that would have occurred. The honest reading is that capital is discriminating between borrowers with precision, which is not a bubble deflating. Our response is that discrimination is what precedes a credit event rather than what prevents one: the marginal borrower loses access first, and the complex is concentrated in the marginal borrower. But we hold this view with less confidence than the rest.

Cross-subsidisation. If AI margins deteriorate, Search and Windows carry the division. The rebuttal is the conglomerate discount. Investors own these companies for growth, not to fund perpetual losses in an adjacent division. Where AI consumes tens of billions without reaching profitability, consolidated returns on invested capital decline, and a high-ROIC compounder that becomes a low-ROIC capital-intensive operator loses its growth premium. The legacy engines are also not infinite: Search faces structural erosion, Retail runs on thin margins, Windows is mature.

Re-leasing the capacity. If a tenant defaults, the provider re-lets the facility. This is the 2006 argument that housing never loses value, restated in gigawatts. A default of this kind will not occur in isolation; it will coincide with broader deceleration, meaning capacity comes online into a softening market. Operators do not re-let into a glut, they compete on price, and the clearing price falls below debt-service coverage. S&P made effectively this point in the Oracle downgrade, flagging leases the company may struggle to exit or re-let on favourable terms. DDTL 5.5 makes the exposure explicit rather than contingent: that facility requires a re-contracting at year three regardless of whether any tenant defaults at all.

We do not disagree with the bulls about artificial intelligence. We disagree about which condition the structure is written on. They are reading the documented terms, all of which are in compliance. We are reading the one that is not in the file.

The thesis is falsifiable and the indicators are observable. In each case we are tracking the gap between the rate a structure requires and the rate it is delivering.

1. OpenAI’s next mark against a 1.7x requirement. Any print materially below 1.23x confirms the shortfall. A print at or above 1.5x, whether through a private round or a listing clearing above $1trn, materially extends borrowed time and would cause us to push out the sequence.
2. Anthropic’s listing and its aftermarket. A successful October offering near or above the last private mark, with a profitable quarter disclosed, would demonstrate that the public pool remains open to frontier labs on credible economics and would isolate the problem entirely to the weaker borrower. A weak reception would close the terminal refinance for both.
3. The market’s reaction to the first announced reduction. A multiple that expands on a genuine capex cut completes the inversion described in Section VIII. This is the single clearest confirmation available and it has not yet occurred.
4. The direction of capex guidance revisions against market reaction. Guidance has continued to rise even as the market punishes the rises. That tension is unstable and must resolve. We watch specifically whether 2027 guidance is revised down against a 2026 base.
5. Rating agency actions across the complex. Oracle at BBB- with mid-4x leverage expected in fiscal 2027, Broadcom downgraded for extending the wrap, and Alphabet and Amazon both levering rapidly. These are the documented covenants, and they are the constraint on the most powerful stabiliser available.
6. Secondary pricing on GPU-backed paper. DDTL 5.0 and 5.5 in particular. Spread widening in distributed paper is the earliest observable sign that credit is repricing counterparty risk rather than forward demand, and it marks the transition from stage three to stage four in Section XII.
7. Re-contracting outcomes on short-duration facilities. DDTL 5.5’s contracts average roughly three years against a five-year maturity. The terms achieved at that re-contracting are the cleanest available test of the entire framework.
8. Frontier-lab revenue growth against token pricing. Adoption and revenue can both grow while the growth rate erodes under efficiency gains and open-weight competition. We track revenue share rather than token share, since the two now diverge substantially.

Consensus holds that AI is a technology cycle carrying a valuation problem. Our view is that it is a credit cycle carrying a documentation problem. That distinction decides how this resolves, and it is the whole of our edge here.

If consensus is right, the downside is a de-rating. Multiples compress, capital discipline returns, the strong compound through it and the weak get bought. Painful, survivable, and broadly what 2000 delivered. If we are right, the mechanism is different in kind. Take-or-pay obligations, GPU-collateralised term loans and syndicated infrastructure paper do not compress. They perform or they stop, and the covenants that would ordinarily warn you are calibrated to conditions that fail last rather than first.

What is priced today is roughly this: demand is real, the platforms are well capitalised, and the labs will grow into their commitments. We do not dispute any of it. What is not priced is that a structure servicing itself by refinancing into growth breaks when growth falls below the rate the refinancing assumed, not when growth turns negative. Between those two points every reported metric looks fine. That interval is the trade.

The asymmetry follows from where the requirement has already missed. OpenAI’s structure has needed roughly 1.7x per round to carry each interval’s burn; the mark available implies 1.23x. CoreWeave’s newest facility matures two years after the contracts servicing it. The spending equilibrium depended on the market rewarding the spender, and that stopped on 22 July. None of these needs a default to matter, and none of them appears in a covenant test anyone runs.

Where that lands is not evenly distributed, and we think this is the most actionable part of the analysis. Equity is the wrong place to look for the convexity. The platforms absorb a tenant failure as impairment and margin compression, which is a de-rating, not an event. The convexity sits lower: in non-recourse paper written against single-tenant contracts, in the neoclouds whose entire economics are a spread with no balance sheet behind it, and in the one corporate credit that has already been downgraded on precisely this concentration. That is where the same deceleration produces a stop rather than a slowdown.

We are wrong in three ways worth naming. If OpenAI’s next mark clears at 1.5x or better, whether privately or publicly, the borrower refinances and we have simply been early on a structure that had more room than we credited. If Anthropic lists well in October, the terminal refinance is demonstrably open and the problem is one borrower rather than a complex. And if platform balance sheets keep absorbing the build without rating pressure, the most powerful stabiliser in the system holds and this runs considerably longer than we think.

That third scenario is the one that should concern anyone positioned around this. Borrowed time is not a short window. In the last cycle it ran roughly a year between the requirement failing and the file registering it, and carrying a position across that interval is expensive in a market printing records the entire way. Being right about the mechanism and wrong about the interval is the standard outcome for this trade, and it is why we have set out observable triggers rather than a target date.