The Teaser Period: Why The AI Boom Is Hitting A Reset Wall
Having laid out, in July, the structural diagnosis that most of the market still refuses to confront: the AI boom is not a technology cycle. It is a credit-driven real-estate-like cycle whose financing architecture depends on the second derivative; the appropriately-named 'Groundbreaker' website has just dropped his next insightful note on what may be the trigger for the market to wake up to the ugly reality beneath the surface of the AI dream.
Trillions in signed compute commitments come due in 2027–2028. The underlying mechanics reveal how the AI boom ends, and when...
I. Past is Prologue
Nothing looked wrong in the summer of 2006. Home prices had risen for the better part of a decade. Delinquencies were near historic lows. Credit spreads were tight, the ratings held, and the securitization machine hummed. If you had asked a hundred people on a trading desk whether the American mortgage market was months from seizing, most would have laughed.
Millions of subprime borrowers were, at that moment, paying the low introductory rate on a two-year adjustable rate mortgage - the 2/28 ARM. A low fixed-rate for two years, then the rate reset to a payment 30% to 50% higher. During those first two years the loan performed beautifully: the borrower paid, the servicer collected, and the bond paid its coupon. Nothing looked wrong because the whole complex - housing, mortgages, securitization - was sitting inside the teaser period.
Every ARM reset was known, dated, and contractually inevitable from the moment of origination. Aggregate those reset schedules and you get the most damning exhibit of the era: the reset wall. Roughly a trillion dollars of adjustable-rate mortgages were contractually set to reset across 2007 and 2008 - thirty to forty billion dollars a month at the peak. Credit Suisse published the chart in March 2007. The IMF reprinted it. It circulated on every trading floor in New York and London.
The mortgage reset wall. Every teaser written in the boom became a dated liability
Few understood it. Paulson & Co. laid out the arithmetic that same month in a comment letter to the FDIC: Over 80% of recent subprime originations, it observed, were two- or three-year adjustable-rate products. The average subprime borrower’s mortgage payments already consumed roughly 40% of their gross income at the teaser rate. Almost none of them could service the reset rate out of income.
The crisis, in other words, was written in advance by the instruments themselves. The market looked at the reset wall and kept buying, because every participant believed the exit would arrive before the reset: home prices would keep appreciating and the borrower would refinance into a fresh teaser before the old one expired.
We have spent the last eighteen years describing the financial crisis as a shock - a black swan, a hundred-year flood, a tail event. It was none of those things. Every reset on that chart was contractually inevitable from the moment of origination. The defaults were not primarily caused by an exogenous macro shock, a spike in unemployment, or a recession that arrived first. They were the scheduled mathematical consequence of loans that assumed perpetual appreciation. The mortgages were built to break.
The AI boom has rebuilt this exact structure, and the market is once again underwriting the teaser.
It has a reset wall of its own - a schedule of dated, contractual, non-negotiable payment shocks - hiding inside the trillions of dollars of compute contracts signed by OpenAI and other frontier labs since 2024.
The take-or-pay compute contract - the instrument at the center of the AI build-out - has a structural feature that almost no one prices: its payments do not begin at signing. They begin at delivery. A lab signs a multi-year capacity commitment today, but the payments do not start until the data center is energized, the capacity is accepted, and the contractual ramp schedule commences - an interval set not by finance, but by construction: siting, powering, and filling a gigawatt-scale campus takes 24-to-36 months from signature - mirroring the two-to-three-year teaser of a subprime ARM.
More than $2.3 trillion of compute contracts now sit on the books of the four largest American cloud providers as remaining performance obligations and contracted backlog - signed, celebrated, capitalized into equity prices, and, critically, not yet billing.
During the teaser period, everyone wins. The seller reports backlog growth that compounds at rates no operating business has ever sustained - Oracle’s RPO grew 363% in a single fiscal year. The buyer - a frontier lab burning cash at historic rates - books no expense because the capacity does not yet exist. The market capitalizes the booked number as if it were revenue and ignores the billed number as if it were a technicality. And then, on a schedule fixed at signing, booked compute becomes billed compute. The take-or-pay clock starts. From that day forward, the frontier labs and the hyperscalers incur those costs regardless of utilization. The invoice is a function of the contract, not of demand. That is the reset.
The parallel to 2006 is exact and it explains the single most-cited absurdity of this cycle: How does OpenAI, a company with some $40 billion of run-rate revenue, sign $1.4 trillion of compute commitments? The same way a household with $60,000 of income signed a $600,000 mortgage: because the terms at signing do not require the payment yet, and because everyone at the table - borrower, lender, and the market - believes the growth will arrive before the payment does.
The 2/28 borrower’s defense was always the same: by the time the reset arrives, my house will be worth more and I will refinance. The frontier lab’s defense is structurally identical: by the time the capacity commences, my revenue will have grown into the obligation.
The compute commencement wall can be made visible in exactly the way the reset wall was visible in 2007 - from disclosed contracts and delivery schedules. The only question is whether the market listens this time
The same chart twenty years apart. Left panel - first-reset principal balances per Credit Suisse and Inside Mortgage Finance. Right panel - announced compute commitments and contract disclosures across every frontier lab.
II. The Anatomy of a Teaser
To see why the structure behaves the way it does, I’ll break down a single contract and walk the lifecycle. The terms below are hypothetical; the architecture is the standard one visible across the disclosed OpenAI–Oracle, Anthropic–Google, Meta–CoreWeave, and OpenAI–CoreWeave arrangements.
A frontier lab signs a $12 billion, ten-year capacity commitment with a compute provider. The contract is take-or-pay, meaning the lab commits to payments once the capacity is delivered, and delivery requires a campus that does not yet exist: two years of construction, procurement, and power work stand between signature and completion.
Now look at what each party’s financial statements show during the two-year teaser.
The seller - a hyperscaler or neocloud - books the arrangement into RPO or contracted backlog on day one - the full $12 billion, disclosed, quoted, and celebrated. The market values it as contractual future revenue. Meanwhile the seller’s cash flow statement hemorrhages: the campus is being built, so capex runs far ahead of receipts. Booked backlog rises; reported earnings feel none of the buildout; financing frequently sits off-balance sheet.
The buyer - a frontier lab like OpenAI or Anthropic - announces access to the compute it needs to pursue its scaling roadmap, and its private valuation reprices on the announcement. The commitment is a future obligation, disclosed - if at all - deep in a contractual-obligations footnote or, for the private labs, nowhere public. No expense hits the P&L because no service is being received. A lab that has committed tens of billions across multiple providers carries a cost structure that reflects only its commenced capacity.
The market sees a seller with explosive backlog and a buyer with secured compute capacity, and prices both as growth stories. Nobody is lying. Every number is GAAP-clean. The structure simply guarantees that during the teaser period, the system’s reported economics and its committed economics diverge by the full value of everything signed and not yet commenced.
Every optical incentive points toward signing more.
Then comes commencement, and the two clocks converge violently. The buyer’s cash obligation steps from approximately zero to the full contractual rate, arriving not gradually but as a step function, tranche by tranche as capacity goes live. The seller begins recognizing revenue, which the market applauds, while backlog begins draining. And here is the asymmetry on which the entire thesis turns: the buyer’s obligation steps up on the construction schedule, regardless of the revenue or utilization that shows up.
The parallel is now clear: the 2/28’s teaser is the construction phase, its reset date is commencement, its fully-indexed payment is the full take-or-pay rate, and its refinance-or-sell assumption is the belief that model revenue will have grown into the obligation by the time it bills - or that another round of fundraising will cover it.
The take-or-pay compute contract is the financing innovation of this cycle the way the 2/28 was the financing innovation of the last one, and it emerged for the same reason: an asset too expensive for its natural buyer had to be made buyable. A frontier lab cannot fund a gigawatt campus out of revenue, just as a subprime borrower could not fund a house at the fully-indexed rate. In both cases the solution was an instrument that splits time in two - a cheap phase that gets the deal signed, and an expensive phase scheduled far enough out that the market ignores it.
In residential credit, the interval between origination boom and reset wall was twenty-four months, because that was the teaser’s term. In compute, the interval is the construction timeline - twenty-four to thirty-six months. The 2025–26 signing boom therefore mathematically guarantees a 2027–28 commencement boom, exactly as 2005–06 originations guaranteed 2007–08 resets.
This is what it means to say we are in the teaser period. The booked figure is enormous; the billed figure is a fraction of it and only beginning to turn up. Everything about the present looks like strength. The obligations that will govern 2027 and 2028 are already signed, already dated, and already sitting in RPO. What has not happened yet is the conversion - the moment booked becomes billed and the take-or-pay clock starts running regardless of the revenue and the counterparty’s ability to pay.
III. Take-or-Pay is Debt
The common objection to the 2008 comparison is simple: this is not 2008 because the leverage is not there.
The leverage is there. It’s simply not booked as leverage.
A take-or-pay contract is, in economic substance, a lease. And a lease is a financing. The defining feature of debt is a fixed payment on a schedule, owed regardless of the borrower’s circumstances. That is exactly what a take-or-pay commitment is. The payment does not flex with utilization. It does not wait for the customer’s revenue. It is owed on the commencement date and every period thereafter, for the term.
This is not a new concept. Rating agencies have treated take-or-pay obligations as imputed debt for more than thirty years - pipeline throughput agreements, ship-or-pay contracts in shipping and rail, long-term power purchase agreements, all routinely capitalized into leverage metrics by Moody’s and S&P. The convention simply has not been applied to compute.
Reported gross debt across the AI complex - the frontier labs, the hyperscalers, and the listed neoclouds - comes to roughly $470bn. The present value of disclosed non-cancellable compute and capacity commitments across the same set comes to roughly $1.66 trillion. The economic obligation is $2.1 trillion. For scale, subprime mortgages outstanding in March 2007 totaled roughly $1.3 trillion.
Three mechanisms keep these contracts off the reported balance sheet.
The first is disclosure asymmetry: remaining performance obligations are a seller-side disclosure under the revenue-recognition standard - the vendor tells you what it has been promised - and there is no symmetric requirement for the buyer to tell you what it has promised.
The second is that the largest buyers are private: OpenAI and Anthropic file no periodic reports, and their obligations enter the public record only when a counterparty announces a deal or books the corresponding receivable.
The third is that the contracts are generally structured as service agreements rather than leases - precisely the maneuver that kept operating leases off balance sheets before the standard changed.
The leverage objection, then, depends entirely on where you look. Look at the line marked debt, and there is relatively little of it. Look at the contractual claims on future cash, and there is more than the entire subprime mortgage market carried at its peak.
So the leverage exists. The question that follows is who owes it and whether they can pay it.
As of the second quarter of 2026, the four largest U.S. cloud providers carry roughly $2.3 trillion in contracted revenue backlog. Roughly $1.0 trillion of that total traces to two counterparties - OpenAI and Anthropic.
Both of those counterparties run deeply negative free cash flow and fund themselves through equity raises and vendor-adjacent financing from the same ecosystem whose capacity they are contracting. The single most important credit fact in the global economy right now fits in one sentence: the largest capital cycle in the history of technology is underwritten, to the tune of roughly one trillion dollars, by two private companies that do not make money.
Now contrast this with the cloud build-out of the previous decade. In the 2010s, bookings and billings tracked each other closely. Capacity was added a step ahead of demand that was already visible. Today backlog-to-revenue multiples across the complex now sit at five to six times the pre-AI software norm - with the vast majority of contracts being take-or-pay contracts signed in 2025-2026 and commencing in 2027-2028.
The multi-year commitments dominating these backlogs are underwritten not by observed demand but by a forecast of demand - a belief about how large and how soon the AI economy comes. RPO has quietly been recast from a risk disclosure into the bull case: “look at all that contracted revenue.” But a backlog is not revenue. It is a promise to pay, and it is worth exactly what the party on the other side can actually pay when the promise converts from booked to billed.
So, as the cloud era transitioned to the AI compute era:
1. Consumption on existing capacity became commitment on unbuilt capacity. The revenue-recognition lag went from one to two quarters to two to three years.
2. Variable service agreements became fixed and contractual. Pay-as-you-go, a flexible operating expense of the cloud era, became take-or-pay, a non-cancellable lease structure that the market has not fully priced as debt.
3. A diversified book became a concentrated one. The cloud-era backlog was tens of thousands of enterprise customers. Today more than half comes from two unprofitable companies.
And, 4. The collateral changed. This one will look obvious in hindsight. In the cloud era, backlog was underwritten to the customer’s operating business. A Fortune 500 firm signing a three-year cloud commitment was going to pay it out of an existing profit stream. In the compute era, backlog is underwritten to the customer’s future funding. It is not credit against cash flow. It is credit against the capital markets staying open - which is exactly the expectation of the 2/28.
“But the hyperscalers have 30%+ ROI!”
The ROI the market is capitalizing is not paid by the hyperscalers’ own operations in any self-sustaining sense. It is paid by the counterparties - by OpenAI and Anthropic and the other labs whose take-or-pay commitments are the revenue line under every one of these returns. The hyperscaler’s return on invested capital is only as real as the labs’ ability to make the payments that constitute it.
When commencement arrives, the payment that pays the ROI becomes a payment the counterparty owes regardless of its own demand. If that counterparty’s revenue has grown into the obligation, the return persists and the bulls were right. If it has not, the return does not gently compress - it inverts, because the same take-or-pay contract that was the hyperscaler’s asset is now a claim on a borrower who cannot cover it. It’s credit risk that looks like an operating return.
IV. The Signing Spree
OpenAI carries the largest compute commitments in the system against a revenue base that is a fraction of those commitments, with no parental balance sheet standing behind the obligation. It signed the most, owes the most, and burns the most, and its exit assumption - raise the next round before commencement, the way the subprime borrower’s was refinance before the reset - depends on a revenue curve inflecting on a schedule that has never been demonstrated at this scale.
Between June and December of 2025, OpenAI executed what may be the most concentrated origination spree in the history of corporate credit.
In less than twelve months, the company signed something close to $1.2 trillion in compute commitments. There was a stretch in October 2025, about three weeks, during which the company announced deals whose combined notional value exceeded the market capitalization of ninety-five percent of the companies in the S&P 500.
Signing was cheap and the re-rating was instantaneous. On the days the largest of these deals were announced, Oracle, Nvidia, AMD and Broadcom added a combined $636 billion of market capitalization.
Every dollar of that $1.2 trillion was signed during the steepest part of OpenAI’s revenue curve and underwritten to its continuation. And almost every one of these deals commences in 2027-2028. The signing spree should be read as an obligation event, not a sign of insatiable demand for compute.
V. Building the Reset Wall
Let us build the reset wall and let us build it the way Credit Suisse built the mortgage wall - in two views:
-
The first is a cash question: how much does the company owe, per year, as these contracts commence? This is the equivalent of Paulson & Co’s arithmetic - which was used to compare the mortgage payments to the borrower’s income.
-
The second is a concentration question: what is the total compute contract amount that resets from teaser to full pay in a single year? This is the equivalent of Credit Suisse’s 2007 reset wall - which showed the principal amounts of adjustable-rate mortgages resetting in a given year.
OpenAI’s committed annual compute cost, built bottom-up from the announced vendor contracts and reconciled to management’s own disclosed plan. The step into 2027 is the reset.
The 2007 reset wall was drawn in notional rather than annual payments - in other words, the unpaid principal balance transitioning from teaser to fully indexed. The compute equivalent is contract notional payable from the commencement date forward.
Credit Suisse could build the 2007 reset wall because securitization documents disclosed every loan’s reset date. Compute contracts are private, so the wall must be modeled - but the inputs are unusually good, because the counterparties keep announcing them publicly.
And, much like 2008, synchronized originations produces synchronized resets. Mortgage origination peaked across 2005 and 2006; the teaser was twenty-four months; the wall peaked across 2007 and 2008. Compute signing peaked across 2025 and 2026; the construction interval is twenty-four to thirty-six months; the wall peaks across 2027 and 2028. Same arithmetic, different collateral.
The mortgage-balance analogue: contract notional still payable from the commencement date forward. $712bn of it recasts to full pay across 2027–2028 for the two frontier labs alone.
Now, replicate the way Paulson & Co. measured the 2/28 borrower: compare the annual cash payment to income and determine the counterparty’s ability to meet these resets. OpenAI has no income, so in this case the comparison is against revenue. Apply four revenue paths, each anchored to the latest reported figures and to what the company itself has told investors.
Run every scenario management or the forecasters will offer - re-acceleration, the management plan, the forecaster median, a slow burn - and set each against the committed compute cost. The bottom panel is the coverage ratio: compute commitments as a share of revenue, before wages, research, sales, or tax.
Even under management’s own plan, compute alone consumes more than 200% of revenue at the 2027 peak. There is no scenario on the chart in which the frontier lab covers its compute bill out of revenue in the year the wall lands. The best case is that it grows back under the line by the end of the decade, and the best case requires the refinancing channel to stay open the entire way.
So, OpenAI’s plan for the reset is to refinance at the reset. Raise the next mega-round, at a higher valuation, to cover the obligations as they commence - exactly as the subprime borrower planned to refinance into the next loan when the teaser expired. This works while two things hold: the capital markets stay open, and the narrative stays intact.
And look at what this implies about OpenAI’s valuation as it moves toward an IPO:
OpenAI’s equity - valued north of $850 billion - is functionally the junior tranche of a capital structure whose senior claims, the take-or-pay compute obligations, exceed any revenue path management itself has articulated.
On those numbers, the equity is effectively underwater, and the market has not priced it that way because it still treats those obligations as service agreements rather than what they are economically: debt.
Even if OpenAI can meet those obligations, OpenAI’s unaudited financial statements - as of March 31, 2026 - disclose $665 billion in non-cancellable compute commitments (management’s more recent plan runs to $750 billion). These commitments are take-or-pay in structure - which, as established above, is debt.
Carry the present value of those obligations as senior debt - roughly $450–500 billion - and a company the market prices as debt-free carries a senior claim worth more than half its entire equity value.
The market is pricing the residual equity as if it were the whole stack.
VI. What the Wall Demands
The labs' answer is the 2/28 borrower's answer: revenue is compounding at triple digits, and by commencement it will cover the payment. It might. The credit point is narrower: the revenue coverage claim is a projection, while the obligation is a certainty.
The claim is not that commencement causes a lab to fail. It is that commencement is the date on which a pre-existing mismatch - fixed obligation against assumed revenue - becomes cash-due, and that, as in 2008, the mismatch is visible in the fundamentals well before the date makes it unavoidable. You do not need demand to fall. You need it only to decelerate below the rate the booked compute was underwritten to.
The compute contracts commencing in 2025 and early 2026 cleared, or very nearly cleared, the required growth rate. This is the crucial point, and it is the reason there is no alarm anywhere in the system: the early vintages worked.
They worked the way the 2005 and 2006 subprime resets worked. The collateral appreciated fast enough. The refinancing happened. Everyone who signed was vindicated, and vindication is the input to the next round of underwriting. Success in the early vintages is the mechanism that manufactures the late ones.
For committed compute merely to equal revenue in 2027 - not to be comfortably covered, simply to reach parity, before a single dollar is spent on wages, research, sales, or tax - revenue would have to compound at 217% annually off the 2025 base. The dashed line at 100% represents revenue doubling every single year and sustaining it, which no company at this scale of revenue has ever done for a multi-year stretch. The obligation is accelerating at more than double the rate of the best case for the cash flow meant to cover it.
The obligation curve is contractually fixed and steep - it ramps according to a defined construction timeline. The revenue curve is a growth rate. If the growth rate rolls over - the two curves cross. That is the reckoning: not a demand collapse, but a demand deceleration meeting a cost schedule that was set in a more optimistic year.
Deceleration alone is survivable if your cost base is variable. If demand growth slows from a 120% to 40%, a company with variable costs simply spends less, earns less, and adjusts. But a take-or-pay obligation is not variable. It is a fixed dollar amount that arrives on a fixed date regardless of what the demand curve did in the interim.
None of this means the company fails. It means the company must raise. Take the base case: roughly $375 billion of cumulative uncovered compute cost across 2026 to 2030, before research and development, before compensation, before every other operating cost of running a frontier laboratory. Round the all-in external funding requirement to the four-to-five-hundred-billion-dollar range across five years, and the exit assumption becomes explicit and testable:
The thesis for OpenAI requires capital markets to fund roughly half a trillion dollars of cumulative operating deficit at a single pre-profit counterparty, at non-punitive terms, through a window in which that counterparty’s compute costs are contractually rising faster than any plausible revenue path.
That may happen. But it should be named for what it is: a refinancing assumption rather than an operating plan, and one that depends on the collateral - the valuation - exactly in the period in which the true cash cost of the build becomes visible for the first time.
Construction timeline slippage can move the obligation - the 2027 peak flattens slightly, the 2028 peak rises, and total obligation is unchanged. The revenue that was supposed to grow into the 2027 obligation now has to grow into a larger 2028 one. This is exactly what happened when servicers pushed resets in 2007. Deferral was a repricing of when, not a cure.
OpenAI has been built as if the AI boom were a venture-backed, technology cycle; when in fact, it has the mechanics of a credit-driven real-estate cycle (as I outlined in The Second Derivative). Every decision executives have made seems to be based on maximizing a single outcome: the next round. While compute commitments are in the teaser period, they are assets - secured compute capacity signaled strength and raised the next round. OpenAI is facing a day of reckoning when those commitments are delivered and, on a schedule indifferent to their revenue or next round of funding, booked compute becomes billed compute.
And time is running out - Bridgewater’s analysis shows OpenAI is burning through their latest fundraise at an extreme pace.
VII. Anthropic and the Whole Stack
The comparison to Anthropic is useful as a controlled experiment. On the same measure, Anthropic’s compute commitments peak at close to 60% of revenue in 2027 then falls - fully covered by revenue with room left to pay operating costs. Undoubtedly stressed in the reset window, but structurally solvent and improving from the peak rather than grinding against it. Two labs, the same instrument, the same commencement window, and coverage ratios that differ by more than a factor of three at the peak.
While in a substantially better position, Anthropic is similarly the equity tranche of a capital structure heavily indebted by take-or-pay compute commitments, which the market has also failed to appropriately recognize as debt.
For sake of clarity, the revenue figures used in this analysis are annual revenue figures not a run rate.
The full system is larger, because the labs are only the top layer. Consolidated across frontier labs, hyperscalers and neoclouds, contract notional recasting peaks at $732 billion in 2027 and $820 billion in 2028. $2.4 trillion recasts from teaser to full pay across 2026 to 2029, with the two-year peak in exactly the window the frontier-lab layer identified.
System-Wide Contract Notional Recasting, Consolidated.
Now, place the full stack side by side with the mortgage reset wall.
The Reset Wall, Then and Now. The whole stack on the right, consolidated and net of eliminations.
It is worth being clear about what these charts imply:
It is not a default forecast. The reset wall did not “predict” defaults in 2008 either. It only revealed the date on which the question would be asked.
It is a statement about synchronization and about arithmetic. It says: on a schedule fixed by contracts already signed, a very large volume of fixed obligations transitions from deferred to due, in a narrow window, for a set of counterparties whose ability to pay the reset depends on a revenue number that does not yet exist - it’s a projection - and whose cash flow today is reliant upon external funding.
That is exactly what the Credit Suisse chart said in 2007. It was right, and it was ignored, and it was ignored for a reason that will be entirely familiar: at the moment it was published, every loan on it was still performing.
When skeptics raised the reset schedule in 2007, the rebuttal was performance data: delinquencies are at record lows. So they were - the vintages were two years old, home prices had risen by double-digits, and the payment being performed was the teaser payment. Today’s rebuttal has the same rationale: AI revenue is compounding at triple digits; utilization is effectively full; every GPU is oversubscribed. All true. All measured during the ramp, while capacity trails demand by construction lag and the billed payments run at a fraction of the booked compute.
VIII. The Second Teaser: Hyperscalers
The frontier labs have a contractual teaser: an obligation that is signed and not billed. The hyperscalers have an accounting one: an asset that is paid for and not expensed.
Under U.S. GAAP, capital under construction sits in “construction in progress.” Depreciation does not begin at expenditure. It begins at placement in service - when the asset is available for its intended use - regardless of whether it is being used. Construction-period interest is capitalized into the asset’s cost and expensed only after placement, through depreciation, over the asset’s life.
Then the tranche goes live, and GAAP flips the switch. Depreciation commences on the full capitalized cost - including the capitalized interest now embedded in the basis. The asset moves, in one accounting instant, from an inert balance-sheet entry to a recurring income-statement charge. In-service to the owner is what commencement is to OpenAI: a reset whose date was fixed by the construction schedule, utterly indifferent to whether demand showed up.
As tranches go live through 2027–28, depreciation inflects upward mechanically and the hyperscalers’ operating margins begin absorbing the fully indexed rate. If utilization and pricing hold, revenues rise in tandem and absorb the scheduled depreciation. If they do not, the industry will discover that depreciation is take-or-pay with the income statement as the counterparty: a fixed charge, contractually scheduled, indifferent to demand, and impossible to renegotiate.
Consider what a live datacenter owes each month whether it runs at 90% utilization or 30%. It owes depreciation, power, interest, staff, cooling, and maintenance. In a representative cost stack for a leveraged cluster, roughly 80% of the monthly cost is fixed the day the meter turns on.
This is operating leverage - a wonderful thing on the way up and a merciless one on the way down. When utilization holds, margins are spectacular, which is exactly the story the teaser period tells. But the same fixed base, spread across revenue that arrives below the underwritten level, produces negative operating leverage. There is a break-even utilization built into every one of these assets - the point below which fixed costs are not covered - and below it, the asset bleeds.
If OpenAI cannot pay, the hyperscalers do not just miss revenue - they absorb a fixed-cost shortfall that their own operating leverage magnifies. A 30% utilization drop does not mean 30% less profit. It can mean the entire facility turns unprofitable.
IX. The Options
When billing commences, unused capacity transforms overnight from strategic optionality to cash burn. A CFO staring at that line item finds ways to mitigate it.
You cannot cancel: take-or-pay is take-or-pay, senior in practice to everything. You can try to grow into the capacity, but demand is largely outside your control. Three mitigants remain: raise capital, renegotiate, or sublease.
Renegotiation is the most likely path. OpenAI’s negotiating leverage is proportional to its systemic importance - perhaps why it proposed handing a 5% equity stake to the federal government. It is too interconnected to fail; every balance sheet in the chain needs the fiction maintained. The renegotiations, when they come, will not be shown as distress. They will look like partnership: volume deferrals framed as capacity rephasing, rate cuts as efficiency-linked pricing.
But the moment one anchor lease is amended, every RPO dollar in the complex carries a demonstrated amendment probability. “Contracted” ceases to be a synonym for “certain” anywhere in the system. The $2.3 trillion only needs quiet contract negotiations to be re-rated as an asset class.
Sublease is the alternative. A tenant subleasing capacity it cannot use will take nearly any rate above zero, because every dollar recovered directly reduces cash burn. The bull case points to premium rates on today’s short-term subleases. But look at the terms: xAI’s arrangements carry ninety-day termination rights; Google frames its leases as bridge agreements; Anthropic takes spot capacity while aggressively contracting bespoke capacity elsewhere. This is bridge demand by construction. It exists only until the 2027–2028 multi-gigawatt deliveries land, at which point it hands the space back - flooding the market with shadow vacancy just as the rest of the $2.3 trillion commitments convert from booked to billed.
Compute does not need to default to break the market. It only takes a wave of quiet contract renegotiations and shadow-vacancy subleases to re-rate the asset class from a scarce strategic commodity to an oversupplied utility.
X. The Index is the Trade
When that re-rating happens, the equity of the entire complex absorbs the loss - and that equity is concentrated in the handful of names that dominate the market-capitalization-weighted indices most of the developed world owns through its retirement accounts. The ultimate holder of the risk is a household that has never heard of a take-or-pay contract.
AI-exposed names now account for roughly 45% of S&P 500 market capitalization. The ten largest companies in the index - themselves overwhelmingly AI names - sit near 40%, against about 27% for the top ten at the dot-com peak. It is the most concentrated the index has been in its modern history
Semiconductors carry roughly 19% of the index and supply roughly 45% of its total earnings growth, the largest share of any sector. That contribution is a function of the order book, and the order book is largely a function of new originations. When the reset lands and the labs spend every marginal dollar servicing commitments already commenced rather than signing new contracts, the next wave of chip orders thins. The vendors are a pure second-derivative play - they book the boom first and feel the deceleration first.
Hyperscalers are roughly another 20% of the index - and the two frontier labs are nearly half of their $2.3 trillion backlog. When the commencement wall hits, depreciation and fixed costs kick in on a schedule indifferent to whether those counterparties can pay, and the hyperscaler’s P&L becomes the backstop for any capacity the labs overbought and cannot cover. Hyperscaler ROI is fundamentally frontier lab credit risk - and the equity market hasn’t even begun to price that in.
Neoclouds are the most levered expression of the wall. CoreWeave and its peers financed gigawatt campuses on debt raised against the take-or-pay contracts themselves - backlogs many multiples of revenue, thin equity beneath, and the bulk of it commencing in 2027–28. Their model rests entirely on booked converting cleanly to billed; the capital structure has no room for a deferred or renegotiated anchor lease. They carry no index weight, but cracks in the take-or-pay complex expose them.
A passive retirement account holding an S&P 500 index fund owns a levered, concentrated bet on the conversion of contracted compute backlog into billed revenue - and on two cash-burning frontier labs’ ability to pay for it Nobody chose that allocation, and almost no one holding knows it.
XI. Living Inside the Teaser Period
The hardest thing to convey about 2006 to anyone who did not trade through it is how good the data was. Record origination, record homeownership, delinquencies scraping decade lows, homebuilder earnings at all-time highs, and every incoming statistic confirming the strength of the American consumer.
What almost no one priced was that every one of them was a teaser-phase measurement: an observation of a system whose payment test had not yet begun, generated by an instrument that mechanically guaranteed the data would look exactly this way until the schedule said otherwise.
An economy of teasers cannot produce bad credit data until the calendar turns, which means the strength of the present data carried no information about the question that mattered. The signals everyone watched were structurally incapable of carrying the signals everyone needed.
Now read the compute cycle’s tape with that in mind. Record RPO backlog, celebrated the way 2005 celebrated origination volume. Capacity sold out, demand insatiable: a construction-phase statement, necessarily true while contracted delivery lags contracted demand. Vendor revenue beating estimates, the way homebuilder earnings were the sound of the mortgage machine consuming its own vendor inputs.
This is the epistemic signature of a teaser period, and it explains the otherwise baffling social dynamics of standing inside one. The bear who cites the future reset wall is answered with the current data. A teaser period does not merely hide the reset wall. It manufactures the exact evidence used to dismiss it.
XII. This Time is Different
Reinhart and Rogoff titled their history of eight centuries of financial folly with the words that recur before every crisis: “this time is different”. And the maddening truth is that the specifics genuinely are different every time.
This is not precisely 2008. GPUs are not houses; take-or-pay contracts are not mortgage-backed securities; OpenAI is not a subprime borrower in Stockton, and artificial intelligence may well be the most consequential technology of the century, which is more than anyone could ever say for a McMansion in the Inland Empire.
All of that is true, and none of it is the point. What repeats is never the surface. What repeats is the structure:
a scarcity thesis that justifies enormous fixed obligations; a teaser period during which those obligations feel costless; a set of commencement dates, fixed at signing, on which the teaser expires and the fully-indexed bill begins; and a bet that the income will have grown to meet the bill by the time it arrives.
The reason many AI skeptics will be right in substance and wrong in the mechanism is that they are often making a valuation argument, and valuation arguments have no clock. What this piece has tried to show is that buried inside the compute contracts is something a valuation argument never has: a reset.
The bull case wins if - and it is a real if - demand scales into the committed supply before the reset wall lands, and the counterparties stay funded through any air pocket in between. The bear case in this piece is not that artificial intelligence will fail, or that the demand is fake, or that the technology disappoints. It is narrower: that the financing structure can break before the demand arrives, because the obligations are fixed and front-loaded in commencement while the revenue is variable and back-loaded in adoption - and a fixed obligation meeting a lagging revenue stream is a solvency problem regardless of how transformative the underlying technology turns out to be.
The industry will spend the next eighteen months debating whether artificial intelligence is a bubble, which is the wrong question, asked at the wrong layer. The technology is real; so were the houses. The question is narrower: what happens when instruments underwritten at the teaser meet their reset schedule, and who is holding the paper when the obligations cannot be met as written. The reset wall is published above and the AI boom sits in a period of fiction.
The Teaser Period.
Read more and subscribe to 'Groundbreaker' here...
Tyler Durden
Sun, 08/23/2026 - 18:15
Source: 
Xu Jiayin, also known as Hui Ka Yan, founder of property developer Evergrande, appears for sentencing at the Shenzhen Intermediate People's Court in Shenzhen, China, on Aug. 20, 2026. Shenzhen Intermediate People's Court /Xinhua via AP
The halted under-construction Evergrande Cultural Tourism City in Taicang, Suzhou city, in China's eastern Jiangsu Province, on Sept. 17, 2021. Vivian Lin/AFP via Getty Images
Construction workers rent shared bicycles as they leave a building site for a new office tower in the Central Business District of Beijing on April 3, 2025. Kevin Frayer/Getty Images
The Rogers camp is wary of El-Sayed using AIPAC support of Rogers as a powerful cudgel in the Michigan Senate race
Recent comments