AI Off-Balance-Sheet Obligations  Hit a Subprime-Like Reset Wall in 2027 and 2028

AI off-balance-sheet obligations, particularly take-or-pay contracts for frontier labs, are set to hit a reset wall in 2027 and 2028, according to Groundbreaker. The article warns that massive AI spending, financed heavily by private debt with leverage on leverage, could lead to a financial crisis similar to the subprime meltdown, as major tech companies face inadequate returns and rely on debt to fund unsustainable capex.

Yours truly has to confess to not having looked enough into what is very clearly unsustainable AI spending and financial because the main players had succeeded in making it much too hard to get to the bottom of things. When someone like the exceedingly diligent and detail oriented Ed Zitron has to write many multi-thousand word posts to demonstrate that the main players ex Nvidia are showing paltry revenues on their huge and growing capex, let alone anything that could be charitably called profit, and no prospect of the needed monster increases in both overall consumption and payments per user and/or task1 Ponzi-like circular financing is a huge red flag. Yet the overall trajectory seems clear: massive and growing and allegedly absolutely necessary increases in outlay, with no prospect of even adequate returns. Even the once very cash-rich Big Tech incumbents who are all in on this break-the-economy wager lack sufficient internal resources. They also recognize that even equity issuance will not provide enough dough. So they have been and expect to continue to rely on debt. As established readers likely know, it is excessive private debt that causes financial crises. If it is merely one layer of leverage think the Japan real estate and stock market bubbles , the outcome even with a very very big overhang tends to be a deep downturn and what Richard Koos has called a balance sheet recession. If the borrowings are also levered, as in the runup to the 1929 Great Crash2 or the 2007-2008 crisis, the result can be a catastrophic fast meltdown that imperils the payment system, and not just wealth. In the subprime mortgage crisis substantial part of the picture were very visible. The problematic subprime loans were in large measure sold into securitized pools, which meant there was considerable information about them. There was also a good deal of data about the residential housing markets, since title transfers and the price of the sale are all public record, albeit at the local level, making good aggregation and analysis costly. With AI financing, the degree of opacity is much greater. Even though the alpha players are public companies, their AI activities are only one part of their multiple lines of business. As we indicated at the top, the Herculean effort Ed Zitron makes to come to some conclusions about their AI economics shows how inadequate their AI disclosure is. It gets even worse when you turn to the debt part of the equation. AI financing comes either from public bond issuance by big tech players for general corporate purposes, and not dedicated to AI. But it relies even more heavily on private debt, where so-called private credit funds are substantial providers. The Financial Times and other major publications who have attempted to get to the bottom of private credit have thrown up their hands, warning that there is leverage on leverage there, but no one has a clue as to how much. For instance, private credit funds buy the risky debt of private equity investee companies. The investors in those funds are often hedge funds or wealthy individuals who have borrowed to increase the amount they can put to work. On top of that, the private credit funds themselves may borrow, through so-called subscription lines of credit. However, the aptly-naked Groundbreaker website, in The Teaser Period: Why the AI Boom Is Hitting a Reset Wall https://www.groundbrkr.com/p/the-teaser-period-why-the-ai-boom , describes how the AI practice of using take-or-pay contracts for “frontier labs” is set to blow up across the industry in 2027 and 2028. Note that Groundbreaker is being faithful to industry nomenclature, so forgive me if my translation is a bit approximate. Critically, it is these “frontier labs” that contract for, lease and build data centers. The very little I have seen about this process has caused me concern. For instance, I read of Meta contracting for a data center for what sounded like 20 years but was 4 back-to-back five year leases. Recall that the once AAA rated General Electric has taken a big fall in investor grace. How strong will the now-mighty-seeming Meta be, particularly if AI bets come a cropper? Groundbreaker has identified a key element of risky AI leverage, which is leases that are not accounted for as debt at the AI sponsor. Those of you who are corporate finance literate will know this is a cardinal sin. When I was a wee young thing at Goldman, where we went through hard copies of financial statements to extract key data which we entered into green accountants’ ledger paper, one of the aims was to determine total debt to assets. One of the footnote items we would root out was operating lease payments. For instance, big companies often use leased mainframes. We would capitalize them and treat them as a long-term liability. The very direct analogy to subprime is that a large swathe of subprime loans has low “teaser” rates, typically for two years, and then reset. The open secret then was that many borrowers could not afford the higher payments and were destined to default if they kicked in. But the remarkable belief, which worked until it didn’t, was that these borrowers would be bailed out by rising real estate prices. They would re-fi into another teaser loan before the reset, with banks and intermediaries taking fees on every turn. Rising interest rates put an end to this party in 2007. Analysts were trying to argue that enough borrower would survive the coming resets that there would not be big loan losses. Yours truly begged to differ then https://www.nakedcapitalism.com/2007/03/does-optimistic-cagan-analysis-of.html . The AI version is that “frontier labs” signed contracts with hard financial obligations to pay for trillions in compute commitments in 2027 and 2028. Groundbreaker describes the AI analogy to subprime teaser rates as to how the industry has cheerily assumed they can somehow grow their way out of them….and why that looks pretty much impossible. The Groundbreaker it is exceedingly carefully argued and substantiated. Shades of Ed Zitron AI economics seem to require that. So I strongly urge you to read the piece in full when properly caffeinated. https://www.groundbrkr.com/p/the-teaser-period-why-the-ai-boom Any shortcomings are likely to be the product of attempting to recap a complex case. Here is the guts of Groundbreaker’s explanation of why the upcoming take-or-pay compute obligations are not properly accounted for and will create a huge shock when they come due: 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 atdelivery.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 costsregardless 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. Groundbreaker goes though a representative agreement to demonstrate the nature of the obligation: 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 itscommencedcapacity. The marketsees 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 reportedeconomics and itscommittedeconomics 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. Later the analysis notes: 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 This is not as bad as it sounds. While $1.3 trillion was indeed the most commonly-cited total for subprime, if you included so-called Alt As, as in slightly less drecky mortgages, the amount rose to $2.1 trillion. US GDP in 2007 was roughly $14.5 trillion. It is approximately $32.5 trillion now. So the AI contribution to private debt to GDP is lower than AI commitments are now. However, AI spending is a far bigger contributor to GDP growth now than residential housing was in 2007. It’s been estimated at nearly half. So a sudden fall in AI capex would drag down any growth, even before getting to the recessionary-or-worse impact of an AI equity bubble implosion. And we don’t know how much damage-multiplying leverage on leverage is at work, but there is some and perhaps quite a lot. Mind you, there is oodles more more detail and analysis at Groundbreaker; again please read the post. At the risk of over-hoisting, a final key bit: 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. Bloomberg’s lead story make clear how realistic this scheme is: So even if the recessionary effects of the Strait of Hormuz strangulation on energy prices does not deal Mr. Market a body blow shortly, the AI “resets” are destined to finish the job. Stay tuned. 1 For the moment I am deliberately avoiding the use of industry nomenclature like “compute” which IMHO is a part of the obfuscation game. 2 The Roaring Twenties stock market not only allowed for high level of margin lending, but also trusts of trust and trust of trusts of trusts, with leverage at each level of the resecuritization. Derivatives expert Frank Partnoy described in his book The Match King how some of the structures strongly resembled CDOs. In the days when I was chronicling the crisis and eventually got insider help to explain how CDOs worked an more important, understand how they worked,The CDO was in turn to solve the problem that the riskiest rated tranche of subprime bonds, the BBB/BBB- tranche, did not pay enough in yield for anyone not stupid to want to buy it. So sausage-like, they were rolled into a new securitization, the CDO, with some non-subprime assets included to make it look better, and tranched. The AAA layer was popular because it paid higher than a normal AAA yield. But like with the original subprime bond, the lower rated layers were unloved and typically rolled into a new CDO, a so-called CDO squared, Not only were there occasional CDO cubeds, but I encountered a CDO to the fourth called Octonian, which is a mathematician’s joke for “end of the line”. From one description https://curtjaimungal.substack.com/p/what-are-the-octonions-exactly : However, following that road leads to a terminus, ending in eight dimensions. Mathematicians would say the list of such number systems closes: you can write down every member and prove that nothing else will ever join it. One comment - Hands Off the Loot So, for the data centers that are under construction, where is that money coming from? Materials have to be purchased, workers paid, etc. Is that all coming from the fundraising / bond sales by the hyperscalers? Reply comment-4460991 ↓