What if the biggest risk in artificial intelligence right now has almost nothing to do with whether the technology works? I’ve been watching the numbers and the deal announcements, and something keeps nagging at me. Levels of spending, revenue, and installed capacity look impressive. Growth rates remain strong. Yet the entire financing machine appears built on a quieter assumption: that growth itself must keep accelerating. When that second derivative turns, the structure does not need a collapse in demand to break. It only needs the rate of growth to stop rising. That is the pattern that defined 2008 far more than falling house prices ever did.
The Quiet Parallel Between Compute and Credit
Most conversations about the current AI expansion still treat it as a classic technology cycle. People debate whether valuations are stretched, whether adoption will match the hype, or whether a handful of leaders will dominate while others fade. Those questions matter, but they miss the deeper architecture that has taken shape over the past two years.
The build-out is increasingly financed like real estate. Hard assets (data centers full of GPUs) sit at the center. Long-duration debt and take-or-pay contracts support them. Special-purpose vehicles and residual-value guarantees appear in the background. Capital is being raised against the expected cash flows of the chips themselves. In short, compute has become collateral.
When a major chip supplier recently signed memorandums of understanding with large private capital firms to create independent compute financing platforms, the message was clear. The goal is to mobilize hundreds of billions of third-party capital. The GPUs are no longer just tools for training models. They are investable assets that can be housed in separate entities, pledged, and financed against projected returns. Residual support from the manufacturer itself, capped in some cases, functions much like an older residual guarantee in a leveraged lease.
This is not a minor customer accommodation. It is the formal recognition, at very large scale, of a financing stack that stays solvent only while the underlying growth rate continues to accelerate. The bond market and private credit are no longer simply funding construction. They are being invited to treat the hardware as the primary security.
Why the Second Derivative Matters More Than Levels
Markets are very good at watching two numbers. The first is the absolute level of activity: backlogs, gigawatts under construction, revenue figures, token usage. The second is the first derivative, the growth rate itself. Almost no one systematically tracks the second derivative, the acceleration or deceleration of that growth.
Yet regime change often lives precisely there. When financing embeds a growth assumption (a reset that assumes refinancing will be available, a covenant that assumes rising cash flow, a commitment sized to continued expansion), the assumption is satisfied not by a high level but by sustained growth. Growth that is still positive but slowing can keep the headlines looking healthy while quietly violating the embedded premise.
There is usually a window of borrowed time between the moment acceleration rolls over and the moment the growth rate itself crosses zero. During that window everything still looks fine. Revenue hits records. Press releases stay upbeat. The machine is already under stress; it simply has not been forced to acknowledge it yet.
That window becomes dangerous in proportion to the convexity of the instruments sitting on top of the underlying quantity. Equity multiples can deflate gradually. Leveraged credit structures are negatively convex. They collect a fixed coupon on the way up and absorb losses that can expand rapidly on the way down. Covenants often act as step functions rather than smooth curves. The difference between a 2000-style equity de-rating and a 2008-style credit seizure is exactly this distinction.
Real Estate Mechanics Disguised as Technology
Walk through the current AI infrastructure story and the real-estate features become hard to ignore. A data center is entitled land plus a structure. It is financed with debt against the asset. Capacity is leased to tenants under take-or-pay terms. Construction lags mean new supply often arrives after the demand impulse that justified it has already changed character. This is not a pure software business that happens to own servers. It is a real-estate-like business that happens to compute.
Real estate cycles rarely break because absolute demand disappears. They break when the rate of demand growth decelerates against a large wave of fixed supply that was committed during the boom. The second derivative again. And the credit instruments that sit on top of those assets do not de-rate gently. They refinance or they seize.
Within the current complex the financing pathways differ. Pure-play neocloud operators often borrow non-recourse against a specific tenant’s take-or-pay commitment. If that tenant cannot perform, the structure can move toward seizure. Hyperscalers fund more through corporate bonds and operating cash flow; for them a tenant problem shows up as impairment and margin pressure rather than immediate seizure. Some large players sit in between, corporate-funded yet highly concentrated on a few counterparties. The wounds look different, but the same deceleration can inflict both.
The Loan Book Hiding in Plain Sight
Look past the scarcity narrative and the agreements take on a different shape. A frontier lab signs a multi-year commitment to pay tens of billions for compute it has not yet used. The counterparty books that promise as backlog and borrows against it to pour concrete and rack hardware. Reduced to its skeleton, the arrangement is a loan. Capital is advanced in the form of a building full of chips. The contracted payments serve as debt service. The structure works only as long as the borrower can keep funding those payments.
For a company that is not yet profitable, that funding comes from continuous capital raising. Each new round must clear at a valuation high enough to service prior commitments and enable the next set of contracts. The markup itself becomes a form of cash flow. A borrower whose liquidity is tied to the rate of change of its own valuation is solvent not because the absolute mark is high, but because the mark is still rising fast enough relative to an accelerating burn schedule.
When the step-up between rounds compresses, the arithmetic tightens even if the absolute valuation still prints a new high. That is not a paradox. It is the natural result of a structure whose serviceability depends on refinancing into growth.
Counterparty Concentration and the Single Variable
Across the large platforms, a meaningful share of the contracted backlog traces back to a small number of frontier labs. Some of those labs have stronger enterprise mixes and structural supports that effectively lift their credit quality. Others remain more exposed, relying on continued private capital and eventual public market access as the terminal refinance.
When one large counterparty sits at the center of so many contracts, either directly or one step removed, correlation rises. A single variable (whether that counterparty can clear its next required mark) begins to influence a wide set of balance sheets. Chip stocks have already shown sensitivity to signals about IPO timing. The market is starting to notice the wiring even if it has not yet fully priced the risk.
Securitization has begun to appear as well. Facilities issued through bankruptcy-remote vehicles, backed by GPU capacity serving non-investment-grade customers, have entered the broader credit complex. Once paper leaves the originator’s balance sheet and begins to trade, the distribution step is complete. The performance of that paper still depends on the same underlying ability of the end tenants to keep raising capital.
Capex, Cash Flow, and the Reflexive Flip
Hyperscaler capital expenditure has climbed aggressively. As a share of operating cash flow the ratio has moved higher year after year and is projected to approach or cross full coverage of cash flow in the near term. Past that point, marginal capacity is funded from the balance sheet through debt or equity rather than internal generation. Debt issuance has already begun to rise, and leverage metrics across the group have increased noticeably.
There is a recursive element that the headline numbers can obscure. A substantial portion of the demand that appears in the system is itself the result of hyperscaler spending recycled through cloud credits, compute commitments, and equity-funded consumption. Strip out that loop and the organic, independent demand is a smaller fraction of the total. The second derivative of hyperscaler capex therefore becomes one of the most important growth rates in the entire complex. And that acceleration has already shown signs of rolling over even while absolute levels remain elevated.
The spending race is a conditional Nash equilibrium. It holds while the market rewards the next dollar of expenditure as a call option on future growth. In that regime every player has an incentive to keep spending because the alternative is to be the one that blinked. The equilibrium is anchored to belief, not to a fixed physical constraint. Beliefs reprice.
The flip occurs the first time a major player signals a meaningful slowdown in spending and the market rewards the stock rather than punishing it. Once discipline is celebrated instead of expansion, the payoffs invert. Spending becomes the move that gets punished; restraint becomes the move that gets re-rated. Because it is a coordination game, the shift can cascade quickly. The day the market cheers a cut is the day the arms race logic changes.
That repricing does not leave the existing contracts untouched. In the boom phase a signed take-or-pay commitment is an asset for nearly everyone who touches it. In the new regime the same contract can become a liability for the lab that must fund it, the provider that holds the receivable, and the lender whose collateral depends on the cash flows. Nothing physical has to break. The market only has to change its collective reading of what the next dollar of spend means.
Who Moves First and What Follows
The first significant cut is unlikely to come from the weakest balance sheet. It is more likely to come from the player with the best information, the credibility to frame restraint as strength, and the balance-sheet flexibility to absorb the short-term optics. Others will then face a different set of incentives. Some will benefit if rivals retrench. Others may find their hands forced by free-cash-flow pressure or concentration risk.
If the capital window for the most exposed borrowers tightens, the sequence is reasonably clear even if the timing is not. Commitments get revisited. Neocloud operators whose entire model is the spread between borrowed money and resold capacity feel the pressure first. Corporate-funded but concentrated players experience deeper impairments. Credit across the complex begins to reprice remaining performance obligations from pure demand visibility into counterparty risk. GPU-backed paper becomes harder to roll. Equity multiples compress as capital expenditure is re-read from growth option to cost center.
The strongest balance sheets with the least direct exposure are positioned to acquire stranded capacity at lower prices and to provide the backstops that remaining leases require. That is the classic endgame of a credit-driven real-estate-style cycle.
The Case That Still Needs Answering
The bull argument is not trivial. Demand for inference is real and still early. Backlogs are large. The cost of under-building a platform that turns out to be generational would be high. Supply remains constrained for years in some categories. Even skeptics can sketch paths to very large annual capital expenditure.
Yet every one of those points addresses the level or the first derivative. None of them addresses the second. A structure this levered and this dependent on continued acceleration does not require demand to vanish. It only requires the rate of growth in the key drivers to flatten. Housing demand was still positive in 2006 when the financing structures began to seize. The break occurred on deceleration, not on an absolute collapse.
Three common defenses deserve direct attention. First, the fortress balance sheet argument: large cash flows can absorb tenant impairments. The accounting impairment is one issue; the operating leverage is another. Capacity carries a large fixed-cost base of depreciation, power, and interest that does not shrink when utilization falls. Margins can compress even if the absolute loss is manageable on a consolidated basis.
Second, the cross-subsidization defense: legacy cash flows from other businesses can carry an AI division through a rough period. Investors ultimately buy growth and returns on capital. If a large new segment consumes capital without delivering commensurate profitability, the consolidated return profile changes and the multiple can re-rate lower. Legacy cash cows are also not infinite or immune to their own pressures.
Third, the re-leasing argument: if one tenant defaults, the capacity can simply be leased to someone else. That works cleanly only if the broader environment remains tight. A default that coincides with a wider deceleration leaves new supply arriving into a softer market. Replacement tenants negotiate harder on price and duration. What was a high-yield, long-duration asset can become a lower-yielding, shorter-duration one. The problem is transformed rather than eliminated.
The disagreement is therefore not primarily about the long-term usefulness of the technology. It is about which derivative the current financing structures are written against. One side watches the absolute size of the opportunity. The other watches the rate of change of the rate of change. Arithmetic eventually settles the argument.
Borrowed Time and the Number Still Ignored
Three errors stacked on top of one another recreate the conditions of the last major credit event. The market prices the AI expansion as a technology cycle while its financing increasingly follows the mechanics of a credit-and-real-estate cycle. It watches levels and growth rates while the structures break on acceleration. And it treats a concentrated set of counterparties as if the exposure were diversified forward demand.
Any structure whose ongoing serviceability depends on refinancing into continued growth does not need an outright decline. It needs only a deceleration below the threshold the refinance requires. That deceleration appears to be underway in several of the key series even while absolute numbers remain at records.
I do not claim to know the precise timing. Borrowed time can stretch. Larger private rounds, strategic backstops, or continued leverage capacity at the hyperscaler level can extend the window. None of those stabilizers is infinite. Rating agencies have only so many notches left. Private capital pools, while deep, are finite relative to compounding annual burn measured in tens of billions.
The public market remains the terminal refinance of last resort for the most exposed borrowers. An IPO process forces disclosure of exactly the fragilities that made the private rounds necessary in the first place. The document that unlocks the deepest capital pool is also the document that prices the risk most transparently. That tension is already visible in delayed timelines and compressed step-ups.
Perhaps the most interesting aspect is how little attention the second derivative still receives in mainstream discussion. Levels and growth rates dominate the conversation. Acceleration is treated as a secondary curiosity rather than the variable on which the financing architecture actually depends. That blindness is familiar. It was present the last time a large set of negatively convex structures sat on top of an assumption of perpetual acceleration.
The current cycle does not have to end in a dramatic collapse to produce significant stress. A period of slower growth in the key drivers, combined with the existing leverage and concentration, is enough to force repricing of contracts, compression of multiples, and a reshuffling of ownership of the physical assets. The technology itself can continue to advance. The financing structures built around the assumption of unbroken acceleration may not travel the same path without adjustment.
In my experience watching credit cycles, the moments that look most comfortable are often the ones in which the second derivative has already turned. The headlines remain positive. The absolute numbers still climb. Yet the assumptions embedded in the debt, the leases, and the residual supports are already being violated. The market has a habit of noticing only after the first derivative follows the second through zero. By then the window of borrowed time has closed.
Investors who continue to focus exclusively on revenue growth, backlog size, and installed capacity may find themselves surprised by the speed with which credit conditions can change. The instruments now being created around compute as an asset class will test that focus. Collateralized compute obligations can extend the game for a period, much as earlier structured products did. Whether they ultimately keep the acceleration narrative intact or simply delay the reckoning is the question the next several quarters will begin to answer.
The architecture is already in place. The second derivative is the number that will decide how it performs. Most of the market is still not watching it closely enough.
