While the market debates whether AI is repeating the tech bubble of 2000, a credit architecture has been quietly assembled, bearing a striking structural resemblance to the 2008 subprime crisis.

This week, Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create a "computing power financing platform," aiming to mobilize over $500 billion in third-party capital. The core logic is that GPU chips can be placed into special purpose vehicles, used as collateral, and financed based on their expected future cash flows. This mirrors the model used for financing buildings and toll roads during the subprime crisis.

Simultaneously, CoreWeave completed its first publicly syndicated GPU-backed financing facility, DDTL 5.0, in May 2026, amounting to $3.1 billion, with underlying borrowers including OpenAI and Cohere. "Collateralized Compute Obligations" (CCOs) have officially arrived.

The credit market is already pricing in the collapse of token costs, yet the stock market remains euphoric about the computing narrative. This gap mirrors the period in 2006 when subprime default rates began to rise, yet home prices were still hitting new highs. Back then, the market mistakenly believed that a price decline was the crisis trigger, ignoring the real catalyst: the turning point in growth rate, not its complete collapse.

The Core Misjudgment: Is AI a Tech Cycle or a Credit Cycle?

According to a July analysis from the Groundbreaker website, the current AI boom is not a technology innovation cycle but a credit-driven capital formation cycle structurally similar to commercial real estate. The logic of their breakdowns is fundamentally different: a tech cycle ends when products are uncompetitive or eliminated, allowing for a slow de-rating. A credit and real estate cycle, however, breaks due to a slowdown in demand growth. It does not require a collapse in demand; only for growth to stop accelerating for the structure to fracture.

This is the essence of the "second derivative." The core variable (S) and its first derivative (growth rate) are the only numbers the market watches. The second derivative, whether the growth rate itself is still accelerating, is almost entirely unmonitored. However, any financing structure with continuously accelerating growth as a built-in premise dies precisely at the inflection point of the second derivative, not when the first derivative reaches zero.

From a data perspective, the ratio of capital expenditure to operating cash flow for hyperscalers has risen from approximately 30% in 2022 to about 60% in 2025, and is expected to hit 100% by 2026 based on consensus estimates. Beyond this point, every marginal investment in computing power must rely on debt or equity financing. The acceleration (second derivative) of total capital expenditure peaked and turned negative in 2025: the growth rate narrowed from roughly +81%, and the acceleration shifted from about +30 percentage points to around -25 percentage points. While the level is still at an all-time high and the growth rate remains positive, the acceleration has already reversed.

Computing as Collateral: Anatomy of the CCO Structure

Nvidia's actions this week amplify this architecture to a $500 billion scale. Jensen Huang has stated clearly that tech chips have become collateral. Nvidia itself provides residual value guarantees of up to 25% on some transactions, a structure identical to leveraged leases.

Every feature of this architecture is a computing version of real estate development: data centers are physical collateral, take-or-pay contracts are leases, the remaining performance obligations (RPO) represented by contracts are the loan book, capital expenditure is the loan principal, and computing usage fees are debt service payments. When the media interprets the record capital expenditure of hyperscalers as "demand confidence," it is actually reading the scale of lending. The market is celebrating loan growth, calling it revenue growth.

The "originate-to-distribute" mechanism is completing its cycle. CoreWeave's DDTL 5.0 is the first publicly syndicated GPU-backed instrument, featuring a bankruptcy-remote financing subsidiary that has entered the secondary market. This means that risk tied to the credit quality of OpenAI has moved off the original holder's balance sheet and spread to the broader credit market, a path highly reminiscent of the CDO distribution route from 2005 to 2007.

The "Naked Borrower": A Credit Anatomy of OpenAI

Every subprime cycle has a borrower that can only repay old debt with new financing and can never truly pay off the principal. In this cycle, that role is filled by OpenAI.

According to Groundbreaker's analysis, OpenAI's revenue structure is about 60% from consumers, making it more fragile than enterprise revenue. Its cash burn rate is about 57% of revenue, with cumulative cash consumption projected to approach $115 billion by 2029, with no path to positive cash flow this decade. Its solvency does not depend on profitability but on each valuation round being higher than the last. Valuation appreciation is its cash flow.

Looking at its funding trajectory, OpenAI's valuation has progressed from roughly $86 billion in early 2024 to approximately $157 billion, $300 billion, $500 billion, and around $852 billion by spring 2026. The corresponding round progression multiples are 1.83x, 1.91x, 1.67x, and 1.70x, while the implied IPO progression multiple is only about 1.23x. This is the lowest in the sequence and the final hurdle that must be met by the public market. The IPO delay is a direct result of this arithmetic logic.

OpenAI has no real co-signer. Microsoft removed all structural support in April 2026, ending revenue sharing, relinquishing exclusive rights, and abandoning priority computing supply rights, retaining only a 27% equity upside. SoftBank is attempting to raise a $10 billion margin loan secured by its OpenAI holdings, requiring personal guarantees, indicating that even the co-signer needs a co-signer. In the eyes of the best-informed counterparty, Microsoft, OpenAI's credit is no longer worth underwriting. Its actions are a signal in themselves.

In contrast, Anthropic's credit structure is fundamentally different: about 80% of its revenue comes from enterprise clients, it generates $1.70 in revenue for every $1 spent on computing, and it has substantial credit backing from Google and Broadcom for its ~$35 billion TPU financing facility, synthetically raising its credit quality to investment grade. The essential difference between these two frontier AI model companies is "secured credit" versus a "naked borrower."

The $2.1 Trillion Bill: Concentration Risk and the Correlation Trap

According to Groundbreaker estimates, the combined RPO of the four major hyperscale cloud platforms is approximately $2.1 trillion, with about half, or roughly $1.05 trillion, coming from OpenAI and Anthropic. These two companies account for about 49% of Microsoft's ledger, around 54% of Oracle's (with OpenAI alone at about $300 billion), about 43% of Google's, and roughly 51% of Amazon's.

This is the same structural root cause of the fatal flaw in senior CDO tranches of 2008: thousands of seemingly diversified mortgage loans were all exposed to the same macro variable, national home prices. When that variable turned, the correlation of all assets instantly went from zero to one. In this cycle, that macro variable is not home prices but whether OpenAI can complete its next funding round. Reports suggest chip stocks fell when OpenAI’s IPO was rumored to be delayed from 2026 to 2027, confirming this transmission chain.

CoreWeave's DDTL 5.0 is securitizing and distributing this concentrated risk, which is highly correlated with the credit quality of OpenAI. The instrument's repayment does not depend on OpenAI's profitability but on its ability to continue raising capital, which in turn is based on the continued reinforcement of the AI capital expenditure narrative.

Reversal of the Nash Equilibrium: Who Blinks First?

The capital expenditure arms race is a Nash equilibrium, but conditional. It exists only as long as the market continues to reward each incremental computing investment, pricing it as a growth option. If a hyperscaler announces a cut in capital expenditure and its stock price rises instead of falls, discipline is rewarded, not punished, and the entire game matrix is rewritten. Spending becomes a penalized strategy, and restraint becomes a signal for re-pricing. This equilibrium reversal is not gradual but a coordinated mutation. The first cutter to be rewarded provides cover and incentive for all other CFOs.

A Goldman Sachs Delta One trading head stated directly: "The first hyperscaler to signal it can slow its spending pace will likely see its stock rewarded (and semiconductor stocks hit). If that happens, others will follow. This is the reflexivity that ultimately stalls the capex cycle, not a lack of demand, but investors deciding the marginal return on the next dollar of spending is no longer attractive."

From stress tests, Google, with its structural cost advantage from TPUs and over $460 billion in cloud backlog, would benefit if competitors pull back and is unlikely to act first. Oracle, with capital expenditure at about 86% of sales, has a highly stretched balance sheet, and its pressure will manifest as a credit event rather than a voluntary choice. Amazon's free cash flow has turned negative, and it may be forced by capital markets. Meta's Zuckerberg controls dual-class voting shares and has no public cloud business selling computing power, making its capex defense the weakest and exit the easiest; it is the most likely to blink first.

The Sequence of Fracture: Mechanism, Not a Calendar

Groundbreaker outlines a clear chain of transmission for the fracture: (1) The final refinancing cannot be priced at the required progression multiple, with the OpenAI IPO delay being the signal. (2) OpenAI cuts computing commitments to preserve cash, triggering take-or-pay contract defaults. (3) Neocloud platforms like CoreWeave and Lambda are hit first, as they are essentially carry trades with no second cash flow stream; a margin compression means structural insolvency. Oracle absorbs the concentration shock, but it manifests as impairment rather than payment cessation. (4) The credit market freezes, with RPOs re-priced from forward demand to counterparty risk, GPU-backed notes cannot be rolled over, and the originate-to-distribute mechanism seizes up. (5) The equity market is severely impacted, with capital expenditure re-priced from an option to a cost, compressing valuations across the supply chain. (6) The most capital-rich participants acquire distressed data centers at low prices, taking over leases that must be continued.

Notably, the analysis acknowledges uncertainty in the timing. A new large private market funding round, sovereign or strategic capital support, or hyperscalers leveraging their own balance sheets can all buy "borrowed time." However, Morgan Stanley estimates that investment-grade hyperscaler total debt leverage has doubled in a year to about 1.8x, higher than the entire energy sector, and this does not include hundreds of billions in off-balance-sheet vehicles. Rating agency headroom is limited.

Historical Echoes: The Triple Misjudgment

Groundbreaker's conclusion is a triple misjudgment, each individually manageable but together precisely replicating the preconditions of the last major credit event. First, the market classifies AI as a technology cycle, while its financing mechanism is that of a credit and real estate cycle. Second, the market focuses on levels and speeds, while the structure fractures at the acceleration. Third, the market prices a highly concentrated loan book, dominated by a single borrower, as diversified forward demand.

The lesson of 2008 has never been accurately remembered. Defaults did not begin when prices fell; they began when prices stopped rising faster, as early as 2006 when subprime default rates quietly rose while home prices were still at record highs. This time, the financing architecture is exquisitely designed, and it fractures precisely at the unmonitored number: the acceleration of the growth rate.