NVIDIA Corporation (NVDA) is reportedly moving forward with plans to securitize its data center assets, a transaction some analysts warn could push the AI computing boom into dangerously familiar territory.
According to Bloomberg macro strategist Simon White, NVIDIA's initiative to convert data centers into investable assets risks triggering severe overcapacity in the AI computing sector, especially as outcome-based pricing models gain traction. The core logic of this warning is that financialization is artificially amplifying demand signals, masking the reality of an already oversupplied market—a pattern strikingly similar to the housing bubble that preceded the 2008 financial crisis.
White highlights that NVIDIA has established a framework with partners including Apollo, Brookfield, KKR, Blackstone, BlackRock, and Goldman Sachs, aiming to channel over $500 billion into so-called "AI factories." Under the terms of some deals, NVIDIA will insure up to 25% of the GPU collateral, effectively guaranteeing its recovery value. The parallels to the housing bubble are difficult to overlook. At that time, American International Group Inc (AIG) sold insurance on the senior tranches of collateralized debt obligations (CDOs), allowing banks to reduce capital requirements and distribute more toxic debt—a cycle that ultimately ended in a government bailout when correlations spiked.
The AIG Moment for GPU-Backed Bonds: Financialization Accelerates
Simon White argues that the scale of the current AI infrastructure boom already exceeds historical expansion cycles like the U.S. interstate highway system (1960s) and broadband deployment (2000s). To sustain this momentum, a series of cross-financing and vendor-backed deals have already been executed, with transactions involving NVIDIA alone reaching approximately $750 billion.
NVIDIA CEO Jensen Huang has explicitly stated that the goal is to establish "AI factory computing" as a distinct investable asset class. The market for GPU-backed debt has already exploded this year, and this new agreement pushes the process further into the realms of asset securitization and distribution. The partners will provide capital, infrastructure expertise, and facilitate the securitization and distribution process.
By making the transactions more attractive, NVIDIA will enable the senior tranches of the debt structures to achieve higher credit ratings. This development coincides with the U.S. Securities and Exchange Commission (SEC) having just relaxed securitization rules for data center bonds. White compares this to the moment during the financial crisis when pension funds finally entered the market—a point typically signaling the beginning of risk spreading to a broader liability base and the onset of the crisis's final stages.
On the surface, NVIDIA's risk may appear limited. However, White cautions that this was precisely the same assessment made by AIG. Over time, AIG took on an increasing volume of obligations it believed to be risk-free and profitable, only to discover that the rising correlation in borrower credit stress was sufficient to cause a catastrophe.
Signals of Overcapacity: Supply-Demand Imbalance Masked by Financial Alchemy
White's central concern is that financialization is systematically obscuring the market's overcapacity signals. Similar to the housing crisis, the current capital expenditure boom is more likely to attract increasing capital flows lubricated by financial alchemy, leading to overbuilding. The signals that computing power is already abundant will likely be difficult to draw attention to.
On the demand side, White points out that the measurement standards themselves are undergoing a fundamental shift. The standard metric for GPU leasing is the hourly cost, while for inference, it is the cost per token processed. However, for end users, what truly matters is the output—the value of intelligence per dollar spent, or more directly, the cost to complete a task.
The cost of intelligence is declining, but this change is not immediately reflected in GPU-level pricing. If a model can accomplish the same task using ten times fewer tokens, while the per-token pricing for the GPU remains unchanged, the effective efficiency of each GPU has increased tenfold.
As the era of "token maximization" fades, evidence is mounting that models can significantly vary their token usage for the same task. Combined with the fact that large language models (LLMs) can be easily adjusted to consume more tokens, outcome-based pricing models are increasingly favored by enterprise users. If this trend becomes mainstream, it will pressure model providers to reduce token consumption, which in turn will depress GPU leasing prices.
Also noteworthy is the reportedly low actual utilization of GPUs. A report from Cast AI indicates that average GPU utilization is only 5%, while ClearML found that nearly half of all enterprises are "wasting millions of dollars" on idle chips.
The Limits of Jevons Paradox: Path Dependency is the Real Market Risk
Faced with arguments about efficiency gains, the usual rebuttal from AI capital expenditure bulls is the Jevons paradox: a decline in the cost of intelligence driven by computing power will inevitably drive demand higher, offsetting lower prices with greater volume.
However, White is not convinced. He argues that the Jevons paradox does not hold true across all time horizons, and for the market, path dependency is the critical factor. He assesses that during the current AI development phase dominated by LLMs, the growth rate of demand for intelligence (rather than demand for tokens) at current price levels will eventually slow enough to trigger a long-term decline in computing costs.
For lenders, risk will rapidly accumulate when the correlation of collateral begins to deteriorate and another key risk—the creditworthiness of borrowers—also weakens. At that point, NVIDIA would be forced to fulfill its insurance obligations, while simultaneously facing potential problems with other transactions it has financed, not to mention the decline in the value of its core product. Other companies, like Google, which are also supporting this capital expenditure boom with their own balance sheets, face similar risks.
As the housing crisis revealed, once correlation begins, it does not wait for anyone. White's conclusion is that NVIDIA's proposed transaction simply brings the market one step closer to witnessing this natural force of finance once again.
Comments