Wall Street wants to turn artificial intelligence compute into an investable asset class. This promises to unlock capital for a costly technological transition, addressing capital scarcity, which represents one of the limits currently imposed on the AI boom . Nearly twenty years ago, the industry also saw an opportunity to bundle subprime mortgages into securities and, in doing so, the financial system began distributing risks that were poorly understood.

In a recent Wall Street Journal opinion piece , BlackRock CEO Larry Fink highlighted the risk to the AI boom presented by a limited supply of capital. He is advancing a vision in which the capacity embodied in chips and data centers can attract capital far beyond the companies building them. This mirrors what Nvidia CEO Jensen Huang told CNBC in August when he stated that the company’s “AI factory platform is really an investable asset.” Financing could be raised against equipment and contracted computing revenues, with standardized instruments eventually making exposures easier to pool, rate and trade.

Strain Is Showing on AI Financing

A look into CoreWeave, a cloud provider that rents AI computing capacity, exposes how strong AI demand can coexist with financial strain and highlights the limits of market risk tolerance. The company used its $104 billion in contracted revenue backlog as proof of a healthy business model, but delivering that capacity will require substantial investments in advance of customer payments. Servicing the debt is widening the gap to profitability, with second-quarter net interest expense reaching almost a quarter of the firm’s revenue. Jim Osman argues here at Forbes that the focus should be on how much cash remains after equipment, power and creditors are paid.

By mid-October, unsecured bonds issued in June with a 9.625% coupon traded below 80 cents on the dollar, yielding more than 15%. The market is signaling growing concern about repayment as the neo-cloud company continues to borrow billions against infrastructure and customer commitments. In addition to concerns about debt volume, its profile is also raising eyebrows. CoreWeave’s August $2.6 billion facility runs over five years, longer than the average three-year contracts supporting it. Lenders are being asked to underwrite contract renewal risk and make a bet on the long-term value of Nvidia’s processors CoreWeave owns.

CoreWeave is financing its own expansion. Fink’s concept differs in scale and purpose, creating a standardized market for AI compute to enable operators to bundle claims on compute and sell them to a wider range of investors. Securitization would create additional capacity to lend.

Pooling these exposures would redistribute that risk without eliminating it.

AI Securitization And The Subprime Analogy

Before 2008, mortgage securitization expanded credit, converting loans into securities. Optimistic valuations, weakened underwriting and leverage concealed correlated exposure to what were supposed to be diversified risk pools.

Compute financing, while not identical, has its own vulnerabilities. Nvidia’s processors depreciate faster than houses. The rental income from the AI factories depends on their utilization, pricing and customers’ ability to turn AI services into profitable businesses. The frontier labs are the largest consumers of AI compute capacity and are still struggling to sustain profitable operations. In a highly concentrated market, several borrowers depend on the same customers, equipment suppliers and all are betting on assumptions about sharply rising future demand.

Chip suppliers invest in infrastructure providers, which in turn buy their chips and AI developers commit to purchasing future capacity. These circular funding relationships may reflect legitimate demand, but they make it harder to assess how much demand is independent of the financing that supports it.

In the past, we allowed mortgages to have a broad effect on household credit and banking balance sheets. To avoid similar risk, proposals to open today’s AI infrastructure financing beyond a specialized investor pool require scrutiny.

What The Post-2008 Safeguards Don’t Cover

Since the financial crisis, regulators require banks to hold more liquid assets, and under the Dodd-Frank Act, firms that package loans into securities have to have “skin in the game” and bear some of the credit risk. Those rules primarily affect the banking sector and its products, but much of the financing of AI infrastructure is happening via private-credit funds and insurers, a market with much less transparency.

Last May, the Financial Stability Board, an international body established under the G20 to monitor the global financial system, warned against the vulnerabilities in private credit markets and signaled a lack of data to execute its oversight function. The FSB urged authorities to “deepen analysis of interlinkages of private credit with private equity and insurers and of liquidity mismatches in private credit funds.”

The recent pressure on CoreWeave illustrates how quickly investors can reassess AI financing. A stressed borrower, by itself, does not constitute a systemic event, but an interconnected and opaque web of vendor financing and concentrated demand can amplify the shock before regulators are able to stop it.

The Case For Transparency In AI Financing

Securitization should not be banned, nor should one assume every new funding instrument carries the seeds of a future crisis. The AI boom requires newer sources of capital and, well-operating and properly regulated markets allocate risk efficiently.

The lessons of the last financial crisis inform us what to inspect. Supervisors and regulators need visibility into exposures across banks and other financial institutions, including guarantees and indirect leverage. They should watch for maturity mismatches, especially in funds that promise investors quick access to their money while holding loans that cannot easily be sold. Lenders should be required to keep enough of the risk on their own books to care whether the loans are sound.

Much is being bet on continued demand growth and on the compute assets maintaining their value over time, despite the typical fast pace of technological obsolescence. An independent third-party assessment of the chip valuation over time would keep lenders from treating currently inflated prices in a supply-constrained market as proof that the hardware will hold its price.

Investors should be informed about how heavily the compute infrastructure is being used, how dependent it is on a small set of customers and how long existing customer contracts last. Stress tests can simulate what happens when computing prices fall, chips become obsolete sooner than planned or when a major customer walks away.

A financial instrument can work for years while spreading a common vulnerability. We learned this in 2008. We should demand evidence before transforming projections of future compute demand for AI into assets marketed as safe.