The most important number in the artificial-intelligence boom may no longer be the price of an NVIDIA GPU. It may be the interest rate required to finance one.
Over the past two years, the AI infrastructure buildout has expanded beyond the balance sheets of Microsoft, Google and Meta into a new class of specialized cloud companies, data-center developers and project-finance vehicles. These companies are building enormous amounts of computing capacity ahead of demand, often using debt secured by GPUs, customer contracts or data-center assets. As long as capital remains plentiful, the model can expand at extraordinary speed. The vulnerability appears when lenders begin demanding substantially more compensation for financing the next dollar of capacity.
CoreWeave provides one of the clearest places to watch that transition. Its August $2.6 billion delayed-draw facility carries interest at SOFR plus 5.5 percentage points. The company simultaneously expects to spend roughly $35 billion to $39 billion on capital expenditures during 2026, against a revenue base growing rapidly but still far smaller than the infrastructure investment required to support it. CoreWeave's backlog reached $104.2 billion in the second quarter, providing substantial evidence of demand, but the business remains dependent on continued access to external capital.
That makes CoreWeave more than an individual-company story. It can function as an early indicator of the marginal cost of non-investment-grade AI infrastructure capital. The question is not whether CoreWeave can finance itself today—it plainly still can. The question is what happens if the next financing costs materially more, requires more collateral or equity, or proves difficult to syndicate.
The first place the damage would appear outside CoreWeave is among companies that have built assets around its contracts.
Core Scientific is an unusually direct example. CoreWeave currently accounts for all of Core Scientific's colocation revenue and approximately 77% of total revenue for the first half of 2026. The contracts have structural protections and CoreWeave remains a primary obligor, but Core Scientific nevertheless acknowledges that its liquidity and cash flows are highly dependent on CoreWeave's financial health and performance.
Applied Digital presents a related but somewhat broader risk. It has contracted roughly 400 MW of capacity at its Polaris Forge 1 campus to CoreWeave, with buildings still moving through construction and commissioning. That means deterioration in CoreWeave credit would immediately raise a second question: not simply whether the leases remain legally enforceable, but what those future cash flows are worth if the tenant financing them has become materially riskier.
This is the first transmission test.
CoreWeave credit deteriorates. Then watch Core Scientific and Applied Digital.
If their financing remains stable, lenders continue treating the contracts as high-quality collateral and replacement customers remain plentiful, the problem may stay contained. If lenders begin repricing those assets, demanding additional credit enhancement or questioning the value of future lease payments, the problem has moved beyond a single neocloud.
The second test occurs in the project-finance market.
Oracle's Project Jupiter is already providing a useful signal. Reuters, citing the Financial Times, reported on September 18 that roughly $18 billion of loans associated with the New Mexico data-center project were being quoted around 89 to 91 cents on the dollar, while banks were having difficulty distributing some of the debt. Those indications do not constitute a default, nor should the project debt simply be treated as Oracle corporate debt. But they show that investors are beginning to differentiate among AI infrastructure credits rather than treating every large data-center project as essentially money-good.
IREN provides an especially useful comparison because it has financed different classes of customers at dramatically different prices. Financing associated with its Microsoft contract was arranged at roughly 6%, while $2.4 billion supporting deployments for non-investment-grade customers costs approximately 9% fixed. Customer prepayments and financing currently cover most of the associated GPU spending, and contracted pricing has actually been rising—important counter-evidence to the bearish case.
But that three-percentage-point financing gap is precisely what deserves attention.
If investment-grade AI infrastructure continues borrowing around 6% while weaker projects migrate from 9% toward 11%, 12% or higher, the economics of the boom begin to separate. Projects do not have to default for the cycle to change. They only have to become expensive enough that anticipated returns no longer justify construction.
That is when the third transmission channel opens: orders.
Data-center developers facing higher capital costs can delay a campus. A neocloud can stretch the replacement cycle of its existing GPUs. A customer can postpone an incremental cluster. Those decisions eventually appear as weaker orders for NVIDIA GPUs, Supermicro and Dell servers, Credo and Astera connectivity products, optical components and networking equipment.
This is where a credit problem becomes an industrial problem.
The sequence matters because hardware stocks can fall for many reasons. A weak semiconductor week does not establish an AI-credit contraction. But if credit spreads widen first, new financings become more expensive, project syndications struggle, developers delay capacity, and only then server and networking orders weaken, the evidence becomes considerably stronger.
NVIDIA itself sits in an unusual position at the center of this system. Its latest filing shows that it is no longer merely selling chips into the AI boom. NVIDIA has agreements under which it may purchase cloud capacity that AI-cloud providers fail to sell to third parties, and it warns that it may not always be able to use or resell that capacity economically. Its supply commitments reached $279 billion by late July.
That introduces an element of reflexivity. Some of the companies buying NVIDIA hardware depend upon financing. NVIDIA increasingly has economic interests in helping the surrounding infrastructure ecosystem expand. Strong demand enables financing; financing enables infrastructure construction; construction creates GPU purchases; GPU purchases reinforce expectations of strong demand.
A contraction could run through the same chain in reverse.
Only after that would we expect the fourth stage to become visible in the physical infrastructure surrounding the data centers: cooling systems, electrical equipment, racks, substations and power infrastructure. Companies such as Vertiv and Celestica sit farther from the initial credit shock because many of their projects are already under construction and difficult to cancel immediately. Their deterioration would therefore be more significant precisely because it would arrive later. Falling book-to-bill ratios, cancellations and shrinking backlogs would suggest the slowdown had progressed from financing into the actual physical buildout.
This gives the AI infrastructure boom a surprisingly simple diagnostic sequence:
CoreWeave credit → Core Scientific/Applied Digital → Oracle/Jupiter and IREN financing → servers and networking → power and cooling.
None of these links currently proves that the cycle is breaking. There is substantial counter-evidence: CoreWeave continues raising capital and expanding backlog; IREN reports increasing contract pricing and largely sold-out 2026 capacity; major AI companies continue committing extraordinary sums to infrastructure.
But the relevant question has changed.
It is no longer simply whether demand for artificial intelligence continues to grow. It is whether the financial system will continue providing increasingly large amounts of capital, at acceptable prices, to the weaker borrowers required to turn that demand into physical infrastructure.
If the answer begins to change, the earliest warning may not come from an earnings miss or a collapsing technology stock.
It may come from the price of the next loan.