On August 10, 2026, NVIDIA announced plans with six financial institutions to mobilize more than $500 billion of outside capital for AI computing infrastructure over time. On Moonshots with Peter Diamandis, in an episode published August 18, guest Alvin Graylin described the arrangement as a way “to essentially securitize chips and compute.” His verdict: “This sounds a lot like the subprime issues.”
The announcement was narrower than that description. NVIDIA had signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR for independent financing platforms serving customers across its ecosystem. The partnerships remained subject to final agreements. This was not NVIDIA borrowing $500 billion itself, or a report of completed securities sales.
But the proposal puts a concrete question behind Graylin’s warning: what happens to infrastructure financed against future computing income if AI services become much cheaper?
His argument is not that AI will disappoint as a technology. It is that the companies paying for the infrastructure may not be the ones who capture its value.
A useful product that loses pricing power
Graylin’s starting point is commoditization: competing products become good enough substitutes that customers have less reason to pay a premium for one of them.
Open-weight models—models whose learned parameters can be downloaded—give businesses an alternative to buying access to a proprietary model. Some can run on a user’s device or a company’s own machines. That process of using a trained model to produce answers is called inference; running it locally can avoid paying a cloud provider for each request, though it still requires hardware and electricity.
“As more and more of the capabilities move to open source, move to edge computing, the dependency on cloud-based premium services will continue to erode,” Graylin said. Edge computing means doing the work closer to the user rather than in a remote data center.
His financial reasoning has several steps. Cheaper alternatives weaken the prices premium providers can charge. If customers also move work onto their own machines, some cloud equipment may spend less time doing paid work. Lower prices, lower utilization—or both—can reduce the income available to cover operating costs, interest and debt repayment.
AI could therefore spread through the economy while disappointing investors in particular laboratories or computing facilities. In Graylin’s account, useful applications and their customers could capture benefits that infrastructure owners had expected to keep.
Future payments are the bet
NVIDIA’s case for the proposed platforms is that computing capacity can serve different customers and generate usage-linked revenue, making it an investable asset. Outside financing would help customers obtain infrastructure without requiring NVIDIA to supply all the capital itself.
Graylin’s subprime analogy targets confidence in those future payments. Securitization means packaging rights to payment streams into investments that can be sold. If the underlying income falls short, the financing built around it can come under pressure. His comparison is a warning about that mechanism, not evidence that NVIDIA’s proposed platforms have already created mortgage-crisis-style securities.
To convey the scale of his concern, Graylin cited a figure he revised during the conversation from $1.6 trillion to $1.7 trillion, describing it as off-balance-sheet debt at the largest cloud companies. The $1.7 trillion should be understood as his claimed estimate of commitments, not an established total of outstanding debt or a valuation of the AI sector.
He also contrasted annualized revenue headlines with revenue actually earned over completed quarters. A run rate extends a recent pace of sales across a full year; it is not a year’s revenue already realized. A fast-growing company can legitimately report a run rate much larger than its earlier revenue. The financing question is whether future sales produce enough cash to meet its commitments, not simply whether those two figures differ.
The hardware can still earn
There is a counterargument from an earlier Moonshots discussion of the financing plan. Emad Mostaque recalled operating 10,000 NVIDIA A100 processors and said a cloud provider had reported customers signing A100 contracts through 2029. A customer with a stable job that fits an older chip does not necessarily need the newest hardware.
His economic argument depended on the equipment having already earned back its purchase price. Once that cost is recovered, a machine can keep earning by serving suitable workloads. Mostaque emphasized electricity as the additional cost of doing more computing, although maintenance and other operating expenses remain.
That leaves room for productive older equipment even in a market with falling prices. It also separates two risks: a chip becoming technologically outdated, and a still-useful chip failing to earn enough to support new borrowing.
Cheaper AI could also encourage enough additional use to offset lower prices. Graylin’s scenario depends on that growth failing to deliver sufficient revenue to the businesses carrying the commitments. Local computing need not eliminate cloud demand for his argument to matter; equally, falling model prices alone do not establish that data centers will struggle to repay their loans.
From a revenue squeeze to a construction slowdown
Graylin reaches for a Cold War analogy to describe the broader danger. In his telling, the Soviet Union lost by spending itself into weakness during an arms race. He worries that competitive pressure could similarly push the United States to build more AI infrastructure than its eventual revenues justify. “I hope it doesn’t happen,” he said, “but I think we are pushing ourselves in that way.”
His July essay, The Great Reckoning Before the Reconnecting, develops a speculative crisis scenario around leveraged infrastructure commitments and declining model pricing power. It acknowledges that vendor financing can help coordinate scarce supply, while judging the demand assumptions too optimistic.
In the podcast, the trigger he anticipates is clearer financial disclosure, including if private AI companies go public. Investors might discover that the value created by AI is not translating into the income they expected at the companies they financed. Graylin’s predicted consequence is a slowdown in construction: fewer new facilities as confidence weakens that their customers can keep paying and servicing debt. In that scenario, AI keeps finding uses, but the next data center becomes harder to fund.