On 10 August 2026, NVIDIA said it had signed memorandums of understanding with six financial institutions — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — aiming to mobilize more than $500 billion in outside capital for AI infrastructure over time. The joint announcement describes proposed independent financing platforms with dedicated pools of money for NVIDIA’s customers: frontier AI laboratories, enterprises and cloud operators.
NVIDIA is not borrowing half a trillion dollars. Nor is the announcement a report of completed bond sales. The plan is to connect customers that need infrastructure financing with outside capital — helping them buy the equipment that NVIDIA sells.
On Moonshots with Peter Diamandis, that distinction opened a harder question. GPUs, the processors used to train and run AI models, can earn money for years. But how far into the future should investors count on that income when a new chip or computing method could undercut it?
Turning computing income into an investment
Dave, the first panelist asked for a reaction, saw the structure rather than the sum as the innovation. Instead of NVIDIA raising all the money by selling more shares, he envisaged standalone investments tied to individual clusters — groups of chips working together. Repeating that structure could attract money that would otherwise stay outside the AI build-out.
The panel took that logic toward what Alex called “compute-backed securities”: investments supported by the money customers pay to use computing equipment. Securitization means packaging rights to future payments into investments that can be sold. Mortgage-backed securities do this with home-loan payments; a compute-backed version would depend on computing revenue.
That was the panel’s discussion of how the market could develop, not a description of completed securities transactions under NVIDIA’s memorandums. Dave argued that prominent financial firms would help attract investors. Their participation alone, however, does not give an investment an investment-grade credit rating.
The objection: the equipment may age faster than the financing
Salim Ismail imagined a structure that packages ten years of GPU cash flows. What happens if a breakthrough arrives well before those ten years are up, making a different architecture much cheaper or more useful?
The old machines might still work, yet no longer earn what investors expected. That is the stranded-asset problem: equipment can lose its economic purpose before the financing attached to it has been repaid.
Salim’s concern was that financial projections need a reasonably predictable decline in an asset’s value, while rapidly changing technology may not provide one. Diamandis amplified the point with a hypothetical investment financed for ten or twenty years that can no longer generate the anticipated revenue. Those were examples of a potential mismatch, not announced terms of NVIDIA’s platforms.
The discussion also turned to credit ratings and the mortgage crisis. But assessing a borrower honestly and forecasting a chip’s useful life are different problems. Even without misleading ratings, an unexpected computing breakthrough could upset the revenue forecast.
The counter: productive assets, and somewhere to hedge
Alex was less worried about a repeat of the mortgage-backed securities crisis. He emphasized that computing equipment is productive: customers pay it to perform work, although it requires electricity and maintenance.
His main answer to technological risk was hedging — taking another financial position designed to offset a loss on the original investment. He argued that a developed market for computing should include options and futures, contracts that let participants manage exposure to future prices.
In his hypothetical downside scenario, a breakthrough turns a $10,000 GPU into a $100 GPU. In the opposite scenario, a supply disruption sends computing prices sharply higher. Investors and computing buyers, he argued, would need ways to manage both directions of risk before trillions of dollars could flow into infrastructure.
Alex disclosed a financial interest in a venture working on such a market. His case depended on liquid, functioning markets in which participants could actually arrange those hedges. He was not saying that chip usefulness would necessarily stay constant, or that the proposed financing platforms already came with protection against obsolescence.
Why an old chip can still earn
Emad Mostaque brought the discussion back to equipment he had operated. He recalled having 10,000 NVIDIA A100s, a chip introduced in 2020. He also said a cloud provider had reported customers taking A100 contracts through 2029.
Why commit to an older chip for several more years? In his account, a customer may have a stable workload that already fits the hardware. It does not need the newest processor to keep doing that job.
Mostaque’s economic argument depended on the upfront hardware cost having already been recovered. Amortization spreads that cost over the period the equipment is used; earning back the purchase price makes subsequent work a different proposition from financing a new fleet. Once that initial expense is covered, he emphasized electricity as the marginal cost — the additional expense of doing more computing — although maintenance and other operating costs remain.
Software can also extend a machine’s useful life. Mostaque pointed to CUDA, NVIDIA’s programming platform, as a link between older and newer generations of hardware. NVIDIA’s announcement similarly argues that software improvements and the ability to serve different workloads can sustain computing assets. Compatibility helps keep older chips useful, but it does not mean every new model fits their memory or runs economically on them.
The A100 example gives the financing debate a concrete test. A paid-off chip with a customer committed to a suitable workload may keep earning long after a newer generation arrives. A newly financed cluster still has to recover its purchase cost, find enough paying work and withstand changes in computing prices. For the proposed capital pools, the question is not simply whether AI demand grows, but whether the particular equipment they finance earns enough before customers move on.