The Information称英伟达5000亿美元AI算力融资计划仍处MOU和结构设计阶段

Nvidia shocked the market earlier this month when it lined up half a dozen financial giants, including Blackstone, Apollo and Goldman Sachs, to finance $500 billion of AI infrastructure purchases.

The eye-popping figure was a signal to the market that Nvidia is moving away from acting as the financier of last resort for AI firms. Nvidia CEO Jensen Huang made an unusually public pitch to sell the idea, including a rare CNBC appearance in a roundtable with senior executives from all six financial firms, and a blog post declaring that “AI factories” built on Nvidia hardware are becoming an “investable asset class.”

For all the fanfare, there were only a few clues about how the initiative would work. Nvidia suggested lenders could make loans against the cash its compute generates. Financial firms would underwrite individual deals by assessing factors including residual value—what assets are worth once a loan is paid off. And Nvidia said it would provide residual value support of up to 25%, covering part of the loss in the event of a default if assets backing loans fall, though it didn’t specify which assets.

I still had plenty of questions, so I spent the past week trying to understand what devil might be lurking in the details. What emerged was a plan still very much in development. The firms have signed memorandums of understanding but haven’t executed anything, and very few people with any knowledge of the actual plans wanted to talk.

An Nvidia spokesperson said the topic will likely come up in the company’s earnings this week, “so recommend giving that call a listen.” Representatives for the six firms didn’t offer any additional statements.

Still, some market participants are filling in the blanks, relying on existing financing models and their knowledge of the industry to make educated guesses. At a high level, the announcement points to Nvidia’s need to find new ways to help customers finance graphics processing unit purchases they can’t easily fund on their own.

“The challenge for the vendors of these GPUs is, how do we extend the runway for growth?” said Andy Li, a semiconductor analyst with CreditSights. “The industry collectively has been trying to build out more of this infrastructure on the back of debt financing.”

And the closest likely analogue is a [$35 billion financing deal](https://www.theinformation.com/articles/inside-broadcoms-bold-move-boost-demand-chips)announced in June to finance Broadcom-designed Google tensor processing units. One person I spoke to suggested that Broadcom’s structure may have encouraged Nvidia to consider something similar but with a bigger splash.

The Broadcom deal tapped techniques from the securitization market, where a special entity raises money to buy assets on behalf of a company. The company using those assets pays for them over time, often splitting the debt up into tranches, or levels of risks and returns for investors.

In this case, Apollo teamed up with Blackstone to raise debt, buy the TPUs and lease them to Anthropic, helping the private, unrated company secure far more chips than it likely could have on its own. Broadcom agreed to backstop $29 billion of the overall debt through a residual value guarantee, making up the difference if the chips were worth less than lenders expected. The riskiest portion of debt didn’t have a Broadcom backstop.

Broadcom’s support was key to making that first deal happen, with its guarantee covering roughly 82% of the overall deal. But that number may be shrinking for future deals, putting Broadcom’s credit support closer to what Nvidia is envisioning. Bloomberg recently reported that a second Broadcom financing under discussion for more than $60 billion could include a $30 billion junior tranche—the riskiest debt that’s first to absorb losses.

What specific deals for Nvidia chips could look like will likely vary across investment firms. But one person close to the consortium said some of the debt issued would likely mature in five years or less and be paid down quickly by cash generated by the compute, giving debt investors a level of comfort with using GPUs to back the loans. Any value the chips retain after the debt is repaid would be potential upside, the person said.

Overall, the goal is to structure financing that relies largely on the underlying value of GPUs themselves, according to two people I spoke to. So far, when GPU purchases have been financed—say, for neoclouds to buy chips—there’s usually a contract or other backstop from an investment-grade customer like Microsoft or Meta Platforms. Smaller AI companies without close ties to such benefactors largely haven’t been able to access the same financing.

“There is still a little bit of a gap that exists,” said Wayne Nelms, founder of compute-data startup Ornn. “Being able to finance thousands and thousands of nodes of GPUs is one thing, but making that available for smaller players with smaller balance sheets is a whole other thing. I think this is the first step, certainly.”

A key part of this overall initiative is proving that Nvidia’s GPUs have a longer useful life than what many lenders currently assume. If the intended user of GPUs can’t pay back the cost for the chips, lenders want to have a pool of customers willing to step in to buy or lease them.

Many people expect GPUs to depreciate quickly, in part because of how quickly Nvidia has introduced newer chips with better performance. And their performance can vary widely based on factors such as power supply, how they are networked, and what optimization software is used. That worries lenders who would prefer a standard chip with a long life that can act as stable collateral to back loans.

Huang, for his part, highlighted in the blog post how Nvidia’s software, networking and broad customer base make GPUs “fungible,” or easy to shift to different customers, while software upgrades can extend how long the hardware can remain useful. Nvidia-backed neocloud CoreWeave, a major buyer of GPUs, recently chimed in on this idea, saying it had signed a contract for A100s, a chip first introduced in 2020, that extends into 2029.

One development that could make lenders more comfortable with that idea is the development of a forward curve, a market view of what GPUs will be worth in the future. Data providers such as Ornn and Silicon Data are already providing indices that begin to provide that kind of visibility.

Meanwhile, CME Group, in partnership with Silicon Data, hopes to launch futures contracts for rental prices of Nvidia’s H100 and B200 chips in October. Such derivatives could let lenders offset some risk by taking short positions that gain value if GPU rental prices fall.

Still, it could be a while before derivatives play much of a useful role in putting the entire financing picture together. One person familiar with the development of past markets suggested that the six firms in Nvidia’s announcement, plus other investors they bring in, would likely need to establish the underlying financing market before a robust market for hedges around those deals can develop.

### New From Our Reporters

[https://www.theinformation.com/articles/nvidia-discusses-perplexity-investment-30-billion-plus-valuation-considered-tech-licensing-deal](https://www.theinformation.com/articles/nvidia-discusses-perplexity-investment-30-billion-plus-valuation-considered-tech-licensing-deal)Exclusive

#### [Nvidia Discusses Perplexity Investment at $30 Billion-Plus Valuation](https://www.theinformation.com/articles/nvidia-discusses-perplexity-investment-30-billion-plus-valuation-considered-tech-licensing-deal)

By Phoebe Liu, Valida Pau and Amir Efrati

[https://www.theinformation.com/articles/nvidia-using-land-electricity-deals-lock-hardware-bundle](https://www.theinformation.com/articles/nvidia-using-land-electricity-deals-lock-hardware-bundle)AI Infrastructure

#### [Nvidia is Using Land and Electricity Deals to Lock In Its Hardware Bundle](https://www.theinformation.com/articles/nvidia-using-land-electricity-deals-lock-hardware-bundle)

By Ann Davis Vaughan

[https://www.theinformation.com/articles/nvidia-ai-chip-prices-rise-17-server-makers-tell-customers](https://www.theinformation.com/articles/nvidia-ai-chip-prices-rise-17-server-makers-tell-customers)#### [Nvidia AI Chip Prices to Rise About 17%, Server Makers Tell Customers](https://www.theinformation.com/articles/nvidia-ai-chip-prices-rise-17-server-makers-tell-customers)

By Phoebe Liu and Amir Efrati

Dakin Campbell is a reporter at The Information, where he covers the money behind the artificial intelligence boom, and other financial topics. He can be reached by email at [[email protected]](https://www.theinformation.com/cdn-cgi/l/email-protection) or on Signal at dakin.11.