Glonce

Nvidia's $500 billion compute-as-asset pitch doesn't add up

businessThe Glonce editorsThe Verge🔥74Glonce rating74Impact72Novelty55Relevance95Surprise62Players88Credibility82
Nvidia's $500 billion compute-as-asset pitch doesn't add up

Nvidia has lined up six major financial institutions, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, on a plan to assemble $500 billion in financing that treats computer chips as a durable, investable asset class rather than fast-depreciating hardware. Nvidia CEO Jensen Huang told CNBC, "This is really the first time that technology chips have become an investable asset class." He described the chips as "revenue-generating assets" that are "productive," "long-lived," "fungible" and "flexible." BlackRock CEO Larry Fink drew a comparison to an earlier financial innovation: "This is the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s, and I look upon this as a next future for financial engineering."

None of this is signed yet. The arrangement consists of memorandums of understanding, and the article notes that Nvidia's last comparable MOU, a $100 billion commitment to invest in OpenAI announced a year earlier, never turned into an actual deal. The piece also argues that compute, in this context, is not the full data center stack of buildings and power supply; it is the industry's existing practice of GPU-backed loans, repackaged under a new name. Huang's own description of Nvidia's offering, a complete AI factory platform spanning computing, networking, software and a developer ecosystem, conspicuously leaves out brick-and-mortar facilities, the part of the buildout that has absorbed most financing so far. Former hedge fund manager Mark Rubinstein estimates that Blackstone alone has built a data center platform worth $185 billion including facilities under construction, and reckons the broader market for long-term ownership of stabilized data centers could grow to $1 trillion over time.

The new framing sits awkwardly next to Huang's own past statements. A year earlier, while promoting Nvidia's incoming Blackwell architecture, he told an Nvidia AI conference, "When Blackwell starts shipping in volume, you couldn't give Hoppers away." Now he points to the older A100 chip, introduced in 2020, as evidence of durability: "Customers continue to commit capacity for multi-year deployments, extending A100's economic life toward a decade." That estimate sits well outside the depreciation figures other people in the piece cite: short seller Michael Burry has put the appropriate cycle at two to three years, and IBM's Arvind Krishna has said five years. The gap is not academic. Depreciation assumptions set loan terms, and CoreWeave, which the article calls the pioneer of GPU-backed loans, can borrow less as its own chips depreciate, according to its corporate filings. A longer depreciation schedule from Huang works in borrowers' favor.

The piece connects the announcement to a smaller deal that came first: earlier this summer, Broadcom put together a $35 billion financing package with Apollo and Blackstone, backed by about a million chips as collateral, meant to boost demand for Broadcom's own chips. Nvidia's arrangement, the article argues, follows the same script. Huang points to real price increases as evidence for his case, including for the Hopper H100 chip, which came out in 2022; rental prices for older chips have been rising and, according to Silicon Data, are projected to keep rising through 2028, driven by demand for inference (running a trained model on new data) rather than for training new models. AI industry analyst Brendan Burke says the result has been a shortage of inference chips that reversed the usual pattern of falling prices; one cloud provider, per the article, nearly doubled its price on Nvidia Blackwell B200 chips for a rental customer at contract renewal. On CoreWeave's second-quarter earnings call, CEO Michael Intrator said the company is still selling GPUs built on 2020-era architecture under a contract that runs through 2029, which the article suggests may be the kind of commitment behind Huang's decade-long claim for the A100.

The article argues the arrangement reinforces an advantage Nvidia already has. Stanford's Vikrant Vig has found that most GPU-backed loans already use Nvidia chips as collateral, which makes them cheaper to finance than loans backed by competitors' chips because the collateral is more liquid, in turn making an Nvidia-based buildout easier to finance than a rival's. "In effect, they made Nvidia's product cheaper without really cutting GPU prices," Felix Wang of Hedgeye Risk Management told Bloomberg. Nvidia has also been investing directly in the neoclouds, companies such as CoreWeave, Crusoe and Lambda that rent out compute and buy its chips. On its most recent earnings call, SpaceX said it works exclusively with Nvidia chips. It later emerged that Nvidia holds a $21 billion stake in SpaceX, a customer the article says had been evaluating alternatives to Nvidia; the piece speculates the investment may have helped lock the neocloud in.

Several questions in the piece go unanswered. It is unclear whether frontier labs such as Anthropic and OpenAI, which drive much of today's compute demand, can turn that demand into profit. Nilay Patel, called in the piece only "our fearless leader," has been asking in Slack how putting a dollar into compute is supposed to return $1.01; Huang's own answer, "the return is in the usefulness of AI," does not really answer that. The article also says the actual contract terms, including whether lenders get any revenue share, are not public. One sign the asset-class framing is spreading beyond Nvidia's own deal: CME Group, a derivatives exchange, has said it plans to introduce compute futures in October, pending regulatory approval.

Key facts

  • Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR are working with Nvidia on memorandums of understanding for $500 billion in financing meant to turn compute into an investable asset class.
  • The arrangement is not signed. Nvidia's previous high-profile memorandum, a $100 billion commitment to invest in OpenAI announced a year earlier, never turned into an actual deal.
  • A year before touting chips as "revenue-generating" assets, Huang said that once Blackwell shipped in volume, "you couldn't give Hoppers away"; now he cites the 2020-era A100 chip as an example whose economic life stretches toward a decade.
  • That decade-long estimate for the A100 sits well outside other depreciation figures cited in the piece: two to three years from short seller Michael Burry and five years from IBM's Arvind Krishna.
  • Broadcom signed a comparable, smaller deal earlier this summer: a $35 billion financing package with Apollo and Blackstone, backed by about a million chips as collateral.

Why it matters

Nvidia and six major financial institutions, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, are trying to establish a new way to fund the AI buildout. Instead of treating GPUs as equipment that loses most of its value within a few years, the plan treats compute as collateral for $500 billion in financing, much the way mortgages once backed securities. If it holds, it would make it easier and cheaper for Nvidia's customers, especially the neoclouds and data center operators that resell compute, to borrow against chips they already own or plan to buy. That matters beyond Nvidia's own balance sheet, since financing capacity, not just chip supply, is increasingly one of the practical limits on how fast AI infrastructure can be built.

Who it affects

Nvidia and CEO Jensen Huang sit at the center of the story, alongside BlackRock CEO Larry Fink and the rest of the consortium: Apollo, Blackstone, Brookfield, Goldman Sachs and KKR. CoreWeave, called the pioneer of GPU-backed loans in the piece, and other neoclouds such as Crusoe and Lambda have a direct stake, since their borrowing capacity depends on how fast their Nvidia chips are assumed to depreciate. Frontier AI labs including Anthropic and OpenAI matter too: they drive much of today's compute demand, and the article treats it as an open question whether that demand will ever turn a profit for them. Broadcom, whose own $35 billion financing deal with Apollo and Blackstone came first, and cloud giants Microsoft, Amazon, Google and Meta, whose pricing power erodes as Nvidia funds independent neoclouds such as SpaceX, round out the competitive picture.

How to use it

There is no product to buy here, but for anyone tracking AI infrastructure or its financing, the practical signal is what to watch next. The $500 billion arrangement is a set of memorandums of understanding, not signed contracts, and the article points out that Nvidia's last big MOU, a $100 billion commitment to OpenAI, did not materialize; the terms that would actually matter, collateral structure, interest rates, any revenue share for lenders, are not yet public. CME Group has said it plans to launch compute futures in October, pending regulatory approval, which would be an early sign the asset-class framing is taking hold beyond this one deal. For anyone financing a GPU purchase today, the more concrete finding in the piece is that Nvidia collateral is already cheaper to borrow against than a rival's chips.

How solid is it

The core facts rest on Jensen Huang's and Larry Fink's own on-record statements to CNBC, plus reporting and quotes drawn from Bloomberg, CoreWeave's public earnings call and corporate filings, and named analysts including short seller Michael Burry, IBM's Arvind Krishna, Brendan Burke, Stanford's Vikrant Vig and former hedge fund manager Mark Rubinstein. This is a Verge opinion and analysis column, not a straight news report, so its skepticism toward Nvidia's framing is argued, not incidental. The underlying announcement is real, but the piece is explicit that it is not a done deal: only memorandums of understanding exist so far, contract details are not disclosed, and a comparable prior MOU between Nvidia and OpenAI never turned into an actual investment.

Risks and caveats

The article's central worry is a direct historical analogy: former hedge fund manager Mark Rubinstein notes that mortgage-backed securities failed when mortgages were overproduced, and the piece points to a similar oversupply risk in an AI industry that is becoming saturated with data centers, compounded by Chinese open-source models that need less compute despite being fairly powerful. Huang's claim that the A100's economic life stretches toward a decade sits well outside the other depreciation estimates cited in the piece, two to three years from Michael Burry and five years from Arvind Krishna, and it sits oddly next to his own remark a year earlier that "you couldn't give Hoppers away" once Blackwell started shipping in volume. Whether frontier labs such as Anthropic and OpenAI, which drive much of today's compute demand, can turn that demand into profit is a question the article raises but does not answer. And because the $500 billion arrangement is only a set of memorandums of understanding, with contract terms undisclosed, there is no guarantee it turns into actual financing at all.

“When Blackwell starts shipping in volume, you couldn't give Hoppers away.”

— Jensen Huang, Nvidia CEO