April – 1805

Napoleon is master of Europe

Only the British fleet stands before him

Compute is now an asset class

I see it is once again time to talk financial innovation. Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR are all working with Nvidia to put together $500 billion in financing to turn compute into an asset class.

“This is really the first time that technology chips have become an investable asset class,” Nvidia CEO Jensen Huang said to CNBC. “These are revenue-generating assets now. They’re productive, they’re long-lived, they’re fungible, they’re flexible.”

Huang said something very different about Nvidia’s own last-generation Hopper chips last year. “When Blackwell starts shipping in volume, you couldn’t give Hoppers away,” Huang told attendees at the company’s AI conference, hyping up its latest GPU architecture. “There are circumstances where Hopper is fine. Not many.” So to now be told that chips are actually “revenue-generating assets” that are “long-lived” is… quite frankly, it’s giving me whiplash.

At least for right now, Huang isn’t wrong. The price to rent old chips has been rising, and Silicon Data projects that it will continue rising through 2028. Here’s a fun anecdote: One cloud service provider nearly doubled its prices on Nvidia Blackwell B200 chips for one rental customer during its contract renewal.

“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,” said Larry Fink, CEO of BlackRock, to CNBC. Now, for some of you, this may make alarm bells go off. As former hedge fund manager Mark Rubinstein notes, mortgage-backed securities failed when mortgages were overproduced. The AI industry is becoming saturated with data centers, and Chinese open-source models require less compute despite being fairly powerful, both of which seem like potential threats to the notion of ever-growing demand for chips. There is also a far more basic question: Can frontier labs such as Anthropic and OpenAI, which are driving much of the current demand, make money?

Before we even get to the Jensen math, I want to point something out: This is not a done deal. This is some memorandums of understanding. You may remember that last year, Nvidia signed a $100 billion memorandum of understanding to invest in OpenAI. You may also remember that it, uh, didn’t happen. But the cool thing about memorandums of understanding is that you get to make a big announcement, and then it sort of doesn’t matter if the actual thing goes forward. Still, let’s assume it’s real, because even as a trial balloon, it’s telling us something interesting.

Let’s back up for a second. Why are we talking about “compute”? Well, according to Huang, “Nvidia compute is not just a chip.” That’s because there is also software, called CUDA. “That is what makes Nvidia AI factories different” from mere dumb silicon, Huang says in a tweet — er, post on X. “Their value is not fixed at installation: CUDA continuously improves their output; the installed base remains productive well beyond its initial depreciation period.”

Okay, but the chips and software alone don’t create compute — they’re only useful if they’re housed in massive data center infrastructure, which requires warehouses and power supplies. Huang appears to be discussing compute without those things, dubbing Nvidia’s system “a complete AI factory platform including accelerated computing, networking, systems software, AI frameworks and a global developer ecosystem.” Notably absent from this list: brick-and-mortar facilities.

Leave aside the risible idea of an “AI factory,” where electricity presumably toils in the silicon chip mine. Huang is downplaying data centers partially because that’s where most of the financing has gone so far. “Blackstone has built a platform valued at $185 billion including facilities under construction, and reckons the market for long-term ownership of stabilized data centers could grow to $1 trillion over time,” writes Rubinstein. Huang doesn’t care about that — a lot of it is real estate and irrelevant to him. Huang cares about people buying Nvidia chips.

So “compute” here isn’t referring to the entire data center stack; it’s a buzzword-y way of talking about our old friend, the GPU-backed loan. I can see why one might want to switch to “compute” over “GPU” because everyone knows that a GPU has a much shorter lifespan than, say, a building — estimates range from somewhere between two and five years. I suppose “compute” also covers TPU-backed loans, so there’s that.

Earlier this summer, Broadcom put together a $35 billion package that looks an awful lot like what Nvidia is offering now, signing a deal with Apollo and Blackstone to fund what we are now calling compute, with about a million chips as collateral. Apollo and Blackstone will make money on interest; Broadcom has provided a guarantee for the two senior notes issued by the special purpose vehicle where the chips live. This deal was meant to boost demand for Broadcom chips. It seems like Nvidia took note — and is doing the same thing, for the same reasons.

So now Huang is cheerleading the long life of Nvidia chips. As a “powerful example” of how compute can improve over time, Huang points to the pre-Hopper A100 chip, which it introduced in 2020, and which “remains in active commercial use,” he says. “Customers continue to commit capacity for multi-year deployments, extending A100’s economic life toward a decade.” My goodness, that’s very different from what he said last year about his flashy new chips, isn’t it!

We’ve talked about chip financing before around these parts. You may remember that no one can agree on a depreciation schedule for chips; it sort of doesn’t matter as long as Nvidia wants to bail out the companies that buy them. You can, in fact, view Huang’s statement as a sort of bailout itself. In the discussion about chip depreciation, short seller Michael Burry has suggested that two to three years is the appropriate depreciation cycle for chips. IBM’s Arvind Krishna says depreciation takes five years. And here comes Huang, saying the economic life of one of his chips is a decade! My, my, my.

This is relevant to the lenders, because it determines loan terms. For instance, the amount that CoreWeave — the pioneer of GPU-backed loans and an Nvidia client state — can borrow decreases as its chips depreciate, according to its corporate filings. So if Huang is out here in front of God and everyone saying that the depreciation schedule is 10 years, then I don’t see why banks wouldn’t believe him. That’s pretty useful for anyone trying to get loans from this consortium, I figure.

Huang cites price increases on compute — including for the Hopper H100 chip, which came out in 2022. He’s not exaggerating about the price increases, as self-serving as his logic may be. They’re driven by a higher demand for inference, which is the industry term for when a trained model analyzes new data, according Brendan Burke, an AI industry analyst. That meant the hourly rates for old chips remained high, and in some cases, even increased, Burke says. “There’s just been a major shortage of inference chips, and that’s reversed the expected trend of decreasing prices,” he told me.

On CoreWeave’s second quarter earnings call, CEO Michael Intrator said that the company has been able to sell GPUs with architecture from 2020 in a contract that extends through 2029. Connecting the dots, since CoreWeave is so tightly wound with Nvidia, I wonder if this is what Huang’s decade depreciation cycle refers to.

Will the demand surges go on forever? Fuck, I dunno. There are all these data centers being built, and it kind of seems like if compute is (or rather, chips are) as fungible as Huang says, that means data center providers are competing on price in a saturated market. But as AI gets integrated into more things, more normal companies — on top of frontier labs — will need to run inference. The pace of adoption matters — if it is too slow, this model may run into trouble.

Our fearless leader Nilay Patel has been running around with his hair on fire in Slack, asking how it is that if you put a dollar into compute, you get $1.01 back. Huang does not exactly answer this question: “The return is in the usefulness of AI,” he writes. But if my understanding of what’s going on is right, and “compute” in this context is just the old, familiar GPU-backed loan, then the return on investment is what it usually is with debt: interest.

So there’s that. We also don’t know what the contracts look like, and the details matter. (In the Broadcom contract that appears to have inspired Nvidia’s announcement, Broadcom is not backing all of the debt, just the higher-priority senior debt, for instance.) Based on previous GPU-backed loans, I’d guess that the contract from whoever is buying the compute is included among the collateral. That contract is better or worse based on who’s behind it — Microsoft will surely pay its bills, but OpenAI doesn’t make money and needs to keep raising, so its contracts are riskier for lenders. Plus, in any agreement, it’s possible that there might be a clause in there giving the debt providers some kind of revenue share or other way of sweetening the deal. What I do know, though, is that this new compute consortium seems like a pretty good deal for Nvidia.

Last year, when I talked to Stanford University’s Vikrant Vig, he noted that the majority of GPU loans were made with Nvidia chips as collateral. That, in turn, made it easier for companies to get new loans with Nvidia chips than with competitors’ GPUs — the cost of financing Nvidia GPU loans was lower because the collateral is more liquid. If the deals between Nvidia and the financiers do get finalized, that will make it even easier to get financing for Nvidia chips. If you’re starting a neocloud — that is, a small company that rents out compute such as CoreWeave, Crusoe, and Lambda — from scratch, buying Nvidia chips gives you support that you can’t necessarily get from competitors such as, idk, Broadcom.

Nvidia has been aggressive about investing in and providing financing to neoclouds in order to expand its customer base. By funding and nurturing neoclouds, Nvidia reduces the bargaining power of the big boys (e.g., Microsoft, Amazon, Google, and Meta) on price. Interestingly, on its most recent earnings call, SpaceX — the big new neocloud player — said it was working exclusively with Nvidia chips; later, we all discovered that Nvidia had a $21 billion stake in SpaceX. SpaceX was evaluating alternatives to Nvidia, but its data center buildout requires a massive increase in spending — so if Nvidia’s investment may have locked the neocloud in.

But there’s also another interesting side effect of this financing, points out Burke. Because it’s in the interests of lenders to have relative uniformity between the loans, that may further standardize the way Nvidia chips get installed in data centers. That may also give Nvidia a competitive advantage in selling chips.

It turns out that GPUs perform differently depending on how they get set up, which can make it hard to reliably project revenue for the lenders taking on the risk, Burke says. Nvidia has started putting out guidance about revenue in the ideal setting, pushing cloud computing providers to use that particular design. That would provide standardization, making lenders’ jobs easier. It also invites more scrutiny on how much customers can make and the accuracy of Nvidia’s modeling. “The forecasts I’ve seen are very bullish,” he says. In some cases, the projections are for $70 billion a year in revenue per gigawatt, which is not what anyone in the field is getting today.

So the terms of the contract may demand specific settings that increase the fungibility of data centers with Nvidia chips, both because it makes it easier to model revenue forecasts and because in the case of a default, that makes it easier for lenders to offload the collateral. “Most data center operators are highly customized and it’s going to be a major sheep herding exercise to get them to follow one approach,” Burke says. Conditions on lending may serve as sheepdogs, corralling the engineers into specific designs.

This also shores up Nvidia against competition — and not just from Google’s TPU and Amazon’s Trainium chips. Inference can be run on old Nvidia chips, sure, but it turns out CPUs can also do this work and CPUs are cheaper — like, one-fifteenth of the cost, Burke says. So if you can use CPUs, you not only can spend less to buy chips, you can also lessen the demand for GPU compute, driving that price down.

Nvidia is bringing in outside capital because it appears to be quite sore about the accusations of “circular financing,” where it’s a major investor in the neoclouds and AI labs that buy its chips. Remember CoreWeave, the neocloud propped up by Nvidia? Nvidia invested and saved its IPO and has promised to buy any extra capacity CoreWeave might have. Nvidia “agreed to spend $1.3 billion over four years to rent its own chips from CoreWeave,” making it CoreWeave’s second-largest customer in 2024.

It’s not just CoreWeave. Nvidia is widely invested in the neocloud companies. Plus, Nvidia is paying $30 billion in cloud service agreements as of its most recent quarterly filing. Jay Goldberg, a senior analyst at Seaport Research Partners, thinks these numbers represent Nvidia’s backstop agreements.

So if the new memorandums of understanding are finalized into deals, we wind up with a different situation. Instead of (say) Nvidia giving CoreWeave a dollar, against which CoreWeave borrows five dollars and then buys six dollars of Nvidia chips, Blackstone is giving CoreWeave five dollars to buy Nvidia chips.

Let’s look again at who’s in this consortium, shall we? We’ve got Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — so exactly one bank, Goldman, and a bunch of private credit companies. Private credit has been financing the AI buildout in a big way.

I am not an expert in finance but it seems to be the case that private assets make more money for the finance bros than public ones. Because AI has been on the rise, there’s been a large-scale freakout in private credit about the threat from AI to software as a service, a sector that a bunch of private credit funds had gotten into pretty deep. So now there’s a new interest in asset-backed securities, and a loan that’s backed by AI chips has the virtue of (1) not being an SaaS company’s debt and (2) potentially being the infrastructure for the thing that kills the SaaS company. That might make GPU-backed loans look attractive.

The AI buildout generally has pivoted to debt. As of the end of July, the hyperscalers “and related companies” like Nvidia had issued about $225 billion in bonds, according to S&P Global. “This segment has seen almost 10-fold growth in its bond issuance through midyear,” S&P noted, and could issue $400 billion by the end of the year.

Apollo, at least, has seen a big opportunity; The Information reported that there’s a new guy there in charge of AI infrastructure financings. Arranging AI deals is “a growing source of revenue.” About 60 people are focused on AI buildout. The development of the data centers themselves is financed differently than the chips.

I am pointing at Apollo because it published something interesting recently. In it, Torsten Slok, the firm’s chief economist, notes that the further away in the AI stack you are from the end user, the bigger your profit margin is. Models and applications lose money. That’s not a problem forever — Amazon lost money before it made money, for instance — but it adds another layer of risk to these loans. The neoclouds are the layer one up from models and applications. Should those companies be unable to figure out how to turn a profit, they are directly exposed to the risk.

AI evangelists compare AI to the internet, as a technology that has the power to totally reorganize their society. Curiously, a lot of these evangelists do not have a good model for what AI’s goals should be. There’s a lot of talk about curing all diseases and “intelligence as a utility,” but the most concrete ones look like “replace customer service agents” and “speed up coding”; maybe there’s some room for, I don’t know, risk assessment in insurance and stock trading. Are those applications enough to justify the enormous capital outlays we’ve all seen? I doubt it.

By contrast, in the early days of the internet, the technology wasn’t ready to deliver streaming music and video, but by 1999 it was obvious to a lot of people, including Larry Ellison, that’s where things were going. Similarly, in the ’90s, we weren’t culturally ready for online shopping, but it was clear to a lot of people that was an opportunity. The problem of the dot-com bubble was not that the evangelists were wrong about what the tech could do — it was that they were wrong about when the tech could do it.

So even if the most ardent AI boosters are right, getting the timing right also matters. The entire model ecosystem is currently being subsidized. It’s not yet clear that if the model makers were to charge the actual price for their services that their demand would be there. Imagine a perfect AI personal assistant, trained on every document in your organization; it costs $100,000 a day. Even if it is very good — nearly perfect! — it is way more cost-effective to hire 100 people who cost $200,000 a year.

OpenAI and SpaceX hemorrhage money. Anthropic has recently been reported to have an annualized run rate of $65 billion — but there’s no word on profit. There’s increasing pressure on the model companies to make money, and to get to a 7 percent return, below which is an “unmitigated disaster” for AI investors, the economic forecasters at Gartner project that AI companies need to cumulatively earn $7 trillion in revenue through 2029. That’s almost $2 trillion a year.

Now, if Fink is right, and compute-as-an-asset is comparable to mortgage-backed securities, we should expect a lot more companies to jump into arrangements like the one Nvidia is touting and Broadcom actually arranged. But if any of the major AI model companies suddenly go belly-up, perhaps because they cannot make a profit, the demand for compute abruptly drops. What’s more, if their contracts to rent chips are part of what secures collateral on chip-based loans, those loans are also in trouble.

One sign that Nvidia is more bullish on “compute as an asset class” than the financiers it’s signed memorandums with is its residual value support. Basically, if the neocloud bails on the loan, Nvidia has agreed to pay up to 25 percent on some of these contracts. This is perhaps meant to reassure investors, since Nvidia’s assets are on the line for investments for companies with little or no credit record. The financing suggests “that Huang believes his ‘investable asset class’ pitch much more than the market does,” says Stratechery’s Ben Thompson.

And although the price of compute has gone up, it’s irrelevant to the question of residual value, which is the resale price of the chips. If someone goes belly-up on a loan, and, e.g., Goldman has a bunch of compute to flip, who’s buying and for how much? The big boys are all building their own data centers, and in some cases using their own proprietary chips. The neoclouds have a ton of debt to service — they may not have the cash to be buyers. That’s not settled. So if the resale value falls below a certain level, Nvidia has to compensate whoever owns the debt.

We don’t have a lot of details, but what will matter here is how much the Nvidia GPUs are sufficient collateral for lending, says Goldberg. In most deals so far, the chips alone weren’t enough — lenders also needed contractual cashflow on those chips. So CoreWeave’s GPU loans are really backed by Microsoft or Nvidia or whoever. If the “AI factory” only needs chips as collateral, and not customer contracts as well, that’s significant. But notice: Jensen Huang isn’t saying that directly.

“Welcome to Jensen math,” says Goldberg in an email. “Jensen is now trying to claim that this is a new investment class - stocks, bonds, mortgages, GPUs. And his tweet is arguing that this is a special asset class because it somehow gets better over time because of software, magic and reasons.”

At a certain point, it begins to feel like this is another way for Nvidia to keep the AI party going. It’s been on a historic run, and investors’ expectations for it are high. Nvidia may have been facing limits on its previous model of endless upgrade cycles, says Leevi Saari, a fellow at the AI Now Institute. “Previously, they were like car salesmen, saying you need a new car every year because the previous generation was so inefficient you’d lose value.” Now, Huang seems to be saying that depreciation doesn’t matter — compute doesn’t lose value quickly, like a car. It loses value slowly, or even gains value, like a house.

The number of companies that can afford to keep buying new Nvidia chips every year is limited, Saari says. For Nvidia to keep beating and raising expectations on its earnings, it has to unlock more ways to fund companies buying chips. Enter the financial institutions it’s cut a deal with. If the chips don’t depreciate the way Huang said they do just last year, they’re suddenly assets that might interest, say, pension funds.

“For the life of me I can’t fathom how they square the circle of ‘you must buy a new chip every year’ with ‘don’t worry about depreciation,’” Saari says. To believe that chips will continue to appreciate in value, you have to believe there’s a totally inelastic market for chips. “That’s naked sleight of hand, in my opinion.”

According to Saari, the actual financial innovation is “how do you unlock safety-seeking capital for Nvidia’s revenue growth” and the answer is the announcement we all saw. The market’s response was muted. In The Wall Street Journal, Jack Ablin, a founding partner at the $260 billion family office Cresset, which invests in Nvidia, noted that compute, historically, is “an asset that’s had the shelf life of lettuce.” Even Stratechery’s Thompson, usually an unabashed cheerleader for tech industry pablum, noted that Nvidia’s strategy was risky.

And who bears that risk? Whoever winds up with the notes issued by these arrangements — and that’s often not the people originating them. “This is originate-to-distribute, and the destination is the general account of a life or annuity insurer,” writes Sascha Steffen, the DWS senior chair in finance at the Frankfurt School of Finance & Management and the director of the Centre for European Transformation, a research group focused on private credit.

There are a few interesting things here. Half of the group making these deals shares an owner with the entity likely to wind up with the loans. That weakens the scrutiny the final holder, probably an insurer, has on the loans. “Nvidia’s announcement is routinely described as resolving the ‘circularity’ of the company financing its own customers, and at the level of Nvidia’s balance sheet it does,” writes Steffen. “One layer down, a second circularity has been created.”

One other thing: The amount of capital required to cover insurers’ risks is tied very closely to debt ratings. If an insurer is downgraded because, let’s say, a ratings agency makes a change in methodology, that insurer may be forced to sell their GPU-backed loans, “an asset with almost no secondary market,” Steffen notes.

The question now is whether the actual arrangement will really come to pass. After all, Nvidia has made public pronouncements before — and then shied away. Just ask OpenAI about their Ohio data center.