Ep. 037 - Who's Funding the $11 Trillion AI Buildout? (Capital Markets)
- SemiAnalysis puts cumulative AI IT and data-center CapEx at $11.3 trillion for 2024-29, far above the familiar one-year hyperscaler figure. The estimate multiplies tracked shipments of GPUs, XPUs, TPUs, Trainium and complete systems by their total cost, then includes facilities often funded outside hyperscaler budgets. The old project “trinity” remains “capital, offtake, and data center,” but financing is becoming the binding leg.
- The buildout could leave roughly $7 trillion of debt outstanding by 2029, with about $1 trillion of net funding needed in the current year alone. From June through October 7, Nishball counted roughly $400 billion of AI loans and bonds either raised or announced, including $260 billion of major completed deals and another $150 billion reported but undocumented. His warning: forecasts only include confirmed capacity, so the estimate “will probably creep up.”
- Hyperscalers and NVIDIA are turning investment-grade credit into “balance sheet as a service” for labs and neoclouds that cannot finance projects independently. Microsoft can backstop Nscale capacity ultimately used by OpenAI, while managed-service treatment may keep commitments outside conventional debt accounting. Nishball’s test is simpler: “If it talks like a duck, quacks like a duck, it’s a duck”—investors and rating agencies will eventually count the economic exposure.
- GPU residual value currently bears little resemblance to textbook depreciation because scarcity has overwhelmed performance-based obsolescence. One-year H100 rentals fell near $1.30 an hour in 2025, yet some buyers later renewed around $2.50 and current deals approach $3, close to early-2023 pricing. With GB300 capacity sold out, demand “cascades down,” making immediately available H100s valuable despite GB300 delivering roughly seven times the inference speed.
- The willingness to finance AI rests on unusually strong economics throughout much of the stack. SemiAnalysis models up to $100 million of annual frontier-token revenue per megawatt against roughly $40 million to install GB300 capacity, albeit under optimistic utilization; its simplified token income statement implies about 90% frontier API gross margins. Nanos described the surprise not as AI fast takeoff, but as “revenue and profitability fast takeoff.”
- NVIDIA’s backstops can expand the market, educate lenders and generate revenue, but they cannot fund the entire buildout. Its AICP structure offers a six-year rental floor and takes 50% of revenue above that backstop; other arrangements secure NVIDIA deployments through residual-value guarantees. SemiAnalysis estimates NVIDIA could support as much as 46 gigawatts—not a forecast—versus roughly 240 gigawatts of incoming capacity.
- The investable endgame is a much broader debt-financing market able to underwrite AI projects without investment-grade guarantees. SemiAnalysis expects AI debt to become the second-largest debt-financing market after the U.S. mortgage-backed market and says 2026 is the last year it remains below 10% of all U.S. issuance. Lenders must learn GPU residual values, customer churn and site-level execution before the coming “zombie” waves move from “wave five” to “wave fifty.”
1. Financing has replaced compute and data centers as the binding constraint
Nishball’s organizing framework is the AI-project “trinity”: “capital, offtake, and data center.” Compute once looked like a fourth leg, especially as memory scarcity threatened rack shipments, but the location of the bottleneck keeps moving.
AI was heading toward roughly 70% of DRAM wafer demand, while rising memory content was pushing accelerator ASPs higher through Rubin Ultra. Instead, vendors began de-speccing systems: ROI mattered, but the immediate problem was simply that “there hasn’t been enough memory to go around to actually ship racks.”
Data-center supply has responded through newly activated powered sites, refurbished facilities and behind-the-meter work. Nishball expects supply and demand to look much more balanced by 2027-28, making facilities less of an obstacle just as funding becomes more acute.
The $11.3 trillion figure covers cumulative 2024-29 AI IT and data-center CapEx. SemiAnalysis derives it from tracked GPU, XPU, TPU and AWS Trainium shipments multiplied by full-system costs—including servers, networking and CPUs—then includes facilities built by hyperscalers, neoclouds and independent data-center companies.
2. The $11.3 trillion buildout implies a $7 trillion debt market
SemiAnalysis projects roughly $7 trillion of debt outstanding by 2029, with about $1 trillion of net new funding in the current year. Because its forward estimates recognize capacity only when confirmed, Nishball expects the total to “probably creep up.”
The existing pools are small by comparison: approximately $3 trillion of investment-grade issuance, around $400 billion of leveraged loans and roughly $400 billion of private equity. Hyperscalers and NVIDIA generally qualify as investment grade; neoclouds such as CoreWeave sit in riskier markets.
From June through October 7 alone, Nishball counted about $400 billion of AI-related loans and bonds raised or announced—roughly $260 billion of major funded deals and another $150 billion reported but not documented. That put the market broadly on track with the annual estimate.
Examples convey the breadth: Blue Sky Funding raised $24 billion plus a $6 billion delayed-draw loan for a Broadcom–Google–Anthropic TPU arrangement; a Meta–BlackRock facility raised $12.5 billion; CoreWeave added a $4.2 billion convert and $2.6 billion delayed-draw loan, while Nebius, Zencore and Anthropic raised $3.5 billion, $3 billion and $15 billion respectively.
3. Five-year contracts make residual GPU value an upside, not a necessity
A GB300 system costs roughly $31 million to $33 million per megawatt, versus perhaps $10 million-plus for the associated data center. Yet the capital side is modeled over roughly five years while a facility may last 15, supporting three IT cycles; that shifts the upfront CapEx split toward approximately 60/40 rather than the 75/25 implied by cost of ownership.
Financing usually requires a long-term offtake, typically supported by an investment-grade counterparty. “If you don’t have a five-year offtake, you’re not getting financing,” Nishball argued, making revenue and project cash flows unusually preordained compared with speculative merchant capacity.
Typical five-year deals are underwritten to roughly 20%-30% annual IRRs, at least under the economics discussed, without assuming any resale or sixth-year revenue. Whatever a GPU earns after its modeled life is therefore “like a bonus,” even though lenders still want a defensible residual-value framework.
4. Off-balance-sheet backstops turn credit quality into infrastructure
Nanos summarized $497 billion of NVIDIA commitments across AI-cloud agreements, third-party and self-use data-center leases, energy and LC guarantees, cloud-service agreements, component supply commitments and an as-yet unvalued residual-value-guarantee category. The variety makes the exposure difficult to reconstruct from headline balance-sheet figures.
In the Nscale example, capacity flows to Microsoft and then OpenAI. OpenAI’s credit alone may not support the construction loan, so Microsoft guarantees usage: “If OpenAI doesn’t pay the bill, Microsoft’s on the hook.” The project becomes financeable because lenders can underwrite Microsoft rather than the lab.
Microsoft gains capacity, may retain an economic cut and accesses outside capital without directly funding every asset. The accounting distinction matters: the speakers described an ASC 606 managed service as off balance sheet, while an ASC 842 operating lease appears on it, even when the underlying economic commitments look similar.
Nishball called the model “balance sheet as a service.” His caveat was that accounting presentation cannot make support unlimited: rating agencies and bond investors will still ask how a company could credibly backstop trillions. “No one has a crystal ball.”
5. Credit is beginning to price both sponsor exposure and execution risk
Amazon’s $25 billion, six-tranche July issue priced with an approximately 100-basis-point ten-year Z-spread and roughly 65 basis points at five years, then widened another 10-20. Nanos highlighted a 105-basis-point Oracle move; Nishball estimated Oracle’s curve had widened 30-50 since summer, with its ten-year spread near 250 versus Amazon around 100-110.
Nanos contrasted CoreWeave, whose spread was nearly 400 basis points, but Nishball separated corporate risk from project risk. A Meta-offtake project priced around 225 basis points over its benchmark—slightly under 200 basis points on a Z-spread basis—against Meta’s roughly 97-basis-point five-year curve. The difference compensates lenders for standing up and operating the cluster.
Nishball proposed ClusterMAX as an execution signal because lenders had already found it useful; a platinum operator should be “less likely to mess up.” Nanos pushed back that managed clusters and bare-metal sites are distinct products, then allowed that it could be a signal toward execution risk—but insisted each SPV, geography, team and construction schedule still requires its own assessment.
Procurement has meanwhile inverted. Buyers once compared five providers through two-week proofs of concept on 256 GPUs; now they “pick a dance partner six months before the GPUs get delivered” and may prepay 30%-50% of contract value. Firms oscillate from renting, to self-building, to exhausting their own balance sheets and effectively “mortgaging” future demand to outside providers.
6. Scarcity has broken the conventional GPU depreciation curve
Pure performance economics should crush older hardware: SemiAnalysis estimates GB300 offers about seven times H100 inference speed and three times the FLOPS. For marginal inference, Nanos asked, “Why would anybody want to buy a marginal Hopper over a marginal GB300?” Availability, however, makes that the wrong near-term question.
One-year H100 pricing fell below $1.80 and $1.40, with some weaker providers offering roughly $1.30. Sellers debated locking two years near $1.25; some later renewed around $2.50, while current one-year deals approach $3 per hour—close to first-half 2023 pricing.
Some researchers still value H100 and H200 familiarity, job throughput and numerical stability, avoiding migrations from BF16 to FP8 or FP4. More importantly, a newly funded startup chooses between waiting six months for sold-out GB300 or beginning immediately on H100. “It cascades down.”
SemiAnalysis’s performance-based curve tracked observed prices within about 2% for 18 months before dislocating around September and October, then “blew out completely” in late 2025. Reversing the model implies demand outran supply; Nanos expects that pressure to persist if construction delays, political moratoriums, endpoint usage or model quality keep demand ahead.
7. Better models and hardware are expanding margins simultaneously
Nanos argued that reasoning shifted progress away from simply increasing parameter counts toward spending more FLOPs through reinforcement learning and post-training. The Astra, Fable or Mythos generations may have grown, but not enough to outrun GB300 NVL72 gains in memory, bandwidth and FP4 performance.
His supporting example was the progression from DeepSeek V3 to DeepSeek V4, 4.1 Pro, 4.1 Flash, GLM 5.3 and Kimi K3. Sparse attention and only two-to-three-times parameter growth produced substantially better outputs without comparable FLOPS growth, allowing higher-value tokens and more throughput per GPU. He put current GB300 NVL72 gross margins for Chinese-model serving around 80%.
The speakers argued that even frontier labs did not anticipate their own profitability, user growth or ability to serve this volume, forcing premium contracts for immediate GPUs. Nanos rejected the “AI fast takeoff” label, but said the evidence looks like “a revenue and profitability fast takeoff.”
8. Frontier-token economics dwarf the cost of physical compute
SemiAnalysis models up to $100 million of frontier-token revenue per megawatt of GB300 against roughly $40 million to install it, though Nishball called the utilization assumption optimistic. Contemporary neocloud contracts were nearer $12 million per megawatt, while SpaceX and Google deals around $48 million initially looked shocking but fit the resale economics.
Its simplified income statement uses about $1.85 per million blended tokens, including cache hits, translating to approximately $38 of revenue per hour. Ownership costs roughly $2.80 per hour and neocloud rental perhaps $4-$5, leaving frontier API gross margins near 90%.
Unlimited subscription plans depend on underuse: their model assumed margins nearer 50%-60%, while heavy users and Chinese gray-market resellers can turn $20 plans into API supply. Nanos’s Planet Fitness analogy captured the incentive—“a profitable gym is the one that’s empty”—even though providers cannot reliably identify low-usage customers beforehand.
Nishball disputed the assumption that token prices are collapsing: frontier and open-model prices were rising, and flat prices would still expand margins as throughput improves. Nanos agreed blended spending is increasing, though token efficiency remains hard to measure because users spend more tokens attempting harder problems that were previously impossible.
9. AI-native demand is strong; legacy-enterprise adoption is uncertain
The speakers divided the funding question into willingness and ability. Profitability and user utility support willingness at the tip of the spear, while backstops, lender capacity and market structure determine whether that demand can actually be financed. Enterprise adoption remains less certain.
Nishball’s team assembled thousands of public debt facilities and relationships, work he estimated would otherwise require four people for four months to produce “half as good of a job.” Nanos cited flight refunds, email triage, travel and routine planning: “You’d have to pry this stuff out of my cold dead hands.”
Nishball remained least certain about traditional enterprises. Many executives display “a lack of discovery in Fortune 500 and lack of imagination,” reducing the question to summarization or “How do I fire people?” Adoption can still hinge on one leader saying embrace AI—or banning it.
Nanos’s harsher answer was that slow incumbents may simply be replaced. He argued that nine of the ten largest companies by market capitalization are AI-, semiconductor- or data-center-related and that successful AI-native companies could grow independently or transform their acquirers. His conditional proof point: an Anthropic IPO at $2 trillion would make it roughly a top-25 company.
10. AI debt is becoming a macro asset class before lenders understand it
In SemiAnalysis’s comparison, AI debt crosses auto loans, student loans, credit cards, non-agency mortgages, CMBS and HELOCs, becoming second only to the U.S. mortgage-backed market. Unlike those slowly growing categories, its line rises abruptly as hyperscalers, neoclouds, labs and data centers all arrive together.
Nishball expects AI financing to exceed $1 trillion and says 2026 is the final year it remains below 10% of roughly $12.7 trillion in total annual U.S. issuance. He linked that scale to higher long-end rates and wider credit spreads: lenders face a “zombie apocalypse” without knowing whether this is “wave five or wave fifty.”
Today, some lenders avoid learning the technology by underwriting NVIDIA, Microsoft or other investment-grade credit. Backstops buy time to study GPU rental prices, residual values, churn, cancellations and execution. Nanos’s Ontario Teachers’ Pension Plan anecdote made the endpoint tangible: pension funds ultimately supply capital, so explaining the risks is substantive work rather than financial ornament.
11. NVIDIA’s backstops are a bridge, not enough capital for the destination
Under AICP, NVIDIA promises to rent unused neocloud capacity for six years at an average floor near $3.25, allowing lenders to price NVIDIA credit plus execution risk. The guarantee is not intended to be exercised; it lets operators finance clusters while building more diversified customer books, including token-as-a-service activity.
NVIDIA also receives 50% of revenue above the backstop, making the guarantee a business rather than charity; Lambda uses a related revenue-share structure. Elsewhere, $105 billion of residual-value guarantees support 4.25 gigawatts—and potentially nearly nine across two phases—while locking the complex into NVIDIA GPUs. Nishball called that function a “land grab.”
For the proposed $500 billion capital partnership, SemiAnalysis illustrated a loss waterfall across equity, an unguaranteed B piece and successively senior A tranches. NVIDIA’s 25% residual-value guarantee could be attached to a selected tranche. Losses would have to exhaust junior capital before reaching the safest tranche, creating products for private equity, private credit, banks and insurers.
SemiAnalysis believes NVIDIA’s balance sheet could support as much as 46 gigawatts, explicitly “not our forecast,” against roughly 240 gigawatts of incoming capacity. With roughly 60-some-odd incremental gigawatts potentially arriving in 2029, the lending market ultimately must underwrite projects without backstops.
Full transcript
For our first ever advertisement, we're going to be doing a lot more of this on future SemiAnalysis Weekly episodes. SemiAnalysis is a reader-funded publication. Over to Dan. What are we advertising this week?
Yeah. Well, we noticed that 90% of viewers aren't subscribed, so if you could hit the like button and subscribe, and be sure to hit the alerts. Seriously, though, the real ad: we're hiring for a new group called Compute Capital Markets. We're going to look at loan demand, how the financing markets are evolving for financing GPUs, and the tools that lenders and everyone in the market need to evaluate how GPU finance is going and how the GPU industry is going: rental pricing, residual value, and all of those things that come with it. So we're looking for credit analysts. We're going to hire 4 slots. We're looking for Singapore and New York City. We're looking for a junior credit analyst, about 1 to 3 years of experience in the credit space, investment banking in Singapore or the U.S. We're also looking for a credit specialist. That could be someone who's been a little more experienced, probably 4 or 5 years in credit trading, credit sales, or credit investing. IG, high yield would be great. And also a credit analyst, anyone who's been a desk analyst, buy side, sell side. We're looking for people who are credit-minded. I think that if you enjoy this episode that's going to come, and all these structures and all these Z-spreads and T-spreads, I think you're going to really like this job. So click the Apply button right here.
All right. Roll the credits. Hello, everyone. Welcome back to SemiAnalysis Weekly. In this episode, we have Dan. We're going to talk about everything to do with capital markets and how we're funding the build-out right now. Dan has been digging into everything credit, capital markets, and compute. Specifically, we put out an article called “NVIDIA Backstop Universe: Heads I Win, Tails You Lose,” where we talked about the 11 trillion AI build-out, NVIDIA's backstop economics, and the limits of NVIDIA's balance sheet.
At the pace things are moving right now, that's already a month old. Without further ado, Dan, why don't you lead us in talking a little bit about NVIDIA's backstop universe, how you came to set up this article, and why you dug into this expanded view of NVIDIA's balance sheet?
1. NVIDIA Backstops The Buildout
Let me show you a bit about NVIDIA's backstop universe. Indeed, it was just a month ago, but things are moving so fast. At this point, funding announced is around 100 billion each month. This is how we looked at the balance sheet.
The whole premise is that NVIDIA has been giving a lot of backstops, and the question is why. There hasn't been a lot of visibility on how much issuance is going to come in or how much funding is needed. So we defined that for the first time. What we did is look at the funding needs and talk about whether NVIDIA's balance sheet would be strong enough to support backstopping a lot of these funding needs. We'll talk about that much more later, but I think it's worth quickly going through the constraints and what led us to this point.
Let me go back to this. If we think about the constraints, right, Jordan? You asked me earlier about the trinity. You remember that?
Yeah.
We used to talk a lot about the trinity. The AI project trinity is basically capital, offtake, and data centers. Earlier, Jordan was asking me, “What about compute? Isn't that the fourth leg?” That's a good point, because compute used to be a pretty big shortfall—a pretty big pinch point—and it used to be a worsening pinch point.
What's happened since then? Memory was one of the biggest pinch points. So let's talk about constraints. The constraints have been moving from compute constraints to data center constraints, and now financing is a new constraint.
Why are those more acute? What's happened with memory? There has been a shortage. You can see that AI's percentage of DRAM wafers was going to be 70%. On the logic side, accelerators are an ever-higher share. But what happened recently is that there has been de-speccing.
Let me show you what happened. From an accelerator perspective, this used to be the pattern: GPU ASP had been going up because of memory, and it was set to go up on Rubin Ultra. But that pattern stopped in its tracks. What we got instead was de-speccing.
Why? For 2 reasons. Let me show you a quick bridge chart. What we saw was that the cost was increasing for the Vera NVL72, and what they did to offset that was decrease SOCAMM. But it's not just about price; it's about availability. The issue is that there simply hasn't been enough memory to go around to actually ship racks. So the de-speccing was necessary to deconstrain that.
That's the chip constraint. Let's move on to the data center constraint. That was one of the first constraints people worried about, but we've found a lot of creative solutions around it. We've mustered a lot of capital to build data centers, activated more powered sites, refurbished old sites, and done a lot of behind-the-meter work. By 2027 or 2028, it's much more balanced, and this is no longer an obstacle.
So what are the new constraints? We talk a lot about funding now, and that's a constraint people have learned about in just the last few months. I'll show you how we calculate the overall constraint. Let's take a look. This is the cumulative total CapEx: 11 trillion in CapEx that needs to be spent by 2029.
Yeah, Dan, explain this chart, just because I think a lot of people are throwing around the term 1 trillion in hyperscaler CapEx. How do you get to 11.3 trillion on this chart?
11.3 trillion is cumulative CapEx from 2024 through 2029, and it's escalating each year. We did this because we track all the GPUs, all shipments, the total cost of ownership of the racks, and the bill of materials. That's not just across GPUs; it's also XPUs, TPUs, and AWS Trainium.
If you take the sum-product of the shipments across the total cost—which includes the server, networking, CPUs, and everything—you actually get to 11.3 trillion. By the way, you also mentioned hyperscaler CapEx, so that's actually a subset of this.
If you look at the legend, this is AI IT CapEx and data center CapEx. A lot of that CapEx is not incurred by the hyperscalers themselves; it's incurred by independent data center companies. That's part of why 1 trillion in a year of hyperscaler CapEx becomes 11 trillion, because it's hyperscalers and neoclouds, and then it's IT and data center.
This doesn't necessarily cover other things in the supply chain, which we do cover and which Dylan has talked about—power plants, industrial equipment, and stuff in the fabs that's a precursor to AI IT in this measurement, like cloud equipment or just a new fab. There's a lot of stuff going into the build-out right now.
You said that there's a constraint related to funding. Talk me through how people are actually funding this build-out and where everybody is finding the capital.
2. The Debt Funding Gap
It's a great question. What we said is that there's going to be 7 trillion of debt outstanding by 2029. What does that work out to?
What I did here is put the net additions, the net new funding. It works out to 1 trillion this year. It's a little flat because whenever we forecast stuff forward, we tend not to add capacity and deliver until it's confirmed. But there will still be some debt that rolls off because it's all amortizing. I would expect this estimate will probably creep up.
The blue triangles are our estimates, and if you add this up, that gets you to 7 trillion. But Jordan, where are they getting the money from? To answer that question, let's look at how much money is put to work in this space.
In the investment-grade market, it's about 3 trillion of issuance. That's the red box here. Investment grade is anyone with a credit rating of BBB− or higher. This is typically going to be hyperscalers.
SpaceX is also investment grade. Typically, great companies like NVIDIA are investment grade, while high-yield companies are a little riskier. CoreWeave is technically high yield, as are various other companies with a lot of debt or a more nascent business model.
That high-yield issuance is denoted down here. It's about $400 billion in leveraged loans, which are loans of a riskier nature and are often used for acquisitions. Private equity is also $400 billion. By the way, this is investment-grade bonds, this is high-yield bonds, and then your loans are separate.
Yeah. So everything's growing. Both high yield and investment grade are growing. Take me through again where some of the individual projects are that are responsible for this much debt being raised.
3. The Deals Financing AI
Absolutely. I can give you a few examples. Just since June alone—we're saying $1 trillion this year, and probably $500 billion to $600 billion, or maybe around $500 billion, was raised through June, so we're pretty much on track. But from June 1 through October 7, in Asia at least, there has been about $400 billion of AI-related capital—loans and bonds—that was either raised or announced.
There have been about $260 billion worth of major deals that have been raised, and another $150 billion that have been reported but not yet documented. Some large examples include Blue Sky Funding, which is a $24 billion deal with a $6 billion delayed-draw term loan as well. This is part of the Broadcom, Google, and Anthropic TPU deal.
Yeah.
Soporia is another. It's a Meta–BlackRock joint venture for an El Paso facility, with a residual-value guarantee for Meta. It's $12.5 billion. Hut 8 raised $4.25 billion of paper.
Yeah.
CoreWeave, like all the neoclouds, has been at it. The ones I just mentioned are all hyperscale, so now we're on to neoclouds. CoreWeave raised a $4.2 billion convertible note. It raised a $2.6 billion delayed-draw term loan in August. Nebius just raised a $3.5 billion convertible note. Zencore in Asia did a $3 billion term loan, one of the first in Asia. Anthropic did a $15 billion pre-IPO revolver.
I could go on, and I haven't even gotten to the hyperscaler bonds, right? Does that help, Jordan? I should talk about hyperscaler bonds, because that's what kicked off this whole firestorm.
Yeah.
I could go through this. I have pages and pages of it.
Well, that's a nice little teaser or taste. I'd love to talk about some of the specifics, both in terms of the actual projects and what they're backstopping, or what they're raising debt for. This is not strictly for GPUs.
Yeah.
I think a lot of people might get confused by this and think that this is strictly people raising money to buy NVIDIA GPUs. But there's so much more that goes into actually building a data center that can accept the GPUs. It seems that NVIDIA, as well as all the hyperscalers, are motivated to backstop people across the whole industry.
4. Data Center Economics Matter
Yeah. It's the data center, too. Even for those who are used to looking at total cost of ownership, let me show you one of the total-cost-of-ownership charts here.
This is the full stack, and what you'll notice is that we always say most of the total cost of ownership is the capital cost, versus the data center cost. If you look here, the operating cost per hour is $0.66, and the capital cost is $2.13 out of $2.80. If I do $2.13 out of $2.80, that's around 76% OpEx.
You might think the data center is not a large part. CapEx for a GB300 is around $31 million to $33 million per megawatt, and the data center is about $10 million or more. If it's $30 million and $10 million, that sounds mismatched. The trick is that the capital side is a shorter-term project: it has a 5-year depreciation period, whereas your data center has a 15-year life.
On a cost-of-ownership basis, it's roughly 75/25, but on a CapEx basis, it's more like 60/40. The data center is even more significant because we're building more upfront, and this data center is going to last 15 years. It's going to support 3 projects' worth of AI IT CapEx. Does that help put into context how much is needed for data centers?
Yeah.
How much of the need is for data centers?
100%, yeah. I think a lot of people don't really understand the breakdown. The chips are expensive, and they make up three-quarters of the capital cost of the project, depending on how you model them. I think that's perhaps the critical question that's outstanding: the residual value of some of these GPUs that are coming off their 5-year life cycle, when people are expecting the warranty to have expired.
Do they keep running them for years 6, 7, and 8? How do you view the cash flows for GPUs after the residual-value period, or after the modeled lifespan has expired, in terms of how that enables them to fund future projects?
Yeah. That's a question people ask a lot. We've gone through this many times, but to directly answer your question, many people doing these projects actually underwrite them to make money over a 5- to 6-year period. They're not banking on the residual value at all.
This gets into the lay of the land now, because most of the clusters are built on long-term offtakes. That's the only way you can get funding, right? Remember the trinity: you need offtake, capital, and data centers. If you don't have a 5-year offtake, you're not getting financing, and that's one of the biggest constraints.
Because of that constraint, because you don't get financing without a long-term contract, everything is typically preordained. On a 5-year deal, you're typically going to get an IRR of about 20% to 30%, at least. It's been higher in the past.
What that means is that by the time your project is over in 5 years, you'll have made a 25% yield per year on your capital investment, without banking on residual value. Most things are underwritten without regard to what happens afterward. It's a bonus.
Yeah. Can we back up a second and talk a little bit about the broader industry? We keep saying “data center,” and we talk about GPUs. But when it comes to analyzing what is being backstopped by NVIDIA, for example, I've got this awesome chart on screen now: NVIDIA's off-balance-sheet commitments.
You summarize the $497 billion of off-balance-sheet backstop commitments that they've made at this point, and you go through 7 categories here. There are AI cloud agreements, where they're signing deals directly with neoclouds; data center leases, where they're signing deals with providers that are building data centers; and guarantees and LC guarantees, again, with more on the energy side.
There are cloud services agreements, which are not necessarily the same as a neocloud agreement in this measurement—basically an accounting difference there.
Take 2 minutes. I'm not—
There are data center leases for self-use, which is different from the previous category.
There are supply and capacity commitments, primarily around component suppliers like memory. Finally, there's the residual-value guarantee, where we haven't assigned any value to it, but it's a category of something that they're guaranteeing for people.
Give me a lay of the land on all of this. It's such a mess. There's so much going on.
Yeah, I'm happy to do that. We can go through each of them. I've got some nice diagrams that can tell you why each exists, how we modeled it, and what the point is.
The good thing is that this is going to be on YouTube, because you're going to want to slow this down to 0.75x when I go through it, for all you listeners. But before I do that, let me talk about the “why.” A lot of people skip the why and go straight to the what. Then they look at the what and say, “Okay, this is circular financing.”
Now let's talk about the why.
Why are we going to hit an obstacle? Why is it a constraint? Other than rates going up—which the market is telling you is going to be tough—and the numbers I just showed you, there’s something more.
Like I said, Jordan, most of the offtakes are 5-year, and most of them have a hyperscaler backstop. What that means is, let’s say you’ve got a Microsoft–Nscale–OpenAI deal. Nscale is selling capacity to Microsoft, which is then selling it to OpenAI. Microsoft is effectively backstopping that sale because it’s guaranteeing that the capacity will be used, right? Some hyperscalers also do that directly, like Meta is directly offtaking from NeoPods, for instance. There are many deals like that.
Yeah. Nscale needs a guarantee to go out and take out a construction loan to buy—
Exactly.
—GPUs, plug them in, and turn them on.
Exactly.
And the guarantee is not good if it comes from OpenAI because they don’t have an investment-grade credit rating. So instead of relying on OpenAI to pay the bill, they’re going to say, “If OpenAI doesn’t pay the bill, Microsoft is on the hook.”
Exactly.
So why does Microsoft sign up for this, Dan?
Well, it’s a good question. I think there’s internal and external use, right? I think they also take a cut of it. It allows them to expand and access capital without directly using their balance sheet, because it’ll show up as upcoming work, right? A lot of it shows up as services of a contract, which is under a certain accounting rule. I forgot—which one is it, Jordan? ASC 606?
606, not 842, yeah.
Okay. ASC 606 is the—yeah, I was about to say ASC 404 if I didn’t remember that correctly.
There’s a number of different designations—the Accounting Standards Codification or something. Anyway, the distinction is very material to these deals because the definition of a managed service is different from the definition of an operating lease. A managed service does not end up on your balance sheet; an operating lease does.
Yep.
The technical treatment of who’s responsible for what in the data center is different from the accounting treatment. Who can log into what, enter what room in the site, or touch what is less relevant than how the accountants want to track where these assets sit and how they depreciate, right?
Yeah. I think, Jordan, it’s kind of like: if it talks like a duck and quacks like a duck, it’s a duck. I don’t think ratings fool anyone. No one’s stupid, right? But it takes a little more work because they don’t do it for you. They don’t say, “Oh,” and add up all this. NVIDIA discloses its obligations, and I think in some of the filings they will disclose them, but they’re not on the balance sheet, right? You have to go looking for them, right?
That’s part of what we’re going to do at Compute Capital Markets, or CCM. It’s something that’s going to be hard for them to walk away from. But my point is, even if it’s not debt, that is not unlimited, right? There’s going to be a point where rating agencies and investors say, “Hey, this is just too much for you to backstop.” How are you going to backstop $3 trillion? No one has a crystal ball, right?
Yeah. You specifically used the term “balance sheet as a service.” So right now, it seems that all of the hyperscalers, when we talk about this $1 trillion of CapEx, are effectively putting their balance sheet to work as a service. That has limits, right?
Mm-hmm.
Yeah.
Well, yeah. If you look at Amazon, for instance, they did a $25 billion deal across 6 tranches. This was back in July, right? It priced well wide of where their benchmarks were, right? This was kind of the canary in the coal mine. People were definitely aware, but you are seeing the hyperscaler spreads blow up by quite a bit, and the CDS spreads blow up by quite a bit.
I’ll actually check for you right now. We can maybe edit this a little bit, but I want to say a number, if that’s okay. You guys don’t mind? Give me a second. Model directory. I can fire up Bloomberg as well. Oops.
Take your time, dude. No rush. We can just clip this section.
Yeah. Okay. So Amazon—if I look at July, I’m just looking at July, right? Their 10-year Z-spread—
I’m sorry, restart, dude. I’m only seeing the financing-market obstacles.
Well, I’m not sharing it. I’m not sharing it.
Oh, I saw your cursor moving like you were sharing something.
I know. No, it’s because it’s overlapping the deck.
Great. Let’s do that again.
I’m moving it to another thing.
Sorry, I interrupted you.
Sure, sure. Yeah. If I look at Amazon, for instance, remember when they issued that bond? Their 10-year Z-spread would have been around 100 basis points. So roughly, if Treasuries used to be at 4% or 4.5%, then you add 100 basis points, or 1%, which means the yield on Amazon debt is about 5.5% or 5.75%, as the case may be. For 5-year debt, that spread was roughly 65 basis points.
What happened right after was that it blew out by about 10–20 basis points from that issue, which is quite a move for investment grade. Things are relatively under control for now, but you’ve definitely seen Oracle move up quite a bit since the summer. Their entire curve probably moved up 30–50 basis points. You are seeing the debt market take note, and rating agencies will see through this stuff.
Yeah. Walk me through this Oracle increase of 105 basis points, and then compare that to CoreWeave, which is well above investment grade at almost 400 basis points higher.
Yeah. Oracle’s spread is currently around 250 basis points on a 10-year Z-spread, versus Amazon, which I’ve got at only 100–110 basis points for 10 years.
This chart is a little bit about relative credit risk—how do we compare credit risk? What this is illustrating is different types of structures. Let’s look at the structure where CoreWeave is selling an offtake contract to Meta. How much am I going to charge to lend to CoreWeave for that?
Someone who’s lending money to CoreWeave will say, “Actually, this is Meta credit risk plus execution.” The Meta curve, at the time we did this, was around a 97-basis-point Z-spread for 5 years. The actual lending to CoreWeave for this project was around 200, right? It was 225, so plus 225, which, when you do a little math, comes in slightly under a Z-spread of 200.
The difference between those 97 basis points and the 200 is the execution risk. Even though Meta is there, offtaking—or Microsoft is there—you still have to stand up the cluster. Which is why we should actually rename this the ClusterMAX risk, right? There’s going to come a day, Jordan, when it’s like, “This is your S&P rating, and this is your ClusterMAX rating,” right?
It’s not, man. Everybody wants to make stock picks from ClusterMAX. Now you want to rate people’s debt from ClusterMAX? It’s one data point.
Don’t tell him—he’s going to shut us down.
Oh, God, man. No, no, no. We’re not branding neocloud debt-product pricing as ClusterMAX, Dan.
Sorry, audience. Jordan does not give investment advice.
There are a lot of things that go into building a high-quality cluster. I think execution risk in standing up a given data center on a given timeline and then maintaining a contract so that there isn’t a cancellation or termination of the contract is typically associated with bare metal, which is a very different product from the managed clusters that we’re rating.
Actually, I’m thinking this through, and I’m not completely convinced by what I’m saying. A lot of the experience that applies to building a great bare-metal service certainly applies to the quality of the cluster. People who build reliable clusters are generally able to maintain and deliver on their contracts.
So, all right, fair enough. We can use this as a rating of something toward execution risk.
Well, Jordan, you say that, but I've been on the roadshow with lenders. I did New York and Singapore, and a lot of them have said, “Yeah, this has been incredibly helpful to their process.” So I think—look, if someone's made it onto ClusterMAX and got to platinum, they're less likely to mess up on execution generally. It's just a stamp of approval. Just because you got good grades doesn't mean you're a genius, but you're probably more likely than not, right?
I think it's a—
You know?
A signal and a data point. But—
Yes.
Managed clusters and large bare-metal data centers—
Sure.
—are very different products. I think just because somebody's in platinum does not necessarily mean that all their data centers are going to be on time, and vice versa. Just because somebody is down the list, or maybe ranked as underperforming or something like that, does not mean that their data centers will be late or that they will be—
Sure.
—low quality in the future.
Sure.
If anything, I would ask people to assess a given project or a given site. We've seen people structure every single site within a neocloud's portfolio as an SPV, and I think that's the right way to think about it. There is a different—in some cases, very different—level of execution risk for a site in one location on one continent that happens to be owned and operated by a given neocloud, even though they can go to a completely different continent for a different project the next week or the next month, raise debt in a totally different way, and have a totally different team working on it. Therefore, things are a little bit different. Dan, let's move on to—
Well, Jordan, that's an interesting point, because a lot of the clusters—a lot of the neoclouds we talk to that are unavailable, they're not unavailable. It's because they're busy doing bare metal. To your point, it's not their focus. They've landed large deals, so it's not like it's just one data point. Some people who are doing great and have landed lots of deals are just like, “Hey, sorry, guys, we just don't have the product for you to test right now.”
5. Backstops Reshape The Market
And the whole market has changed. We're constantly going back to this, but when ClusterMAX started, buyers had 5 different providers offering them 2-week POCs on 256 GPUs so they could make their decision in an informed manner and sign up. Now you have to pick a dance partner 6 months before the GPUs get delivered, and you have to make a prepayment to these people that's worth 30% to 50% of your contract value in some cases.
You said this earlier on this episode: effectively, the buyers are just paying for people's cost of capital directly. You're taking money that you've raised, which has a cost of capital, and then you're giving it to somebody else, and you have to pay for their cost of capital. It's stacking on top, which is driving a lot of people to try to do self-build work.
It's almost like this oscillation. You start out early and you do rental because you don't want it on your balance sheet, you want to move fast, and you want the biggest cluster possible. Then you start to grow and realize, “Hey, this is a bad deal. I'm getting ripped off. I should own this stuff myself.” Then you start doing these self-build sites, and you reach this limit of what your balance sheet can handle and how much money you can raise.
Then the frontier labs go, “Hey, guys, I need help. I need to basically mortgage my future right now, and I need you to build some stuff. I need to guarantee you some returns on this work because I don't have the balance sheet to do it, and I need to put your balance sheet at risk.” So it's an interesting curve where the frontier labs, the way they're doing deals right now and trying to get everything off their balance sheet—and even the hyperscalers are trying to get stuff off their balance sheet—it kind of looks like an early-stage startup, right? They want the most that they can get right now.
I think you've touched on the why, right? That's where we kind of got here. Why is there all this backstop? You described the problems: you have to book 6 months in advance, you can't get the compute you want, and you have to put 50% down. That's kind of why they've done the backstop.
Let me share. I did up this slide that describes some of the why. Broadening compute availability is what you just talked about. Supporting the financing market is another thing, because we just talked about how, unless you have that IG backstop, you can't do anything.
There's no incentive as a business model to say, “Hey, I'm going to fill my book with half of various inference services,” because your lenders will go, “Hey, I'm not going to—if it's not IG credit, I don't understand this business well enough.” That goes back to this earlier slide that I had on the obstacles. They're still on the learning curve. They don't understand the business well enough to underwrite anything other than IG, and then they don't have the tools.
A lot of what Jordan and I are working on is Jordan working on the tools to assess the operational quality, and myself on assessing what's happening in the market. That's why we have the GPU index. That's why we're trying to create tools that lenders need to monitor risk.
You mentioned residual value. This is one of the huge asks we had on our roadshow, because banks think in terms of residual value. They're lending against assets. They're used to lending against property and cars, and they have to know residual value, which is very, very hard with GPUs. It's not like you straight-line depreciate them. I think the residual value is what you can earn on tokens, or what you can reasonably rent it at.
6. Demand Rewrites GPU Depreciation
And it's so distorted right now because I think in a normal market where you have this contention between supply and demand, where people produce enough GPUs based on the demand for them and they can always go a little bit higher, then you have real pricing and real settlement. But what's happening right now is nobody has an understanding of what the true theoretical H100 depreciation curve looks like as GB300 comes online, as Vera Rubin comes online next year, and as it outperforms that.
How can I run an H100 profitably? We already have models with InferenceX that show that you really can't. If a GB300 is outperforming your Hoppers by 7 times, why would anybody want to buy a marginal Hopper over a marginal GB300? You'd never want it. You'd want the latest and greatest.
But this isn't the question people are asking about Hopper. Instead, the question is, “I have no option to purchase a B300 GPU because it's all sold out, or I have to wait 6 months. I want to do research right now. What option do I have to buy?” And the answer is, “Renew my H100 contract with 4-year-old GPUs for double the price of what I signed it for previously.”
This is on screen. Dan's got our pricing chart for the 1-year term-length H100 deals that we know about. They're approaching the prices at which they were signed in the first half of 2023 today: 3 bucks an hour for H100 GPUs on a 1-year term.
If you believe in your startup being successful, you have to imagine that somebody's going to do something OpenAI and Anthropic can't, and it's going to be you. So, flippantly, you have to believe in AGI here, right? And if you don't—
Jordan, I had a question, because you said that no one doing inference would ever rent an H100. Why is the price going up, then?
Well, because I think there's so much demand for it that the price goes up strictly because it's the only option for a lot of people. Second of all, some researchers prefer H100 because it's familiar. They're more productive by the metric of jobs per day rather than tokens per second.
I don't need to spend time debugging. I don't need to port stuff from BF16 to FP8 or FP4. I'm doing training, and I need numerical stability. I don't care about these low-precision data types that are used for inference. I'm doing some new model where, again, I care about numerical stability. I don't want this risk.
So there are reasons why people want to do research with Hoppers—H100 and H200.
But I think the biggest reason why prices are increasing today is simply because GB300 sold out, so people go to the B200 and B300. Those are sold out.
It cascades down.
Exactly. If your answer is to wait 6 months and have nothing, or get started with H100s today, a startup that just raised money has a pretty easy decision. They go for the H100s they can get access to right now.
Yeah, that makes so much sense. That's why, when we do this index, we do it survey-based, because there's a story behind every data point, right? This is not just scraping a bunch. There is a story, and those matter. And you know what I did, Jordan? I actually decomposed—do you want to go, or can I talk about this?
Yeah, sorry. Well, just one last thing. If you go back to our previous one—
Yeah.
We would have had a more academic, theoretical way to forecast the depreciation in the H100 pricing curve. If you look at this chart, sometime around September or October of last year, it just kept going down below $1.80, below $1.40.
Mm-hmm.
I remember some deals being signed at that time for $1.30 for an H100. People were offering stuff that low.
Yeah, that was the bottom. But those were not great providers.
For sellers.
Well, no, I mean—
As a buyer.
The guys who were selling at $1.30 were not great providers, right?
That's fair enough. Yeah, exactly.
Yeah.
But the point is, I know some sellers at that time who were evaluating $1.30 and asking, “Should I do a 1-year contract at $1.30, or should I do a 2-year contract at $1.25?” Something like that. Those guys are kicking themselves for not doing a 2-year contract then, because they just renewed at literally $2.50 in some cases—way, way more than they were paying this time last year.
The point is, we would have a real understanding of H100 residual value if we had seen the pricing curve depreciate like some other asset, like a car or something. But 4-year-old H100s going into next year are likely going to be selling for higher prices than they are today if the trend continues. I expect it to, because I expect demand to continue to rise.
Let me make 2 more cases for this that are not based on an AGI future. You don't need to believe in that. You need to believe in data center delays or cancellations due to political posturing, moratoriums, and execution risk, which we see a lot of. You also have to believe in more demand from consumers on Anthropic and OpenAI's API endpoints, which I think you can believe in even if you don't believe the models are getting better.
You just need to look at things like Instinct growing or Muse from Meta and imagine a world where more people use AI-driven products, even if the models stop getting better. The combination of data centers taking longer to build, politicians getting in the way and stopping them, more demand for the existing models, and the models getting better are all a recipe for increased prices over time. That's generally the case I've been making to a lot of people.
Yeah. The thing is, if you did the scientific version, we started this in 2024. We actually had a forecast, and that looked downward-sloping. It started detaching. We were within 2% for a good year and a half, or a good 18 months, but then it started dislocating in September.
The way we calculate the theoretical depreciation curve is based on the relative power, the relative compute power, of new GPUs. For instance, a GB300 had 7 times more inference speed and 3 times more FLOPS, right? That new capacity would push down the price of your old capacity. In theory, that's what should happen.
But what we always said is that we understand supply and track price, but it is hard to track demand in real time. So what we do is flip it on its head and say, “Let's observe price and let it derive demand.” In this slide, that's exactly what I did.
If you have a balanced, neutral supply-demand environment, what you see is the value depreciating, or the rental price going down, in a scientific way—in line with new capacity. But what happened in October is that it dislocated. If your demand is higher than supply, then your price surprises to the upside.
So I reversed it and said, “What does this imply for the demand-supply balance based on price?” What we saw is that it blew out completely in late 2025 and has never looked back. Does this tie in with what you've seen, Jordan? I assume so. Maybe not the theoretical stuff, but what you observed in the market.
Well, a couple of things happened around that time last year. I think models basically stopped getting bigger. It took a while for the Chinese models to catch up, but up until around this time last year, when reasoning models were taking off, people discovered that we can trade FLOPs for parameter count and continue to improve the quality of the model with a lot of innovations around RL and post-training.
They've just continued to exploit that. We may have gotten bigger models. I think we likely did get bigger models with the Astra and Fable generation, or the Mythos generation. But we didn't get significantly bigger models that outpaced the improvement in hardware that came when the GB300 NVL72 came online.
Significantly more memory, memory bandwidth, and FP4 performance were able to keep up with the increased size of the models. What that meant is that these companies could continue increasing the price per token on their APIs without necessarily increasing their costs to serve the model significantly. The result is that they're way more profitable—way, way more profitable.
You don't even need to know anything about the Anthropic or OpenAI models to make this case. You just need to look at the Chinese models, and you need to look at the performance of DeepSeek V3 on an H100, H200, or even a GB200 when those were being brought online last year. Then you need to compare it to DeepSeek V4, 4.1 Pro, 4.1 Flash, GLM 5.3, or Kimi K3.
You look at all the innovations in sparse attention. You look at the parameter count growing a bit—around 2 times, or 3 times in some cases—but FLOPs not increasing significantly and performance increasing so much. They can command a pricing premium over the previous tokens because the tokens are more valuable, but they're cheaper to serve. You get more throughput per GPU with these.
People serving Chinese models are enjoying gross margins of 80% right now on a GB300 NVL72. It's obvious that there's going to be more demand for this compute, this asset.
The question, or what gets confusing to me here in some cases—and maybe you have an answer for this—is how nobody forecasted this. We did our best. We understood it, right? I think even the frontier labs that really deeply understand how these models work did not expect themselves to be this profitable, to be able to serve this many users, or to grow this fast.
Even they were caught by surprise and had to do crazy things and sign big contracts at very high premiums just to get access to GPUs tomorrow, serve their users, and go for the bigger research jobs. I don't view this as an AI fast takeoff, but it seems to be a revenue and profitability fast takeoff for a lot of these labs.
Yeah. I think we've really hit on what the debate is going forward, and it's the willingness and the ability. I think the ability is the profitability. A lot of people who are new to this don't realize just how profitable it is. That's one of the things that will determine whether we overcome this obstacle, this funding wall—
Well, I think it's also confusing which layers of the cake are profitable, which is to say—
Right. Yeah.
Nvidia—
We looked at that, yeah.
Nvidia clearly had durable profit margins for a long, long time. But—
Mm-hmm.
Your layer-cake chart is great for explaining that there has clearly been an increase in the profitability of the neoclouds, because—
Mm-hmm.
The amount that they can rent the GPUs for to people has increased recently, while nothing has changed in the operational costs of the GPUs. People are selling tokens with more efficient models or at higher prices, and therefore their profitability is increasing. My personal experience, and I think a lot of other people's experience, is that the value or utility you're able to derive from using these AI applications is also increasing over time as they get higher quality.
And so I don’t look at this and say, “We’re taking from one to give to another.” It seems like every layer of the stack, in some ways, is increasing profitability, except for the server OEMs, where it’s still a pricing war and they can’t get 9% margins.
Well—
—9% margins.
No, no, no. There’s still a bit for everyone, but, you know—
The trick will be the tin box and the people to plug it in.
Well, I can get into that, but I want to make sure everyone stays—
Yeah, sure. Yeah.
Organized, right? I think, again, everything Jordan and I have talked about so far fits into 2 buckets: the willingness and the ability.
What we’re just talking about now is the willingness. What is the return? Will the returns remain high? How high are the margins? Earlier on, we were talking about the ability. We talked about the balance sheet as a service. We talked about NVIDIA’s backstop, which is trying to address the ability, and who’s going to fund them. We talked a lot about that.
Why don’t we dip into the willingness? That’s what we were just talking about. We’ve done analysis where you can actually earn up to $100 million per megawatt generating frontier tokens off a GB300. It may be a little optimistic in terms of utilization, but when this thing costs about $40 million per megawatt to stand up, you make this back in half a year. It’s incredibly profitable.
Let’s compare that against what neoclouds are typically renting for a GB300, at least a few months ago: about $12 million per megawatt, or $12 billion per gigawatt, for a 5-year contract. If you’re doing frontier inference, like I said, you can do $100 million per megawatt.
People were shocked when they saw SpaceX and Google renting at $48 million per megawatt. But that makes sense, right? They’re buying here and selling here, so of course they would. People were used to seeing this, so the profit is there.
Jordan, I think you mentioned where the profit stays in each layer. What we did here is organize the entire business as an income statement. Imagine you’re Anthropic. This is your revenue, and we put it in a couple of different ways.
At about $1.85 per million tokens, which is a good blend of Sonnet and Opus with cache hits, this is actually our own blended spending. Even if Opus is $5 per input million tokens, the cache hits will take it down. So, about $1.85 per million tokens, which is equivalent to generating around 9,000 tokens per second—I can’t remember exactly what we used. It’s equivalent to $38 per hour.
We already talked about the cost per hour of $2.80 and then a neocloud. The cost to own the system is $2.80, but the cost to rent from a neocloud is around $4, probably north of $4 by now—$4.50, almost $5.
If we simply look at that, what’s the gross margin? It’s 90%, isn’t it? You’re selling at $38, your cost is $3.80, and this is your profit. It’s incredibly profitable.
Yeah.
At least for frontier models. By the way, this was really cool. Everyone should read this. Drop everything and read this now because—
By the way, this was really cool. Everyone should read this. Drop everything and read this now because—You know, Max and team—
We recorded a pod earlier today. You don’t know that. This is my third pod of the day. I’ve been cheating on you, Dan. But Max and Andrew went through this on—
Oh, my God. Gee.
Previous episode in case people missed it, do this one again. Maybe for people who missed it, do this one again. For the all-you-can-eat consumption plans on some of these businesses, you can get a lot of value.
But it depends on how much people actually use the plans that they’re given access to, if they’re going to be profitable at this layer or not.
Yeah. We peg a 90% API margin, and then I think in the model we use 50% or 60%. But the question I have is, how many of these people are really using 10%?
There’s a whole cottage industry in China that takes apples and makes applesauce. They take these $20 plans and just sell them as APIs on the gray market. And you know me—I’m on a $20 plan. I actually plan my sessions around utilizing it. I have loops so that I use it even when I’m not there.
I’m definitely in this bucket, but I don’t know where everyone else is.
Yeah. Max made the case earlier that there are a lot of people who maybe get these through their business or the company they work for, and they’re not really worried about the value trade exactly. You blend it over a bunch of users, and it works out.
The loud people are the ones pushing for more limits. They’re the ones you’re losing money on. But you keep them around because you don’t know how to separate the wheat from the chaff and only sell the all-you-can-eat plans to the most profitable segment of users, which is the ones who don’t use it.
The whole thing about Planet Fitness is that it markets gym memberships to people who come once in January for the new year and then never come back. A profitable gym is the one that’s empty. They don’t want anybody to show up.
That’s kind of cynical, but that’s the incentive structure for these guys to do all-you-can-eat plans. The exact opposite is true for per-token pricing, where they’re just trying to get people to consume as much as possible.
Yeah. The next question becomes, okay, it’s so profitable now, but is that going to be the case in the future? People ask, how fast is token price falling?
Let me find our token pricing. Where did I put the token pricing?
Yeah.
Sorry?
Yeah.
I mean—
Not to cut you off, but let me—yeah, okay. Do this one. Yeah, yeah.
My question for you, Jordan, is that people ask me, “How fast is token price falling?” What do you mean it’s falling? It’s going up.
Even if it’s staying flat, gross margin is going up because while token pricing is staying constant, they’ve been generating more tokens. Then they ask, “Okay, well, great. I see frontier models have been holding up, but how fast are open models falling?” I’m like, “Well, they’re going up.”
The other thing is, when you look at token expenditure, it’s a changing basket. These indexes that track token expenditure don’t adjust for that, in my opinion. I don’t know how they work, but they don’t adjust for that. With agentic systems, it’s been all about effective price and cache hits.
What do you think token pricing is going to do in the future, Jordan?
I had a little tweet thread where I went at David Friedberg from the All-In Podcast. He didn’t respond, of course.
A lot of people are saying that there’s this overtaking—open models overtaking closed models—but I think the dynamic is just that people are consuming more of everything, and they’re consuming it at higher prices.
You’ve got the price of input on screen here, which I think is a decent metric, but that little dropdown in the middle, you can—
You should do the effective one, right? I mean, you could. It’s just a different approach.
Yeah, you can do it for cache read, which I think is the bulk of our spend. It’s actually on the cache-read tier because agentic coding has so much cache-hit usage.
Output is the most expensive tier, even though you don’t actually generate a bunch of output. That’s what takes up all the time on the GPU and runs into your memory-bandwidth constraint.
Anyway, the point is that I totally agree that model pricing is going up over time on a blended dollar-per-million-token basis. I think token efficiency is a question mark because you have to trade off the ability to efficiently solve simple problems with people’s more grandiose expectations as they try to solve more complicated problems that weren’t possible before with unlimited tokens and now are possible with a lot of tokens.
I don’t know how you measure that trade-off. But none of these dynamics point to anybody deciding, “I’ve had enough of AI. I’m going to use less of it.”
I've had enough of these models. Let's make them worse and less efficient. I've had enough of these new GPUs. Let's make the next one worse and less performant. It all seems to be going in the opposite direction, right? They're going to raise prices over time, with more demand for them. The GPUs are going to be more efficient, and the models are going to have higher performance.
So we'll see. There are a lot of ways in which this can break if people decide to YOLO for the 20-trillion-parameter model with the new NVL72 and really use it. If they start introducing ultra-fast mode and some of these low batch-size optimizations and stuff that can jack prices up. But I think all of that is just artificial constraints on what, to me, feels like continued overwhelming demand for these products, which drives the usage of the GPUs underlying them.
Yeah. We talked a bit about the explicit income statement, but what about the income statement above it? What about the users? That's part of the equation too. What value are they getting? I don't see why anyone would step back when these are the kinds of efficiencies they're getting.
We built an entire database of thousands and thousands of debt facilities and all these relationships purely from the public domain. It's been myself, Oliver, and Terrence doing this, and we're all part of Current groups because we're hiring, like I mentioned in the PSA. But it would have taken me 4 months and a team of 4 to do half as good a job as we've done so far in building up this database and relationship table.
We've shown this slide a lot. We had a podcast on this. Again, I don't see why anyone who is paying $6 to investigate a company would ever go back.
Jordan, would you ever go back? Or would you just stop using fast mode? What would you do?
I don't use fast mode.
Me neither.
No, you'd have to pry this stuff out of my cold, dead hands, man. Even in my personal life now, with some of these support agents and stuff.
It's like, “Pick a wine. Pick a wine to bring to dinner with my friends.” It's like—
That's not what I'm talking about.
“Okay, Mom's for $3.”
Some stuff still has a personal touch, Dan, okay? I'm not writing birthday cards.
Speak for yourself.
I'm not that far off the deep end. But I just got a refund on a flight that I was delayed on for 4 hours. There's no way I would sit on the phone trying to get that refund ever again, right? I'm checking the times at the gym, booking a flight, looking for hotels, checking the grocery list, or just checking my email—what came in last night?
It's a little weird, but I talked to some friends who aren't enlightened to the possibilities that having a personal assistant can provide. It does feel like nobody can ever go back once they experience how much time they save and realize how much time they were wasting.
Yes. And I think Muse is going to be eye-opening. It's like, “Why did I ever sit there for 2 hours and browse all these airline booking sites? Why did I ever...?” You know? I think there's going to be that moment, right?
Yeah. It gets a little egotistical when you talk about it this way, I think, because you kind of anthropomorphize the agent. Then you view people who have personal assistants today as exploiting this thing.
But my personal view is that it's shocking how much time we waste on this monotonous stuff. It's shocking how some people I'm close with do not value their own time. I'm not going to go into all the details there, but you can—
I think we're sold on willingness, right? We're sold on AI being useful, right?
Well, I—
Book it.
I feel like, since working at SemiAnalysis, we've been at the tip of the spear for adopting a lot of this stuff and feeling the ability to use it. It makes our research better and informs the underlying research that motivates some of the data we see.
It's really nice to see other people we work with, and people in my personal life, adopt these things and start to regain control over a lot of things in their lives. It's really cool to see. I just get really happy when I see people have these good experiences using it and start to get motivated to build a more positive, high-quality future together. I don't know how to wrap it.
But talk to me about the willingness to finance. Do you think—
Yeah, yeah.
There's—
We just talked about that. I think we're sold. What I've got to convince you on is willingness and ability, right? I've got to sell you on ability.
Yeah.
I might have scared you with this, but—
Talk to me about the human element here, because one thing I've realized working at SemiAnalysis is how much of this stuff that seems so data-driven always comes down to some human being making a decision. Satya said “pause,” therefore pause. Jensen said “build,” therefore build, or backstop, or whatever.
Where does this break down to genuinely one of the leaders of these companies having this experience using a personal assistant or an AI, or just having that personal moment where you're like, “Oh my God, I can see the math discovery before it comes out and gets published,” or, “I can see this progress we're making on this cancer drug,” or whatever? And then, “I just really want to bet the company on this. I want to do this work because it means something to me, and this is my life's work,” as opposed to rationalizing all of the financial details to make it work.
Can you even explain that balance between those two things?
That is one of the big questions because we're still under willingness, right? We've covered that we're sold. I think what you're saying is, are other people sold too? That's what this chart is a little bit about.
I would make it slightly different. I would stratify it into consumers, API tip-of-the-spear people—AI natives—and then the rest of the enterprise. I think we're convinced on the tip of the spear. I feel like Manus is going to prove the case in consumer, but I don't know about enterprise.
I sit at these tables. I do a lot of these events, and I meet a lot of folks. We do talks and dinners on AI, and then Fortune 500 executives are like, “What would I use it for? Just summarizing stuff?” No, it's so much more than that.
I feel like there's a lack of discovery in the Fortune 500, a lack of imagination, and a lot of people don't come at this with an open mind. They come at it like, “Oh, great, AI. How do I fire people?” That's all they think about.
Yeah. It's odd.
The enterprise is the one of the three that I know the least about. It's so much about whether the CEO says, “Hey, let's embrace it. Let's trust Amazon. Let's go full blast,” or, “No AI in your work.”
Have you talked to Fortune 500 companies? Where do you think they are on this?
I think the dynamic that I imagine happens with the Fortune 500 is that the Fortune 500 slowly gets replaced with companies that are all-in on AI technology.
That's dire.
Look at the top 10 companies in the world by market cap. Nine out of 10 of them are AI-, semiconductor-, or data-center-related at this point, right?
That's possibly true.
When we say hyperscalers, all 5 of them are in there. Then you've got NVIDIA, Broadcom, and TSMC. The memory guys are getting close to the top 10.
So when we use the term “Fortune 500,” if you actually mean the 500 biggest businesses in the world measured by revenue, I think a lot of them are just going to—
Maybe I should say corporate America.
Legacy businesses, corporate America—
Legacy businesses, yeah.
Government—just people that are slow to adopt AI. Fair enough, I think there are lots of people who will be a lot slower to adopt AI than the biggest companies in the world, which I would lump all the hyperscalers into.
Soon to be the AI labs—if Anthropic goes and IPOs at $2 trillion, they're going to be a top-25 company in the world by market cap or something like that, right? So that's the proof point right there. But you have to make this case of going out 2 or 3 years. Do you expect that this next crop of new AI labs pursuing all of these innovations will just completely get swallowed up and acquired, and none of them will pan out?
I actually expect many of them to do quite well across a number of different industries, continue to grow, and, if they do get acquired, probably help those businesses grow and take over. I don't view a future where Visa, Bank of America, and AT&T are so critical to the success or failure of the AI-native companies that are trying to grow.
Maybe what I'm most interested in is the unlocking of capital. The companies that may have been funding those companies in the past may turn and shift over and say, “Hey, I work in infrastructure finance. I work in credit, and for the next round of this, I'm not actually going to do another project for another bridge or some debt for a credit card payments provider. I'm actually going to go into this AI thing because I think that's upstream of all those other businesses I've done, and it actually is infrastructure. I'm going to get these returns, and I'm going to have these guarantees and things like that.”
That's where I ask you, because you know these guys better than me, about the human element. For people who really don't understand this and don't have experience using AI, they represent so much capital that's still sitting on the sidelines. We joke about stuff like NVIDIA's revenue: if it hits $500 billion, it's going to be 2.5% of the M2 money supply in the U.S., right? Just unbelievable numbers we're playing with—trillions of dollars of CapEx we talked about earlier in this episode.
But there's still so much left over. What's the other 97.5% of the M2 money supply? So—
7. AI Debt Reaches Scale
Yeah, I can share some stuff, because now we're on the ability, right? One of the key things on the ability side is understanding, and one of the key obstacles is that lenders are on the learning curve. The people making decisions on lending are doing it because, okay, it's NVIDIA credit risk, it's Microsoft credit risk—I don't have to know. But that's going to change, and that's why all these backstops are there: so they have time to understand this industry.
Let me size the problem for you a little bit. Let's talk about where this is going, right? Obviously—
All right. Here we go.
It's going to be the second-biggest debt financing market, right? It's already surpassed auto loans and everything this year. But everyone has done it without having to have a residual-value model for GPUs because it didn't have a backstop, and it's only second to the U.S. mortgage-backed market.
Dan, move on from that chart.
Sure.
We've got some audio-only listeners, but I just want to describe it. It's such a compelling chart. Auto loans, student loans, credit cards, non-agency, CMBS, and HELOC—these U.S. asset-backed credit markets are just flat lines, slowly going up to the right.
You can't even see them going up, actually.
Yeah. They're just 5 flat lines at the bottom of the chart, and then there's this massive exponential line up to the right. More of a straight line, but, yeah, a straight line up to the right called AI debt financing. Explain to me the implications of AI debt financing in 2026 crossing over auto loans and student loans on this chart.
The implication is—and we already see it—we already see the long end. We already see 10-year rates going up. We already see credit spreads going up. The reason the credit market is going to be hard is that they don't have the tools to understand it, because it's all coming from a different front. It's a war on so many fronts, because you're getting hit by hyperscalers, neoclouds, and data centers—everyone at the same time.
It's like a zombie apocalypse, but you don't know how many waves of zombies are coming, right? The AI is like the Terminator. You know how many Terminators?
Give me your money.
Anyway, there are many zombies coming, right? You don't know whether you're on wave 5 or wave 50.
To put things in context, I've got a slide here that shows you what this is. Like I told you, I showed you what it is in terms of the corporate bond market, right? That's just up here. But on the next slide, let's talk about overall U.S. issuance.
AI debt is at $1 trillion, and it's going to cross $1 trillion. If you think about it, total U.S. issuance—this is federal, which is in blue—for those of you listening, that's about $5 or $6 trillion a year. You've got MBS and RMBS. Just ballparking that, it looks like $3 or $4 trillion a year of issuance, maybe $2 to $3 trillion. Corporate bonds, which we already talked about, are about $3 to $4 trillion. The total is $12.7 trillion.
2026 is the last year that AI debt is going to be less than 10% of all issuance in the U.S. So it's going to take believers. How do we handicap that? Maybe what we can do is jump back to the question you asked earlier: Why is NVIDIA doing these backstops?
Let's just look at—oh, sorry, I'm going to go. Okay, here. Yeah. Why do they have all these different backstops? They see this gap coming, and they're figuring out how to plug it.
Let's talk a bit about their AICP, right? By the way, to start off, we projected NVIDIA's balance sheet, and we were looking at its ability to continue doing these backstops and how much it can do. But let's talk about an example of one of the backstops and how it accomplishes those goals.
This is the AICP program. This actually started in Asia. The first one was ShareAI, then Firmus followed with another one. There's been Aolani[?], and there have been many others. GMI did one.
The whole idea here is the way this works: NVIDIA goes to a neocloud and says, “Hey, I'll give you a guarantee. If you can't rent your cluster, I'll rent it from you, and I'll give you that guarantee for 6 years at $3.25 on average.” The actual table is different; each year has a different minimum. But if you can't find clients, you can always go to NVIDIA and rent to them at this price.
The intention is not for that to actually happen. The idea is that you use this as a guarantee to then get financing. Remember the trend, right? So now you have an offtake, and now you can get financing, and then with that, you can get a data center.
With this in mind, lenders see this as being priced at NVIDIA risk. It says, “Hey, I'm really lending to NVIDIA plus execution risk.” What that allows the neocloud to do is say, “Oh, now I don't have to sign a 5-year offtake with whoever. I can do tokens as a service. I can do all these things.” It really opens up the market. That's the whole idea.
They don't need to be investment grade. That's the important part. But this is only an intermediate step, because what it does is lenders have to understand that they're also taking the platform risk—whether the neocloud can build a good book of clients. As they lend to the neocloud, they say, “Can they build an independent business outside the offtake?” They get to understand the customer profile, the churn rates, and the cancellation rates. They get to do a free look, or a dry run—a practice run—until the real thing. That's why this is really important. Does that make sense, Jordan? Any questions?
Makes perfect sense. Yeah, man—balance sheet as a service. Maybe the one thing to talk about is that at some point they introduced a take on the percentage of revenue that the neocloud generates above the minimum it promised.
Oh, I forgot about that.
And then they took that away.
No, no, no, it's still there, right?
Okay. Explain that.
It's also a revenue model, and it's fair—we're not communists, after all. In exchange for the backstop, NVIDIA gets a 50% revenue share above the backstop. For example, let's say you rented at $6.68. The difference is $3, so you have to pay $1.50 to NVIDIA out of the $6 in exchange for that.
So it’s actually a meaningful revenue share for NVIDIA. If you look at the other structures they’ve done, this is the Lambda structure; they’re also taking a revenue share out of Lambda in exchange for that guarantee.
A third reason, Jordan, why one of these backstops does this is not just to facilitate these deals and allow them to happen. In this instance, they’re allowing Lambda to get financing because Anthropic hasn’t yet IPO’d and isn’t IG, right? But with NVIDIA supporting the lease for the data center, they can actually get this done. There’s another reason, right? Let’s look at Ports Pike [?]. This is a little complicated, but the bottom line here is NVIDIA is giving $105 billion of residual value guarantees to the Ports Pike Data Center complex [?] in order to support this. The catch is—what they’ve done is guaranteed that 4.25 gigawatts, or up to almost 9 across 2 phases, will be running NVIDIA GPUs. So it also locks in. It’s like a kind of land grab, another reason for it.
We looked at all these different models, and I’ll talk about one last one. This is the most interesting one: the $500 billion capital partnership. This gets into how we can actually—to your point, right? You talked about lenders as one homogeneous group, but it’s actually very different. There are bank lenders and insurance lenders; they don’t want to take a lot of risks. You’ve got private equity—they want to take a lot of risks. You’ve got private credit—
Yeah, they all—
— they’re in the middle.
They all syndicate into these things as well, right?
Well, it’s about creating different products for different types of investors. So the $500 billion capital partnership—this is what we think it could look like. You can have an AI lab; they don’t have to be IG, but you tranche the risk. You have equity, then you’ve got a B-piece, which doesn’t have any residual value guarantee. Then you have an A-1 and an A-2; you can put one senior to the other. NVIDIA is giving a 25% residual value guarantee in this $500 billion partnership. Maybe you can attach this 25% to this tranche or this tranche.
But let’s say there are losses in this program, in this business. The equity will take the loss first, then the B-piece, and then part of the A-2. It will have to go all the way down to this. The loss has to be so severe to affect the senior tranche, which is how you can create different risk profiles and open up the market.
Anyway, all this depends, right, on NVIDIA having the capacity to do this. So we analyze the balance sheet, and if you love talking about this, this is all we’re going to do. Here’s the conclusion of this, Jordan: we think they have the capacity. This is not our forecast, but we think they have the capacity to support up to 46 gigawatts of capacity. We haven’t done it for the hyperscalers yet, but that’s still not enough. There are 240 gigawatts coming in. What I’m saying here is the lending market has to evolve to take risks outside of backstops. That’s what we’re all about: understanding that question. It’s really the biggest question of our time, to be honest, which is why we’re starting an entire group around this question.
Makes sense, man. If there are 2 tangible things I’m taking away from this, it’s, first, more education for lenders on what they’re actually getting into here—what the real deal structures are like, who the end-user customers are, and why they’re structuring things this way. And then, second—well, I guess the second is twofold. It’s always incredible to see the size of these numbers and how much has already been done, but also how much bigger it needs to get from here.
Mm-hmm.
Incremental 60-some-odd gigawatts of capacity additions in 2029. Yeah, let’s see.
These are incremental capacity additions.
So I do 40—
With your usage. Yeah, I know. I know. I know. I was looking at the numbers.
Let’s wrap up with one little funny story. When I got started working at SemiAnalysis, I didn’t understand anything about financing and how this stuff works. I was trying to figure out who a lot of these lenders were and what private equity was. I’d known these names and things like that. I remember working on this one deal—I won’t say who was setting up the deal—but it was private equity and—
Did you think EBITDA was like a musical at some point, Jordan?
Yeah, dude, the beast.
Right?
Yeah, yeah. I still get confused when you talk about project versus equity IRR. I don’t know. I just know that it’s really important, but—
Project versus equity IRR.
Equity IRR. Yeah, okay, okay. I just know IRR is—You spend enough time with Dan, you know IRR is important.
Yeah.
And it doesn’t get easier to say I-R-R.
IRR, yeah.
That was the pirate, right, Dan? When he started, it’s like he needs an eye patch because he’s talking about his eye, or IRR.
IRR.
Anyway, okay, let me tell this quick story before we wrap. This private equity company was setting up this deal. We had to go do this technical DD on it. We had to present to some of the lenders, and the lenders were from all sorts of different institutions. I remember at the beginning learning, “Oh, private equity doesn’t have any money for themselves. They have lots of assets under management, but they set up the deals, and then the lenders come in and actually fund the deal, and private equity makes carry off the top of that,” right? If anybody in PE who’s listening to this is like, “Yeah, of course that’s the way,” I don’t know.
Anyway, I had this meeting with this pension fund to explain the work that we did, and I remember sitting on the call. It was the Ontario Teachers’ Pension Plan, and I was like, “Oh, it’s my mom’s money.” This is what’s getting in here. So the whole loop back was thinking, okay, yeah, Canadian pension, Ontario Teachers’ Pension Plan, lots of stuff in Quebec—I’m Canadian, so anyway. I ended up feeling kind of proud of the work that we did, not like I was helping people who were just rent seekers trying to make a little bit of money off this. It actually is pretty hard to set these things up, make people comfortable understanding what they’re investing in, explain the risks, understand the technology, and figure out what the return is. There’s going to be more to do in the future because lots of people just like that are probably getting in and evaluating their first AI cloud opportunity or neocloud deal, and it’s going to be a big part of the future.
We’re on a learning curve, and I think the more we help people get up there and talk about everything, the more it helps everyone make great decisions.
Yeah. Cool, man. Okay.
So, one last joke. You know why it’s IRR?
No. Why?
’Cause private equity.
Pirate equity. Hey, everyone, if you're still listening after this hour-and-fifteen-minute-long conversation about credit and markets, you should consider applying to work at SemiAnalysis on credit and markets. Dan is a very fun guy to work with. Dan, who are you hiring for?
Well, we're hiring credit analysts in New York and Singapore. We're hiring junior credit analysts, so anyone who's done 1 to 3 years of experience, which could be investment banking or DCM. We're also looking for a more senior credit analyst, anyone who's been a desk analyst or investment analyst on the buy side, and also a credit specialist—someone who's quite in touch with markets, has a great Rolodex, and has worked in credit trading or credit sales, buy side or sell side. We'd love to have you apply. Just click here to find out more. I also need a video editor. If anybody knows about a video editor, shoot me a text.
Okay, this is your third interruption. We're experimenting with ads on the SemiAnalysis Weekly podcast. We also need a video editor. Akash is hiring.
Yeah, so if you know how to edit videos, please hit me up. We're looking for people.
Please hit me up.