The Subprime Data Center Crisis

**Thanks for reading this week’s free Where’s Your Ed At newsletter. Friday’s premium newsletter will ask the simple question: Is Oracle dying? **

It’s been one year since I launched the premium newsletter, and I’ve decided to extend the discount on annual subscriptions. Between now and 12AM ET, July 26, you can get a permanent annual rate of just $60— a $10 discount on the usual price of $70 — for life. Click here for the offer.

In addition to getting access to the entire back catalog of premium posts, you’ll also receive one additional post each week — usually anywhere between 10,000 and 20,000 words — covering the most pressing topics in the AI bubble — the best value in tech analysis. Highlights include the Hater's Guide To The Memory Crisis, a guide to how AI made everything more expensive, How OpenAI Kills Oracle (which pairs nicely with the Hater's Guide To Oracle), The Hater's Guide To NVIDIA, The Hater's Guides To Private Credit and Private Equity, and how the entire AI Compute Demand Story Is A Lie.

Soundtrack: Dillinger Escape Plan — Black Bubblegum (2007) In The Big Short, Mark Baum shook with anger as a CDO manager told him that the market for insuring mortgage bonds was about 20 times larger than the mortgage bond market, realizing in real-time that speculation driven by greed and hype had set up a massive systemic weakness under everybody’s noses.

To get specific, Baum (played by Steve Carell) is giving a short, dramatic summary of a much greater problem — that there were trillions of dollars of synthetic collateralized debt obligations ted(effectively bets on whether somebody else’s bucket of mortgages (well, mortgage bonds) will actually pay up) that allowed multiple people to bet on the same mortgages again and again, meaning that once said mortgages went belly-up, the carnage would be widespread and hard to contain.

This became even more chaotic when it became clear that the same mortgage bonds were attached to many different CDOs — one study found that 5500 different mortgage bonds had been placed or referenced in CDOs over 36,000 times. A mortgage bond (or mortgage-backed security) is a slice of a pool of payments from thousands of mortgages, with each slice sold off to different buyers at different levels of seniority, the most-senior ones getting paid first and taking losses last.

In the end, the only thing you really need to know is that financial institutions built CDOs that threw together bonds in ever-more complex and dangerous ways, selling synthetic CDOs to bet on the outcomes, with different CDOs having different bonds covering the same pools of mortgages — bonds that were routinely rated by agencies at a higher grade than they should’ve been. When IMF Chief Economist Raghuram Rajan

at attempted to warn the financial services industryabout the instability of the system, former US Treasury Secretary (and

the Kansas City Fed’s 2005 Jackson Hole symposium) Larry Summers referred to his concerns as “misguided.”

__close friend of Jeffrey Epstein__Meanwhile, the industry was handing out awards. On July 1, 2005 Lehman Brothers would receive one of Euromoney’s “ Awards For Excellence,” where it was named the “Credits Derivatives House Of The Year.” Euromoney also referred to Lehman,

, as “one of the more conservative credit derivatives houses.” It added that the company,

a financial institution that was leveraged 25.3x in 2005, was being able to take on the heavy burden of synthetic CDOs because it “...understands the arbitrage-driven economics of cash CDOs, the way that loan deliverable credit default swaps track the loan markets, how high-yield CDS trade (like bonds), and so on.”

__which routinely overvalued its CDOs__Three years later on January 1, 2008 — nine-and-a-half months before its collapse — Risk Magazine would name Lehman Brothers’ “Point” risk management system as its “In-House System of the Year,” saying it “...stood out for the breadth of its coverage and depth and quality of its functionality.”

All of this started because of a flood of overseas money in the early 2000s buying up U.S. Treasuries as a result of a “global savings glut” — a

— pushing yields down, leaving investors with far fewer places to get those all-important yields.

__fancy way of saying that there was too much money floating around__Low interest rates in the early 2000s (a direct response to the collapse of the dot com bubble) dropped mortgage rates to “generationally low” levels, and financial institutions realized they had an opportunity, as

at exactly the time that foreign investors were desperate for places to park their money — mortgage-backed securities, and their associated derivatives. More mortgages meant more mortgage-backed securities, so banks made it

government policies had allowed them to loosen underwriting standardsincrediblyeasy to get a mortgage, to the point that in 2006,

. __20% of all new mortgages were subprime__You’re probably wondering why nobody feared they’d get burned by this endless stack of different interconnected debts, and that’s because they’d* *“spread all that risk out” across credit default swaps with insurers, not realizing that insurers could and

become insolvent if everybody tried to make a claim at once.

__would__This was all avoidable, and there were many warnings, and just as many people lining up to protect the grift. In June 2005, Larry Kudlow would say that housing bears were “wrong again,” dismissing those concerned with increasing default rates as “bubbleheads” that “don’t do their homework.” In September 2006, financier Michael Milken would refer to CDOs in

as a “financial innovation” that “helped to spread risk and create tens of millions of jobs by freeing up investment capital for growing businesses,” saying that they would “increase prosperity by multiplying the value of human capital, social capital and real assets.”

__the Wall Street Journal__In other words, the argument was that the “financial innovation” of ever-expanding financial speculation was good for the economy because it created more money out of thin air, with the “risk” spread out somewhere, in a *way that you shouldn’t think about because everything is going to be fine. *Everybody would keep building houses forever, the numbers would only ever keep increasing, every new house would add a new mortgage to a new mortgage-backed security, and the line would only ever go up.

To put it all very simply, the great financial crisis was caused by inflated demand for housing caused by a mixture of historically-low interest rates and banks incentivizing bad habits as a means of increasing the value of speculative assets. In the end, “mortgage-backed securities” stopped existing as ways to invest in large swaths of mortgage payments, and more as high-risk financial vehicles that promised to be an infinite money glitch where nobody could lose because there would always be more demand for mortgages and, by extension,* collateralized debt obligations made up of mortgage-backed securities.*

It all broke because eventually those speculative assets had to interact with the real world, by which I mean mortgage defaults began to spike starting in 2005 with the expiration of teaser rates and

throughout 2006 making adjustable-rate mortgages creep upwards. As mortgages collapsed, CDOs — and their connected synthetic CDOs — collapsed with them, crushed by the weight of the consequences of offering so many people so many mortgages under volatile and unrealistic terms, and assuming that nothing bad would ever happen

multiple fed rate hikes*because nothing bad had happened yet.*Where bankers should’ve seen

And, fundamentally, the great financial crisis was caused by massive speculation based on demand that was, in and of itself, an illusion created by the financial institution itself to justify further investment.

Say, that kinda reminds me of something!

I realize that the comparison between an AI data center and a CDO might seem a little ridiculous, but they’re actually remarkably similar. I’m going to generalize here, because each of these deals has weird little unique terms that make them, well, *more dangerous. *

• When somebody decides to build an AI data center, they form a special purpose vehicle (much like a CDO), which then raises debt, in some cases slices it into tranches and, in most cases, sells them to institutional investors, asset managers or banks.

• Think of the SPV as its own little company (owned by the holding company, CoreWeave for example), and when somebody signs a contract with an AI data center company (say, OpenAI), they actually are signing a deal with the SPVrather than the company itself. - When the SPV receives the funds from the debt raise, it makes payments to contractors and suppliers (EG: NVIDIA for GPUs), and receives the revenue from the customer contract, assuming said customer is paying (or has anything to pay for).

• During construction (IE: pre-revenue), interest payments are taken out of the SPV from a pre-funded interest reserve account.

• When a customer pays, the SPV uses those funds to pay for the operating expenses of the data center, then creditors (based on their seniority in the debt), then, if anything’s left, the holding company. All of this money counts as revenue.

• These SPV-based data center debt deals also have a few fun little features:

• A DSCR (Debt Service Coverage Ratio) which means that the SPV must bring in a certain amount of revenue compared to its debt. For example, if an SPV’s debt had a DSCR of 1.15x and a monthly payment of $1.5 million, it needs to bring in $1.725 million in revenue after paying its operating expenses.- These often don’t begin until a date when the data center is theoretically operational, and yes, this absolutely could go horribly wrong with the amount of delays there are.

• A minimum liquidity requirement that, when breached, requires the holder to refill it or face default.

• A Debt Service Reserve Account (DSRA) set up after construction as a buffer if payments fall through.

• A DSCR (Debt Service Coverage Ratio) which means that the SPV must bring in a certain amount of revenue compared to its debt. For example, if an SPV’s debt had a DSCR of 1.15x and a monthly payment of $1.5 million, it needs to bring in

Put simply, every time somebody builds a data center, they form a completely separate entity that owns the chips, owns the debt, and, in many cases, owns most of the risk. These SPVs only pay out to their creditors in the event that customer revenue flows in, which means that they are dependent both on the speed of construction of said data centers and their customers’ ability to pay.

CoreWeave is the main offender in the SPV no IT loads refused cash-dump, with a different SPV for each of its Direct Draw Term Loans (DDTLs), most of them non-recourse, meaning that if their customers fail to pay, investors get screwed to varying degrees based on their seniority in the debt, and CoreWeave’s assets can’t be pursued in court, though it is on the hook for the payments on the debt.

For example, CoreWeave’s $8.5 billion DDTL 4.0 loan was raised using its contract with Meta and the underlying data center assets as collateral with funding coming from banks like MUFG, Deutsche Bank, and US Bank, with funds being deposited into an SPV called CoreWeave Compute Acquisition Co VIII LLC, with another filing showing that the funding and provided to an “investment-grade customer” that Wells Fargo believes is Meta.

__would be used to lease space from Applied Digital in Ellendale, North Dakota and fill it full of GPUs__Similarly, CoreWeave raised its $2.6 billion DDTL 3.0 loan last year to “accelerate delivery of services from OpenAI,” funding two different SPVs

that, in turn, signed a deal to provide compute to OpenAI through the main CoreWeave entity. The deal also features a “cash trap” that means that if either CoreWeave screws up (IE: doesn’t deliver the compute) and OpenAI can quit or OpenAI doesn’t pay for three months, the SPV stops feeding

called CoreWeave Compute Acquisition Co. V and VII, LLC.anymoney to CoreWeave until the situation is cured, and if OpenAI (or someone else) doesn’t start paying, things start to break, as the deal has a contract realization ratio of .85x,.

To be clear, “non-recourse” does not mean “CoreWeave gets off scot free if these SPVs collapse,” just that creditors can jump on the SPV’s assets first and cannot immediately go after CoreWeave’s assets, though because each of these deals is guaranteed by the parent company (CoreWeave itself), it will eventually be forced to make them whole.

You can probably guess how that goes badly.

The nature of these SPVs makes it difficult to quantify the exact scale of data center debt, but Bloomberg estimates that there’s over $500 billion in outstanding AI data center debt, with (per Garima Kapoor of Elara Securities Research) at least $200 billion of it held by private credit, making up roughly 8% of outstanding private credit loans.

That being said, the number is likely much higher. Nikkei Asia reported this week that Meta, Google, Amazon, Microsoft and Oracle have accrued around $1.65

trillionin outstanding debt in the last five years, with an additional hundreds of billions of dollars’ worth of “off balance sheet” debt, meaning that the corporate structure allows the company to not include it as part of its liabilities.

For example, BlackRock is currently raising $12 billion to build a data center for Meta, which in practical terms means BlackRock has invested in and is raising debt for a holding company called “Project Sopaipilla Holdings,” of which it owns 80% and Meta owns 20%. This holding company will then buy NVIDIA GPUs and pay construction firms to build the data center, and Meta’s (theoretical) payments will be used to pay down the debt. Despite the fact that Meta will (theoretically) own and operate as the exclusive tenant of this data center, the actual debt — $12 billion or more! — won’t appear on its balance sheet, much like its $27 billion Hyperion Data Center that belongs to an SPV called Beignet Investor LLC which is 80% owned by Blue Owl, 20% owned by Meta, and funded using bond sales to PIMCO and

. __BlackRock__The problem with these SPV-based deals is that they allow companies to, at least on a balance sheet basis, hide the scale of their debts. Meta’s long term debt sits, as of its latest quarter, at around $58.7 billion. It’s as if the $39 billion in debt for gigawatts’ worth of AI data centers doesn’t exist out of the payments it’ll eventually have to make.

This is all legal, worrying, and yes, *a little bit Enron. *

Per Amanda Iacone of Bloomberg: Enron Corp. exploited US accounting rules to hide from investors and lenders hundreds of millions in debt it had bundled into off-balance sheet entities — obligations that contributed to one of the biggest corporate collapses in US history.

Alphabet Inc. and Meta Platforms Inc. each have turned to vehicles known as variable interest entities (VIEs) as part of the financing mix needed to construct data centers and related energy infrastructure.

Meta, the parent of Facebook, last year formed a joint venture, a VIE, to build a Louisiana data center through a partnership with Blue Owl Capital. The social media titan’s maximum exposure for the venture is $46 billion, according to its filings with the Securities and Exchange Commission. The company announced last week that it would expand its planned campus and is expected to spend as much as $250 billion on the project, Bloomberg News has reported.

To be clear, a* Variable Interest Entity *is a type of SPV where you have control over the entity, and you *must *consolidate it into your balance sheet…unless you are not considered the “primary beneficiary,” which Meta argues isn’t the case despite being the primary tenant and reason that Hyperion is being built. Per Bloomberg:

Meta determined it shouldn’t bring billions in debt from the Louisiana project onto its own balance sheet because it isn’t responsible for finding tenants to replace or join it at the nearly 4,000-acre campus — a critical job that impacts the entity’s economic performance, the social media company said in its most recent quarterly SEC filing. Meta said its role is limited to construction management, along with administrative and property management services.

Auditor Ernst & Young raised a “red flag” ( per the WSJ) about this arrangement, flagging it as a “critical audit matter,” adding that it “...was especially challenging due to the significant judgment required in determining the activities that most significantly affect the VIE’s economic performance.” Nevertheless, it was approved, it happened, and everything is

fineand

normal.

This is why Google backstopped Fluidstack and Cipher Mining’s 300MW data center and . Both will, eventually, operate as data centers that Google will lease to provide compute to Anthropic, booking revenue for doing so, acting as the sole tenant and the entire reason that the debt was raised, yet because

another for TeraWulfFluidstackand

TeraWulfand

Cipher Miningare the actual entities involved, nothing shows up on Google’s balance sheet.

What’s also important to note is that none of the money going into these SPVs counts as capital expenditures. For example, across the space of five quarters ( Q1 2025 through Q1 2026), Meta spent around $88.6 billion in capital expenditures, but that doesn’t include any of the debt or purchases of GPUs or anything else

done in its name as part of the Hyperion SPV, despite it having (per its own fillings) $45.95 billion of exposure.

To be clear, even “on balance sheet” obligations are off-balance-sheet until the leases begin. Bloomberg has a truly horrifying chart that illustrates its scale:

Much like the Great Financial Crisis, nobody has seen any of these data center SPVs (or the greater data center bubble) as a problem yet because

I want to be very blunt about something: we do not, at this point, have a firm hand on exactly how much demand there is for AI compute, and evidence suggests that it’s much, much smaller than we’ve been led to believe.

I estimate that 70% or more of Microsoft, Google and Amazon’s compute capacity is taken up by OpenAI and Anthropic, and

, I struggled to find any customer

in my analysis of non-hyperscale compute providersotherthan them that was spending more than $50 million a year on compute.

That’s because real, diverse demand does not exist for AI compute, as evidenced by the fact that the same four or five companies are the only ones interested in renting it at scale.

Sidenote: Furthermore, any demand for compute that exists currently is inflated by the effective subsidization of paying subscribers to ChatGPT and Claude, who can burn tens of[.]thousands of dollars worth of tokens while only paying $200-a-month

For example, on July 1, [ Bloomberg reported](https://archive.ph/submit/?url=https%3A%2F%2Fwww.bloomberg.com%2Fnews%2Farticles%2F2026-07-01%2Fmeta-is-building-a-cloud-business-to-sell-excess-ai-compute&ref=wheresyoured.at) that Meta (

that “selling capacity was on the table if it overbuilt”) was creating a cloud business to rent out its AI GPU capacity. A mere two weeks later,

mere months after Zuckerberg saidthat it was in talks to rent capacity to Anthropic. __the New York Times reported__While one might argue that Meta is taking advantage of a wealthy buyer, one has to ask: if there was such insatiable demand for compute, why wouldn’t it want to sell it to a diverse set of customers who would likely pay a much higher rate than a years-long contract?

It’s because ** those customers do not exist at a scale that would actually make it worthwhile!** If they did, we’d see massive bursts of remaining performance obligations from neoclouds like Nebius, IREN and CoreWeave that were unrelated to new contracts they’ve signed with either hyperscalers, OpenAI or Anthropic. Companies like Lightning, Runpod, and Lambda would have billions in revenue. Instead,

[,](https://www.runpod.io/press/runpod-ai-cloud-surpasses-120m-in-arr?ref=wheresyoured.at)

__Runpod has $120 million in ARR__[, and](https://www.forbes.com/sites/iainmartin/2026/01/21/ai-startup-merges-with-a-billionaire-backed-data-center-operator-in-25-billion-deal/?ref=wheresyoured.at)

$500 million in ‘annualized’ revenue, with a little less than half of that coming from Microsoft and Amazon.

__Lambda had $114 million in revenue as of the second quarter of 2025__While the counter-argument is that these companies are all GPU-constrained, and that demand is simply waiting in the wings…except surely that would mean that these companies also had massive remaining performance obligations?

To be clear, the point I’m making is not that there’s no demand, just that the vast majority of that demand is coming from either Anthropic and OpenAI — two companies that cannot afford to pay for it long-term — and hyperscalers, who are mostly buying compute on behalf of OpenAI and Anthropic.

And I’m not sure that people are taking me seriously when I say that AI compute demand does not exist at the scale that it needs to, will likely never reach that scale, and data center construction is a debt-funded asset bubble with ruinous consequences.

So, let’s set some table stakes.

Per my own analysis, NVIDIA’s predicted $1 trillion in Blackwell and Vera Rubin GPU sales (by the end of 2027) represents around 40GW of data center capacity, which will, assuming a PUE of 1.35, result in around 30GW of usable capacity. At a cost of around $12 million a megawatt, that works out to around $435 billion in global annual compute revenue to make these data centers necessary.

Right now, there appears to be roughly $100 billion or so in annual compute spend, with OpenAI representing around $50 billion ( per their statements in the Musk trial)

, and Anthropic likely spending similar amounts. Microsoft and NVIDIA represent a combined 65% of

, with the rest likely taken up by OpenAI. IREN, another neocloud, CoreWeave’s $2.08 billion in (latest) quarterly revenueit was targeting a year-end cloud ARR of “over $4 billion,” or around $333 million a month, with a customer base that

recently announced, unsurprisingly, Microsoft and NVIDIA, as well as companies like Perplexity, Figure AI, and Together AI that are unlikely to be spending more than $50 million apiece given their funding status and revenues.

__includes__Sidenote: It’s important to note thatspendisn’t necessarily the same asutilization. Much of the compute spend comes in the form of long-term contracts, where the price-per-hour for a GPU is (naturally) lower than it would be if it was rented to deal with a momentary spike in utilization. The vast majority of AI compute spend comes in the form of these long-term contracts.

Another concerning anti-demand signal is the fact that NVIDIA has committed to $30 billion in multi-year cloud compute agreements, spending $6 billion or more a year through 2028 to rent back its GPUs,

that explicitly states that NVIDIA is “is obligated to purchase the residual unsold capacity” through April 2032, suggesting that there would be residual capacity that had gone unsold to the tune of billions of dollars.

__including a $6.3 billion backstop for CoreWeave__Oh, and NVIDIA owns 9.3% of Nebius too. It’s also invested in

[, CoreWeave and](https://www.reuters.com/business/nvidia-invest-up-21-billion-iren-part-ai-data-center-deal-2026-05-07/?ref=wheresyoured.at)

__IREN__[.](https://finance.yahoo.com/news/nvidia-signs-1-5-billion-150328535.html?ref=wheresyoured.at)

__has both invested in and rented capacity from Lambda__If you’re wondering why these deals keep getting signed — as mentioned previously — it’s because a financial guarantee from NVIDIA is sufficient collateral for a bank to lend money to these companies to buy more GPUs.

I imagine a conversation in the Big Short 2 might go a little like this scene.

So there’re these companies I invest in that, at least in theory, build data centers using my AI GPUs, but I need them to buyHUANG:moreGPUs, so I sign a contract saying that I’ll rent the GPUs back from them in the future. Because NVIDIA has such a strong balance sheet, these companies can raise billions of dollars to buymy GPUsjust because Ipromised to rent them in the future, and the best part is all the risk is held by the companies and the investors. When I need more money, I just sign another contract, they raise more debt, I sell more GPUs.So — just so I have this clearly — you, the guy who makes the GPUs, invest in companies that exist pretty much to buy GPUs from you and rent them to customers. Except you’re the customer too, and a big one.BAUM:That’s right. We call them neoclouds.HUANG:[.]S&P just revised CoreWeave’s outlook to positiveThat’s fucking crazy.BAUM:It’s not crazy — it’s awesome.HUANG:

Those Meta and neocloud deals exist explicitly to lower their capex and debt — by which I mean that if Nebius or IREN takes on the billions in debt to buy all of those GPUs, Microsoft and Meta only have to worry about the ongoing leases, assuming that construction is ever complete. These deals also regularly include a clause that allows them to be terminated in the event that delivery milestones are not met,

Microsoft.

__as is the case with Microsoft’s $17.4 billion deal with Nebius__This means that hyperscalers take on effectively no risk, and investors are left holding the bag. For example, Nebius’ recent $775 million debt facility is “backed by contracted cashflows and deployed GPU infrastructure,” meaning that if things fall apart, the only entity that can be sued would be a company that explicitly exists to buy NVIDIA GPUs and rent them.

To be abundantly clear, the vast majority of the AI data center compute revenue is contingent on the continued ability of two unprofitable, unsustainable AI companies’ to raise tens or hundreds of billions of dollars a year. This is not an overstatement, this is not hyperbole, it is the quite literal situation we’re stuck in.

Putting aside whether data centers are profitable or not ( they aren’t), if the demand does not exist at this remarkable scale, the vast majority of AI data centers and their associated SPVs will collapse.

If we take February’s Sightline Climate report at its word, there is 190GW of data center capacity in planning, or 140GW of IT load if we take a 1.35 PUE, for a total of $1.68 Trillion. If we assume — and I’m being nice! — that there’s $120 billion in annual compute demand, and take into account that tens of billions of dollars’ worth of data centers have been announced since, this means that there’s over 15 times more data centers being planned than the demand that actually exists, and 70% to 90% of that demand is from Anthropic and OpenAI’s unprofitable services.## The Global Savings Grab, Or How Everybody Misrated The Risk Of AI Data Centers (An Misunderstood Demand)

The hunger for speculation has vastly outpaced the actual demand for AI compute, much like it did in the great financial crisis, and for many of the same reasons. Back in May, JP Morgan’s Karen Ward brought up the global savings glut that I mentioned in the intro as part of a discussion of what she calls a “global savings grab”:

In short, a number of Asian countries were either scarred from years of financial crisis and balance-of-payments turmoil in the 1990s, or seeking to depress their currencies to focus on export-driven growth. This created a glut of savings that travelled abroad, with US government bonds top of the wish list.

[US Fed Chair] Bernanke highlighted that this was acting as both a blessing and curse for the US. It provided the US Treasury secretary with abundant cheap financing and so too for American households and businesses. But this cheap capital was too tempting, running the risk of overspend and bubbles in the US. Within a few years, this warning proved correct.

Well, good thing that the world is different now, right?

That wasn’t the end of the savings glut story, however. Stage two happened when key parts of Europe shifted from being net spenders to net savers after the Eurozone sovereign crisis.

The global savings glut got bigger, as did the investment flows into the US, one of the few areas of the world still happy to spend and grow. At the end of 2025, America had a negative net international investment position — the difference between US-owned foreign assets and foreign-owned US assets — of $27tn.

However, we are now transitioning to what could be termed a global savings grab. Governments and companies around the world want to spend more. In doing so, they are competing to offer the most attractive opportunities to the world’s savers.

The difference between a savings glut and savings *grab *is that there’s incredible demand for cash rather than an excess of capital to invest, at a time when banks (and private credit funds)

but have tons of cash… and investing in AI data centers, which they consider to be the “cheat code” — high-yield, low-risk investments in infrastructure that have “guaranteed” customers.

__are pulling back from investing in software and healthcare companies due to AI-related risk__It’s a perfect storm that mixes dangerously with the $400 billion or so in private infrastructure funds waiting to deploy, much of which is funded by pension and insurance funds (

) drawn to private credit — get this! — as I covered in the Hater’s Guide To Private Creditbecause banks were restricted from making the same kind of reckless bets that caused the global financial system to implode.

because they needed new things to invest in after the Great Financial Crisis made yields difficult to find*Are you beginning to work out why I’m a little concerned? *How about the fact that the amount of dry powder within these retail-focused financial institutions is shrinking — suggesting that more and more cash is being deployed, or withdrawn as a result of diminishing confidence within households.

Anyway, much of the assumption of how “safe” investments in AI data centers comes down to three ideas:

• AI data center demand is infinite and all compute will be used.

• AI data centers allhave “locked-in customer demand.”- This is, to be clear, fundamentally untrue. The only guaranteed, locked-in customer demand I can find is from Amazon, Google and Microsoft (for OpenAI and Anthropic), Meta, and Google and Anthropic. While there might be some random AI firms or inference companies that have “locked up capacity,” their dollars are only as good as their access to venture capital, much like Anthropic and OpenAI.

• That these are “safe” investments, backed by the richest companies in the world.

• This idea comes from the child-like belief that because Microsoft, Google, Meta and Amazon are the richest companies in the world and are signing 10-to-20-year-long leases, that *tons of other companies will do the same,*and that AI data centers and their debt should be valued as such.

• This idea comes from the child-like belief that because Microsoft, Google, Meta and Amazon are the richest companies in the world and are signing 10-to-20-year-long leases, that

Financial institutions have built entire models based on logic that borders on childish.

The “proof” that it’s worth investing in data centers mostly comes down to seeing that hyperscalers are spending a lot of money on them, and that OpenAI and Anthropic have lots of demand for compute.

They have also mistaken the ability for hyperscalers to keep funding data centers out of cashflow as a sign that all data centers are a good investment, when what’s actually happening is that they’ve run out of hypergrowth ideas and had so much free cash sloshing about that they were able to spend a trillion dollars in four years. Hyperscaler demand for NVIDIA chips has been so significant that it made it

looklike NVIDIA had an insane amount of demand, which in turn created a degree of FOMO and speculation, with everybody assuming that because

hyperscalers were getting rich(they weren’t, they have never disclosed their AI revenues, but people just assume they wouldn’t do this without making a profit) that

they too would get rich by buying GPUs and building data centers.

NVIDIA has been a big part of creating this fake demand story with its investments in — and backstop contracts with — CoreWeave, Lambda, IREN, and Nebius. Much like people assume hyperscalers wouldn’t make a huge, trillion-dollar mistake, they also assume that NVIDIA wouldn’t invest in companies that weren’t going to see incredible demand, somehow ignoring the very obvious point that **NVIDIA doesn’t give a shit about any neoclouds outside of their ability to generate more GPU sales. **

This is where the media and analysts could’ve done their jobs, but because none of the neoclouds have done yet, it’s totally fine that CoreWeave is sat on $30 billion in debt, most of it impossible to pay if any client drops out of a contract, *because, much like the great financial crisis, nothing bad had happened yet, *by which I mean that clients leave their invoices unpaid, and CoreWeave finds itself in financial distress.

NVIDIA’s naked self-dealing and circular financing are only made possible with a completely captured tech and business media. While mildly-concerned stories have run for the last year or so about the “massive circular financing under the AI bubble,” none of them treat the situation as anything else other than a curiosity.

I cannot adequately express for those that have hand-waved the danger of this bubble, or tried to minimize the risk created by the overbuild of AI data centers.

Much like a subprime mortgage, AI data center debt is being poorly-underwritten, virtually-uncollateralized and issued to projects that have extremely low likelihoods of repayment, all based on flimsy information and hype-driven mania.

Their collapse is inevitable because their ongoing payments are made out of customer revenue that is, in the vast majority of cases, entirely theoretical or contingent on payments from unprofitable and unsustainable AI companies.

What differs this from the subprime mortgage crisis is that the systemic risks aren’t driven by derivatives or complex financials but by the sheer scale of costs to build an AI data center, a catastrophic misunderstanding of the AI industry itself and the dangerous lending standards of private credit. When every single debt deal is over $500 million and usually numbering in the billions, we don’t need a vast web of different contracts to create a systemic risk, just clusters of projects that either fail to keep up with their SPVs’ debt or bonds that go unpaid by destitute or defunct data center developers.

Also, please remember that it didn’t take massive losses to begin the great financial crisis — just hard hits to a few load-bearing pillars of the industry. Lehman Brothers suffered two sequential quarters of losses ($2.8bn in Q2 CY2008 and $3.9bn in Q3 CY2028) before it entered liquidation. Those losses weren’t what killed it, but rather, what those losses did to the broader market — as well as the perception of Lehman with potential saviors.

Similarly, Bear Sterns failed after two hedge funds under its umbrella collapsed. While the monetary losses from these funds weren’t insignificant, they were something that, if everything else was fine, Bear Sterns could recover from them. Sadly, they occurred at a time when the market was spooked, and Bear was spectacularly over-leveraged, meaning that marginal losses would have a disproportionate impact on its balance sheet.

For AI, those “hard hits” to the “load bearing pillars of the industry” means a large amount of capacity flooding the market at once — such as that caused by the failure of a major compute customer, [ most likely OpenAI ](https://www.wheresyoured.at/the-openai-bubble/)— or the slow arrival of new capacity that can’t find revenue to pay for it. Perhaps we get both.

While the data center debt market might be much smaller than the trillions of dollars of (at least theoretical) securities that broke the back of the financial markets (I estimate somewhere between $500 billion and $750 billion), the risk — the actual underlying financing — is spread across the entire financial system, with every major bank and financial institution and the vast majority of asset management firms having billions or tens of billions of dollars’ worth of debt tied up in an impossible situation.

The scenario I’m talking about is one where the vast majority of AI data centers go unused, and because the vast majority of data centers are paid out of customer revenues, **80% or more of the funds invested in AI data center debt will be lost. **None of this is written to be alarmist or hyperbolic, and represents a rational position when compared to the fact that we have over 15 times the amount of data center capacity than we need, and there is little compelling evidence that there’s more than a few billion dollars in total demand.

Sidenote:If anything, I consider it a radical and actively deranged position to believe that there is demand for all this compute, and find it distressing that nobody else is making any substantive attempt to measure the actual demand for AI compute, or have any kind of meaningful discussion of what “we overbuilt capacity” means.

This all makes me feel a little crazy. I’ve looked! I think I’ve looked harder than anyone else has! I’ve tracked down as many sources as I can of the actual amounts that people are paying to rent AI GPUs. The demand isn’t even remotely there!

I even founda bizarre mid-2025 presentation from the owner of the land that Stargate Abilene is built on top ofthat prices the 1.2GW of capacity at around $10 billion a year. I imagine nobody wants to talk about this because it suggests that, at scale, AI compute only sells for about $12 billion a gigawatt annually, which makes the whole “data centers[” mathematics a little hard to reconcile with.]cost $100 billion a gigawatt

I need everybody to realize that the overbuild has real consequences, and that there’s no post-bubble economy to replace this one, because it’ll be just as expensive to run these things in 2030 as it is today.

This will mean that effectively every single financial institution in the world will have to write off or mark down hundreds of millions or billions of dollars’ worth of loans — and when they go to sell the underlying assets, they’ll be dumping aging hopper and Blackwell GPUs into a market saturated with them, meaning that the salvage price they get — assuming they get one at all — will be negligible.

This is the data center equivalent of subprime loans defaulting, except instead of *hundreds of thousands *of loans being the trigger, all it takes is ten or fifteen of them to send the industry into a panic.

It’s easy to dismiss this entirely as “rich people problems,” but AI data centers are increasingly funded — both directly and through private credit — using pension and insurance funds that rely on these (theoretical) payments for future yield to pay out premiums.

I’ll give you some examples.

• Australian AI infrastructure company Morrison, backers of CDC (the largest data center operator in the country), .recently convinced Japanese bank SMBC to allow it to invest its pension funds in AI data centers in the country - IPI Partners, , has a limited partner (read: people funding it) base, per Deutsche Bank, split into equal 25% chunks made up of sovereign wealth funds, family offices, public pensions, and insurance/private pension endowments.a one of the largest private data center investment firms that is now owned by Blue Owl - The California State Teacher’s Retirement System is the biggest investor in Blue Owl’s publicly-traded Blue Owl Capital Corporation fund.

• Blue Owl funded Meta’s Hyperion data center and Stargate Abilene, amongst other deals.

• In 2024, ,Blue Owl acquired insurer Kuvare__for funding Meta’s Altoona-based data center, which was, at the time, Meta’s largest data center.which won the Great Des Moines Partnership’s “Deal Of The Decade” Award in 2023 - CDPQ, one of Quebec’s largest pension funds, .invested in CoreWeave’s $7.5 billion DDTL 1.0, as part of its CDPQ American Fixed Income V Inc fund - A few weeks ago, Asset manager , and did so funded by billions of dollars of insurance annuities it’s able to play with as a result of__Apollo Global raised $35 billion for Broadcom to build Google TPUs for Anthropic to lease. If these payments aren’t made (though Broadcom has backstopped them), it will directly hit Athene’s ability to pay out insurance and retirement premiums.its acquisition/merger with insurance and retirement firm Athene- One worrying quote from the piece, emphasis mine: “What also sets Apollo apart is its homegrown trading operation, further blurring the lines between the alternative asset manager and Wall Street banks. It has also become one of the largest forces in insurance, prompting concerns that the firm and its peers are ramping up risk in a once-sleepy part of finance, and at a pace that makes it difficult for regulators to keep up.”

• One worrying quote from the piece, emphasis mine: “What also sets Apollo apart is its homegrown trading operation, further blurring the lines between the alternative asset manager and Wall Street banks. It has also become one of the largest forces in insurance,

The other problem with the “private” part of private credit is that we don’t really know how much data center exposure pension and insurance funds and the insurance/retirement funds of asset managers actually have. What we do know is that private credit is sinking hundreds of billions of dollars of people’s retirements and insurance premiums into deals based on

. __obfuscated valuations and questionable underwriting standards__For an example of how lax those standards are, here’s a quote from The Information about Blue Owl’s due diligence on Stargate Abilene, emphasis mine:

Blue Owl’s willingness to make fast decisions has made it a favored partner for the developers racing to produce giant data centers.It took just 15 minutes for Blue Owl executives to agree to invest up to $10 billion in future projects alongside real estate firm Primary Digital Infrastructure during their first in-person meeting two years ago, said Primary chief investment officer Bill Stein. Later, Stein introduced Lipschultz to developer Crusoe, which led to Blue Owl investing in the Abilene data center.

To make matters worse, Moody’s estimated a few months ago that banks had around $1.4 trillion in exposure to private credit,

. __with $300 billion of that exposure held by big banks__And because neither banks nor private credit funds nor asset managers are forced to keep any level of reserves, all it takes is a few bad apples — a few billion of data center deals — to go pear-shaped for there to be a cataclysmic unwinding of the AI trade.

So, the reason that *nobody *is really worrying about this situation is that we’re still waiting for the vast majority of data centers funded so far to complete construction. Once that happens, the assumption is that either A) the client in question will start paying or, more likely, B) that the data center provider will simply expect said customer to appear.

Eventually, these data centers ( which are taking 18-36 months to complete) will start turning on, which will require them to start having paying customers at a scale that the market can’t actually support. While subprime mortgage defaults were a kind of slow, ugly boil, it’s much more likely that the collapse of the subprime data center bubble will happen in fits and starts as capacity comes online and, assuming Anthropic and OpenAI don’t swoop in, goes unused.

Sidenote:The most frustrating part of working out when this all collapses is the deliberate obfuscation by just about every party of the current state of data center construction.[that there were only 5GW of data centers due in 2026 actually under construction, but that number dramatically increased when you added 2027 and 2028 — 7.6GW and 2.1GW for those years respectively — which doesn’t actually line up with the amount of GPUs sold by NVIDIA (around 4GW to 5GW a year), or any attempt to line it up the publicly-announced data centers.]Sightline said in February

This is a problem of information asymmetry I’ll get to shortly.

I think we’ll see a few rescue missions to try and keep the con alive. Hyperscalers will do everything they can, scooping up capacity anywhere they see it, to avoid the perception that AI data centers will go unused.

You see, hyperscalers are currently in their own confidence game as a result of their ruinous expenditures creating the illusion of demand. Microsoft, Google, Meta and Amazon are stuck in a terrible situation where building *more *capacity will cost them tens of billions of dollars, but stopping building capacity will be an immediate signal that they’ve overbuilt capacity, sending anxiety-strewn shockwaves through the industry and killing data center debt issuance.

Yet what also might kill issuance is the market itself. Per Bloomberg, AI data center debt has “hit a wall,” with 80% of data center securities issued since early 2025 quoted at a wider spread than issuance, meaning that investors are valuing them as worth less than when they were initially issued. If the market continues to sour on AI data center debt, it will eventually become difficult to impossible for hyperscalers to keep issuing bonds, leaving them with only equity sales (

) that are equal parts limited and desperate. __like Google’s $85bn stock sale__Any decline in appetite for AI bonds will be immediately obvious, given the scale in borrowing, with (per Goldman Sachs) AI-related bonds accounting for nearly one-quarter of all US-investment grade debt issuance.

NVIDIA’s continual circular funding of neoclouds and anyone who wants to buy NVIDIA GPUs continues only as a marketing function, and an attempt to conjure up the illusion of insatiable demand for AI compute. These deals are acts of desperation themselves, and tacit admissions that without NVIDIA, none of these neoclouds would exist — though, to be clear, Jensen Huang has quite literally said this on camera.

This, in my view, represents another troubling parallel between the AI bubble and the subprime mortgage crisis. In the early 2000s, loose lending standards — combined with the securitization of mortgages, which allowed lenders to offload their risk to third-parties — made it possible for people who shouldn’t have been able to obtain a mortgage to buy a home, albeit often at worse terms than so-called “prime” borrowers.

As I outlined in Coreweave Is A Time Bomb, Coreweave has been able to borrow tens of billions of dollars, despite having a business that is fundamentally reliant on a single customer — OpenAI — and on terms that would make a mafioso loan shark and say “hold on, that’s a bit harsh.”

In a sane world, Coreweave should not have been able to borrow as much as it did — and the same applies to the countless other debt-laden neoclouds that are, for the most part, cookie-cutter versions of Coreweave. These are the subprime homebuyers of the AI bubble, and the backstops and freebies offered by NVIDIA (and the other hyperscalers) has allowed said companies to raise more debt, without actually changing the fundamentals of these companies that made them so inherently risky to begin with.

Another obvious trigger is the insolvency of OpenAI or Anthropic, who have $1.1 trillion in compute commitments across Microsoft, Google, Amazon and Oracle, and, more dangerously, another $50 billion across Cerebras and CoreWeave.

In fact, maybe insolvency is going too far. CoreWeave’s own master service agreement with OpenAI will breach investor covenants if OpenAI fails to pay for three months straight, and

, it’s going to either need to raise more money or not pay its bills. __with OpenAI delaying its IPO to 2027__In any case, the sheer scale of AI data centers coming online massively outpaces the demand for AI compute, and **to be clear, the AI bubble doesn’t have to burst for the subprime data center crisis to begin, because all it takes is for the revenue to not exist to pay for the compute. **

As the vast majority of AI data center debt is project financing funded by compute revenues, the cutesy and half-assed retort of “even if it’s an overbuild, it’ll be alright” doesn’t really matter very much. The second a debt-backed data center is built, it must immediately produce revenue, because otherwise creditors will be left unpaid.

While the sheer scale of who those unpaid creditors might be is hard to quantify, what we *can* quantify is that the risk of the subprime data center crisis is everywhere — your bank, your community, your pension fund, your insurance company, everyone has, on some level, *some* exposure to the bubble.

For me to be wrong, there will have to be dramatic amounts of AI compute demand — hundreds of billions’ worth — within the next 3 years, at a time when there’s little more than $120 billion, with 80% or more of that coming from two companies that can only afford it because they have near-infinite sums of venture capital behind them.

Oh, and for some context, the entire global software market is estimated to be around $779 billion in 2026. It’s unclear where that money will come from, and nobody seems to want to talk about it.

The AI bubble is, like the great financial crisis, a product of information asymmetry — companies intentionally obfuscating how much revenue they have, how much data center capacity they have, how much revenue that capacity generates, how much demand they have for their services, and how many real dollars actually flow from the AI industry outside of investments in semiconductors.

And much like the great financial crisis, modern tech and business journalism routinely defaulted on its responsibility to demand this information, or to present a lack of information as suspicious, choosing instead to fill in the gaps and assume that whatever bullshit a rich person peddles is the truth. While many media outlets — as they did in 2007 — are

nowtrying to

somewhatquantify the risks, most business publications

, act as if AI’s continued growth is a certainty, and bring up financials only as a hesitant warning about an otherwise-solvent and successful industry.

__continue to celebrate every time NVIDIA sinks billions of dollars into a neocloud that exists only as a means of selling more GPUs__Similarly, entire research groups — regularly quoted by the media — exist to inflate the bubble further. Exponential View’s speciously-sourced and questionably-founded research paper on AI revenues was used

and multiple other media outlets as a means of saying that “AI had started to pay off” because “the quarterly revenues were now outpacing depreciation,” a completely nonsensical statement that compared

by Bloomberg*revenues from the entire AI industry (unspecified and undefined, by the way)*with the

depreciation of GPU hardware by hyperscalers.

It’s hard to see this as anything other than some tech and business journalists having a vested interest in seeing the AI industry win, which means, by proxy, that some writers will be fundamentally responsible for what follows when the bubble bursts. This is not a deliberately hyperbolic statement (and, lest I be accused of tarring the media and analysis industry with the same broad brush, there are many exceptions, and many good reporters calling bullshit where justified), but it seems as though there is a concerted effort to support industry narratives that I find repulsive.

What’s most horrifying is that OpenAI and Anthropic don’t even have to die for this all to end horribly. For one company to be able to afford even $200 billion a year in AI-related operating expenses is ludicrous.

Microsoft, a company that makes about $318 billion a year in revenue, has

, and that includes the cost of running OpenAI’s compute. The idea that we’re going to have multiple Microsofts-worth of opex entirely focused on AI compute in the next four years is absolutely fucking ridiculous, yet it’s one of the most commonly-held beliefs in the tech industry.

__about $169 billion of operating expenses a year__As I hope I’ve made clear, I believe the vast majority of AI data centers are the AI bubble’s subprime loans, and will collapse when they face the cold, harsh reality of “someone actually paying money for AI compute,” much like subprime mortgages collapsed when teaser rates ended and homeowners were forced to pay their actual bills.

This is an inevitability — and something that’s very obvious when you sit down and actually try and work out how much capacity there is versus how much people are actually paying for AI compute. The fact that I, a random guy, albeit with a (recently-acquired) Bloomberg Terminal, am the one to say this is a sign that the media is not trying hard enough to protect consumers.

Every time that the media has accepted a spurious announcement or a questionable run rate or a circular deal or the outright refusal of hyperscalers to disclose their AI revenues, they help inflate the AI bubble and endanger the futures of millions of people, especially those tied up in a stock market increasingly-dominated by NVIDIA and other tech stocks.

Unlike the great financial crisis, the calamity to follow will be easily-traced to a complete failure of anybody to measure or demand measurements of the actual demand for AI services and AI compute.

There will be attempts to claim it was too complex or multi-faceted to pry apart, and those attempts will likely be made by media outlets that failed their readers, viewers, listeners, and the general public.

Bubbles can only inflate in an information-poor — and information-deprived — environment.

They inflate much faster and more-dangerously when that information is poisoned by marketing spiel and misinformation peddled by those who are meant to tell people the truth.

**If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 10,000 to 18,000 words, including vast, detailed analyses of the biggest events and companies in the AI bubble. **

As a reminder, if you sign up between now and 12AM ET, July 26, you’ll get $10 off a subscription. Click here for the offer.

source & further reading

wheresyoured.at — original article