Your 401(k) Is Collateralized by AI Hype and Nobody Told You

This is the second piece in a series. The first, The AI Industry Is Moving Money in Circles and Calling It Growth, traced the circular capital flows between NVIDIA, the hyperscalers, and the AI startups they fund. That piece was about rich people playing games with each other's money. Absurd, probably going to blow up, but in isolation, mostly their problem.

This piece is about where those games connect to yours.

The Pipe Nobody Is Talking About

Politicians and regulators keep repeating the same reassurance: the risk in AI is concentrated in privately held companies. The IPO window is closed. Average Americans are not exposed. Relax. This is wrong. Not a little wrong. Structurally wrong. And the reason it is wrong is a financial instrument that almost nobody outside of structured finance is paying attention to, which is exactly how these things work every single time.

Banks have securitized $48.69 billion of data center debt into tradeable bonds since 2018. Eighty-eight deals. The market is projected to hit $30 to $40 billion in annual issuance by 2026. And I need you to understand what that means in plain terms, because the entire game depends on you not understanding it. A data center operator, typically backed by private equity like Blackstone or Brookfield or DigitalBridge, builds a portfolio of data centers with long-term leases to tenants like Amazon, Microsoft, and Google. They set up a special purpose vehicle, a legal shell designed to be bankruptcy-remote. The SPV issues bonds backed by the lease revenue. The bonds get tranched: senior notes that get paid first, mezzanine notes that absorb losses first. Rating agencies rate them. And then they get sold to insurance companies, pension funds, and sovereign wealth funds. The people managing the money you are counting on to retire.

This is not small. Switch has issued roughly $3.5 billion in asset-backed securities plus a $2.4 billion CMBS deal. Latham & Watkins, one of the most powerful law firms in the world, represented the underwriters. Both offerings oversubscribed. DataBank has done five securitizations totaling $3.23 billion. Blackstone's QTS Realty issued a $3.46 billion deal in November 2025, likely the largest CMBS deal of any type that year. The lowest-rated tranche was 23 times oversubscribed. There is more demand for this paper than there is paper. Meta created a $27 billion SPV called Hyperion, arranged by Morgan Stanley through Blue Owl Capital. $27 billion in A+-rated debt, anchored by PIMCO at $18 billion and BlackRock at $3 billion. Meta retains 20% and leases the data centers back. That is off-balance-sheet financing at a scale that has drawn direct comparisons to pre-crisis structured vehicles. And in early 2025, someone issued the first-ever GPU ABS. A $1.1 billion deal securitizing the AI chips themselves, with AAA-rated notes. We are now securitizing the graphics cards. If that sentence does not sound insane to you, I need you to read it one more time.

The Ratings Are Written in Pencil

Now let me tell you about the ratings, because this is where it stops being abstract and starts being genuinely frightening. S&P published its first data center ABS criteria on June 13, 2024. The first deal was issued in February 2018. That is six years of bonds getting rated before anyone published formal rules for how to rate them. Moody's published theirs in February 2025. Fitch had zero outstanding data center ratings as of July 2025 despite tens of billions already on the market. And in February 2026, Moody's assigned the first-ever AAA rating from a Big Three agency to a data center deal. AAA. The highest rating you can give. For a seven-year-old asset class with zero loss history, backed by technology that becomes obsolete every two to three years, in an industry whose demand is driven by five companies all making the same bet. AAA.

The asset class has seven years of performance data and zero credit losses. That sounds reassuring until you think about it for more than five seconds. Of course there are zero losses. Data center demand has done nothing but grow for the entire period these securities have existed. The models are calibrated to a world where the line only goes up. This is the exact same problem that killed mortgage-backed securities. The Gaussian copula model used to price CDOs was calibrated to less than a decade of credit default swap data during which housing prices only rose nationally. The model literally could not imagine a world where prices fell everywhere at once, because that world did not exist in the training data. And here we are again with rating models calibrated to seven years of monotonic growth, being asked to assess instruments that will need to perform in a world where demand could contract.

The Most Obvious Correlation Risk in Financial History

And the correlation risk is arguably worse here than in MBS, which is a sentence I did not think I would ever write. The fatal assumption in mortgage-backed securities was that individual defaults were approximately independent. They were not. Every mortgage depended on the same factors: housing prices, interest rates, employment. When those factors moved, everything moved. But at least the illusion of independence was somewhat understandable. There were millions of individual borrowers. You could squint at it.

Data center bonds are backed by five companies. Five. Amazon, Microsoft, Google, Meta, and Oracle account for 73% of all data center leasing. They are all making the exact same bet on the exact same technology at the exact same time. There is no hidden correlation to discover. The concentration is sitting right there in the prospectus. Everyone can see it. Everyone is choosing to believe it does not matter because these are investment-grade tenants with hundreds of billions in cash.

And that is true. Amazon is not a subprime borrower. Microsoft is not going to miss a lease payment because they forgot to check the mail. But that is not the risk. The risk is what those five companies are doing to their own balance sheets right now to fund the AI bet, and whether they can keep doing it.

"Investment Grade" Is Doing a Lot of Heavy Lifting Right Now

The top five hyperscalers are projected to spend $600 to $690 billion on infrastructure in 2026. That is a 36% increase over 2025. Capital intensity has hit 45 to 57% of revenue, levels that look more like a utility company than a tech company. And here is the part that should really concern you if your retirement is backed by bonds whose payments depend on these companies continuing to spend: they can no longer fund it from cash flow. Aggregate capex, after buybacks and dividends, now exceeds projected cash flows for the big five. Morgan Stanley expects hyperscaler borrowing to top $400 billion this year alone, more than double 2025. JP Morgan projects $1.5 trillion in total tech sector debt issuance over the coming years to finance AI infrastructure. The key players are expected to spend about 90% of their operating cash flow on capex in 2026. These are the investment-grade tenants whose lease payments back the bonds in your retirement account, and they have transformed themselves from cash-rich software companies into leveraged infrastructure plays in about 18 months.

Against that $600 billion in spending, AI-related services are expected to generate about $25 billion in revenue in 2025. That is a 24-to-1 ratio of infrastructure spending to actual AI revenue. Bain calculated that AI needs $2 trillion in annual revenue by decade's end to justify current investment. Best-case forecasts say $1.2 trillion. That is an $800 billion gap. Sequoia's David Cahn calculates a $500 to $600 billion annual revenue gap between AI infrastructure investment and what the AI economy actually produces. Only 25% of AI initiatives have delivered expected ROI. Fewer than 20% have been scaled across enterprises. MIT's NANDA initiative studied 300 public AI deployments and found that only 5% of generative AI pilot programs achieve rapid revenue acceleration. Five percent. The vast majority stall, delivering little to no measurable impact on the income statement.

The Center Cannot Hold

And the anchor tenant of the entire ecosystem just told us something important about how this is going. This week, The Information reported that OpenAI's $500 billion Stargate project, the one Trump announced at the White House in January 2025, has not staffed up and is not developing any of OpenAI's data centers. Thirteen months. Zero employees hired. The partners cannot agree on anything. OpenAI tried to build its own data centers independently and could not get financing because lenders would not back a company that might run out of cash by mid-2027. OpenAI also revised its cash burn projections upward by $111 billion this week, now expecting to burn $665 billion by 2030. Gross margins at 33%, down from their 46% target. They spend $1.69 for every revenue dollar. Sam Altman quietly walked down total infrastructure spending from $1.4 trillion to about $600 billion. Nobody in the press asked about the 57% reduction.

This is the company whose demand for compute is supposed to underwrite the entire data center expansion. This is the tenant whose growth backstops the lease revenue that flows through the SPVs into the bonds that insurance companies and pension funds are buying for yield. And they cannot build their flagship project, they are hemorrhaging cash faster than any company in history, and they might be insolvent in 18 months.

CoreWeave is the single best illustration of how every thread ties together. They went public in March 2025 and carry $14.2 billion in total debt, some at 15% interest. Seventy-one percent of their revenue comes from a single customer, widely understood to be Microsoft. They lost $863 million in 2024. The debt is structured through bankruptcy-remote SPVs collateralized by NVIDIA GPUs and customer contracts. Kerrisdale Capital's short thesis argues that on a look-through basis, the revenue is overwhelmingly dependent on one underlying customer: OpenAI, since Microsoft's reserved capacity largely supports OpenAI workloads. NVIDIA guarantees to absorb unused CoreWeave capacity through 2032. OpenAI entered deals with CoreWeave worth tens of billions, renting chip capacity in exchange for CoreWeave stock.

Read that chain one more time and really think about it. NVIDIA invests in CoreWeave. CoreWeave buys NVIDIA GPUs. CoreWeave rents compute to Microsoft. Microsoft uses it for OpenAI. OpenAI, the company that just revised its cash burn upward by $111 billion and might run out of money by mid-2027, is the end-demand customer that the entire revenue chain depends on. The GPUs collateralizing the debt depreciate by 50% or more within two to three years. And D.A. Davidson analyst Gil Luria compared the SPV structures to "Enron-era structures" and warned that if the AI market even steadied in its growth, "pretty quickly we will have over-built capacity, and the debt will be worthless."

The Collateral Is Rotting

Let me talk about the technology underneath these bonds, because there is a layer here that has no parallel in mortgage-backed securities. A house is a house. It has inherent utility. If the mortgage defaults, the bank forecloses and sells it. A data center full of GPUs that were cutting-edge in 2024 is not cutting-edge in 2027. NVIDIA releases new architectures annually. Hopper in 2022. Blackwell in 2024. Rubin in 2026. Frontier economic utility lasts two to three years. Physical lifespan under high utilization may be one to three years. The bonds have anticipated repayment dates of five years and legal final maturities of 24 to 30 years. The collateral has an economic life of two to three years. Just sit with that for a second.

Power density requirements are changing so fast that facilities built three years ago are approaching structural limits. Traditional data centers ran 5 to 10 kilowatts per rack. Current AI workloads need 40 to 130. Projected 2027: 250 to 600 kilowatts per rack. Air-cooled facilities physically cannot support the next generation. Retrofitting for liquid cooling costs $1 to $5 million per megawatt. Man Group's assessment: "Depreciation schedules are too long, collateral values in default are illusory, and cash flow assumptions are fragile."

You Are Already in This Trade

Now here is who is actually buying these bonds and why you should care. Insurance companies are the dominant purchasers. They buy senior tranches for the yield: 145 to 180 basis points over Treasuries, which is 65 to 100 basis points more than comparably rated corporate bonds. Pension funds are the second largest buyer. Canada Pension Plan has committed billions across multiple data center investments. Blue Owl sold $1.4 billion in private credit fund assets including data center exposure to North American public pension and insurance investors in February 2026. CalPERS and CalSTRS have indirect exposure through their private equity commitments to Blackstone and KKR. Sovereign wealth funds are in at enormous scale: Kuwait Investment Authority and Temasek anchored BlackRock's AI Infrastructure Partnership, which bought Aligned Data Centers for roughly $40 billion. Abu Dhabi's MGX is targeting $100 billion for AI infrastructure.

The direct securitized ABS exposure is still mostly institutional for now. Rule 144A. Not available to retail. But the market is "pondering SEC registration," which would blow the doors open. And you already have exposure whether you know it or not. If you hold an S&P 500 index fund, you hold Equinix and Digital Realty, both S&P 500 components. Data center REIT bonds are in the Bloomberg US Aggregate Bond Index, which means they are in iShares AGG, Vanguard BND, and virtually every target-date retirement fund in existence.

Nobody Is Watching

And nobody is watching. No U.S. regulator has issued guidance specific to data center securitization. No congressional hearing. No SEC framework. The OCC actually rescinded its leveraged lending guidance in 2025, making it easier for banks to pile in. More than 50 commercial banks are now active in data center lending. Greenfield construction volume went from roughly $200 million a year before 2020 to $30 billion in 2025. JPMorgan projects $300 billion per year in AI and data center-related debt deals. And not a single congressional hearing.

The Snake Finishes Eating Itself

Now pull it all together, because these pieces only work as a system. NVIDIA invests in AI startups. Those startups spend the money on NVIDIA chips and data center compute. Data center operators use those revenue contracts to securitize their debt into bonds. Banks sell the bonds to insurance companies and pension funds. The rating agencies rate them based on seven years of data during which demand only grew. The institutional investors buy them for the yield. The capital flows back to the hyperscalers who lease the data centers, who are now borrowing $400 billion a year to fund capex they cannot cover from cash flow, who invest in the AI startups, who spend the money on compute. At every node, someone is booking revenue. At every node, someone is claiming growth. And at every node, the underlying economic activity is the same pool of money making the same trip around the same circle. Except now, at the end of the chain, the risk has been packaged, tranched, rated, and distributed to the people who are least equipped to understand it and least able to absorb the loss.

And the technology those bonds are ultimately betting on? Ninety-five percent of enterprise AI pilots fail to deliver measurable ROI. Eighty-five percent of AI startups will be dead within three years. The anchor company of the entire ecosystem might run out of cash by mid-2027 and just revised its spending projections down by 57% from what the CEO was saying publicly months ago. The $500 billion flagship infrastructure project has not hired a single person. The hyperscalers backing the leases are spending 90% of their cash flow on capex and borrowing the rest. And the revenue gap between what AI infrastructure costs and what AI actually earns is somewhere between $500 billion and $800 billion per year.

The bonds are priced as if none of this is true. The ratings assume the line keeps going up. The buyers are chasing yield in a zero-loss asset class that has never been tested under stress. And the regulators have not even looked.

If you are a leader trying to make AI decisions for your organization right now, this is the information environment you are operating in. The urgency you feel is partly manufactured by companies that need you to spend to justify their valuations, which justify their debt, which justifies the bonds in your pension fund. The signal-to-noise ratio is zero. The technology is real. AI works. We see it every day. But it works in specific, measurable, often unsexy ways: automating a claims workflow, reducing inventory waste, cutting review cycles. It does not work the way the press releases suggest, the way the fundraising rounds imply, or the way the bond prospectuses assume. The gap between what AI can do for your organization and what the financial machinery says it is worth has never been wider. And the entire system, from the VCs to the startups to the hyperscalers to the banks to the rating agencies to the bond buyers, has a financial incentive to keep that gap invisible.