Honey, They Might've Shrunk Our Pension

By Kenan Machado, May 7, 2026

Deep in the northeast corner of Louisiana, on 2,200 acres of farmland in a parish most Americans couldn't place on a map, construction crews are building what may become the largest artificial intelligence computing complex in the world.

The facility will eventually draw as much electricity as four million homes. It will house nine data center buildings covering four million square feet. And it will be, in part, financed by the retirement savings of teachers, nurses, transit workers, and civil servants in California, Ontario, and British Columbia — people who almost certainly have no idea.

Pension funds, under pressure to earn returns that outpace inflation, have poured money into a corner of finance called private credit — essentially, loans made by investment firms rather than banks. The firms take that pension money and deploy it in ways that can be hard to trace: into loans to private companies, into complex financing structures for infrastructure projects, and increasingly, into the debt that is funding the AI boom.

Last week, an international financial watchdog published its first major report on the risks building up in this market. It warned of potential "sizeable losses" if AI-related asset values fall sharply. The same week, three major banks were reportedly struggling to find buyers for $38 billion in AI data center construction loans — investments they had been trying to offload for more than six months.

The people who own much of this risk, directly or indirectly, are the ones saving for retirement.

How Retirement Money Ended Up Here

To understand how pension savings reach an AI data center in rural Louisiana, you have to follow the money through several layers.

The most straightforward layer is familiar: pension funds give money to investment managers, who lend it to companies. The loans are typically secured against assets, which means lenders get paid first if anything goes wrong.

Investment firms like Blue Owl Capital, Apollo Global Management, Inc., Ares Management, and BlackRock have built large businesses doing exactly this. Much of the lending is to mundane, stable businesses — auto repair chains, aerospace suppliers, commercial cleaning companies. The kind of businesses whose revenues do not depend on whether the latest AI model wins a chatbot competition.

This type of lending has performed well.

When Blue Owl, one of the largest private credit firms, ran into trouble earlier this year and needed to quickly sell a portfolio of those everyday loans, it found buyers immediately. Three major pension funds, California's California Public Employees' Retirement System or CalPERS, Ontario's Ontario Municipal Employees Retirement System or OMERS, and British Columbia's British Columbia Investment Management Corporation or BCIM, bought $1.4 billion of the loans at nearly full price. The loans were worth what Blue Owl said they were worth.

The second layer is less familiar and considerably more complicated. Some of the same investment firms have also been financing AI infrastructure — and the structures they use to do it are unlike anything in mainstream investing.

Here is the basic idea.

A technology giant like Meta wants to build a $30 billion data center. But it doesn't want that debt showing up on its own balance sheet, where shareholders and analysts would scrutinize it. So instead of borrowing the money directly, it creates a jointly owned company with a private investment firm. That jointly owned company takes out the loans. The technology giant signs a lease to use the building and promises to cover losses if things go wrong. The debt stays off the tech giant's books. Pension money, routed through the investment firm, ends up financing the building.

This is not illegal. It is not even unusual in infrastructure financing. But it is new at this scale, and it is new to AI — a technology whose economic life expectancy is measured not in decades but in processor generations, which now turn over roughly every year.

The Louisiana Deal

In October 2025, Meta and a private investment firm called Blue Owl completed what was, at the time, the largest single private financing deal ever recorded. The deal had a name: Project Beignet, after the fried pastry popular in New Orleans. The numbers were staggering.

The jointly owned company issued $27.3 billion in loans. PIMCO — which manages retirement assets for millions of Americans — took $18 billion of that, roughly two-thirds of the total. BlackRock took another $3 billion. The loans carry a fixed interest rate of 6.581 percent and don't come due until 2049. That is a 23-year commitment.

The building being financed, called the Hyperion campus, will sit in Richland Parish. It is expected to be operational around 2029 and will eventually draw up to five gigawatts of power — enough, at full buildout, for roughly four million homes — from a dedicated network of gas plants, substations, and transmission lines that Entergy Louisiana is currently seeking permission to build.

To get the $27.3 billion in loans rated at investment grade — meaning a quality high enough that pension funds are permitted to buy them — the deal relied on a single promise from Meta. If the data center lost value and the lenders needed to sell it, Meta agreed to cover the gap between whatever it sold for and what was still owed on the loans. That promise, called a residual value guarantee, is what persuaded the rating agency S&P Global to assign the debt an A+ rating, one notch below Meta's own corporate credit rating.

S&P called the guarantee "the linchpin" of the rating. Remove it, and the deal would not qualify as investment grade by any standard measure.

The catch — and it is a significant one — is that Meta only committed to lease the building for four years. The loans run for 23.

That gap is what we at Icarus Asia spent months analyzing. "Remove the guarantee," we concluded, "and this credit is structurally below investment grade by every standard convention."

The Guarantee Is Not Quite What It Sounds Like

When most people hear that Meta has promised to cover losses, they picture a direct, legally simple backstop: if something goes wrong, Meta pays. The reality is more complicated, and the complications matter.

The guarantee does not run directly to the lenders. It runs to an intermediate holding company sitting between Meta and the entity that actually issued the debt.

In plain terms, if Meta's guarantee were ever called upon — say, because Meta decided not to renew its lease and walked away from the building — lenders would not simply send a bill to Meta. They would first have to pursue a claim through that intermediate company, which could open the door to legal disputes, delays, and competing claims from other creditors.

Icarus Asia estimates that a contested guarantee could take 18 to 24 months to resolve through the courts. [Inference; based on comparable legal timelines] During that time, the interest payments on $27.3 billion in loans — roughly $1.8 billion a year — would need to come from reserves held in the deal structure. If those reserves run out before the dispute is settled, the loans could technically be in default even if Meta ultimately honors the guarantee.

The precise size of the guarantee itself has not been made public.

Based on analysis of available documents, Icarus Asia estimates it covers roughly $20 billion of the $27.3 billion in loans — leaving a potential gap of $7 billion to $8 billion that would depend on whatever the building could be sold for. [Icarus Asia estimate; exact amount not confirmed from deal documents]

For a data center in rural Louisiana, that is a harder question than it sounds.

The Technology Problem Nobody Has Fully Priced

The computers inside an AI data center do not age the way a bridge or a building ages. They become obsolete.

NVIDIA, which makes the processors that power most AI systems, now releases a new generation of chips roughly every year. The Blackwell generation arrived around 2024. Rubin followed in 2025. Future generations are expected on a similar schedule. Each new generation is significantly more powerful and efficient than the last, which means each old generation becomes less useful — and less valuable — faster than anyone planned for even five years ago.

The Hyperion campus is being built to the specifications of 2026-era chips. The loans financing it matures in 2049. By then, NVIDIA will have released approximately 20 more chip generations. The computing hardware inside the building — which represents roughly 70 percent of the construction cost — will have been obsolete for years.

The physical structure of the building, the electrical infrastructure, the substation — those things might hold their value. Steel and concrete do not go out of date.

But our calculations show that the physical shell and power infrastructure of a large data center retains roughly 65 percent of its value even in a worst-case scenario. But that only accounts for about 30 percent of what was spent building the place. The remaining 70 percent — the computing hardware — is worth approximately nothing to anyone else once newer chips exist. [Icarus Asia components-based recovery model]

Put those numbers together, and in an exit scenario eight years from now, the building might be worth roughly $5.9 billion in a liquidation. The loans outstanding at that point would be approximately $19 billion. The guarantee would need to cover a gap of $13 billion. That is an amount that may exceed the undisclosed cap on Meta's guarantee. [Icarus Asia estimate; illustrative only]

None of this means the deal will go wrong.

Meta is one of the most profitable companies in the world and has every incentive to keep the building running. PIMCO's analysts understood the structure when they committed $18 billion to it. The A+ rating — despite its dependency on a single guarantee — was not assigned carelessly.

But the rating and the actual price investors accepted tell different stories.

Investment-grade bonds with an A+ rating typically offer interest rates just 0.75 to 0.90 percentage points above what the U.S. government pays to borrow money. The Beignet loans pay 2.25 percentage points above that government rate. Investors demanded that extra yield precisely because they understood the risks the rating doesn't fully capture. The price was honest even when the label wasn't.

When Investors Tried to Get Their Money Out

Not all pension exposure to private credit runs through sprawling Louisiana data centers. Much of it runs through simpler vehicles — pooled investment funds that lend money to private companies and pay investors regular distributions. These funds have been enormously popular, partly because they offered something that felt like a reasonable balance: higher returns than conventional bonds, with the ability to withdraw money periodically.

The "periodically" part turned out to be more conditional than many investors realized.

By February 2026, investors in two Blue Owl funds had requested to withdraw a combined $5.4 billion. In one fund, Blue Owl Technology Income, withdrawal requests covered 40.7 percent of the total money invested. In another, Blue Owl Credit Income, it was 21.9 percent.

The trigger was a misreading of AI's effects.

Investors assumed that the rise of AI tools would hurt the companies Blue Owl was lending to — primarily enterprise software firms, the kind that run payroll systems and manage supply chains. That assumption was largely wrong. Companies that are already embedded deep inside other businesses tend to become harder to displace as AI spreads, not easier. But the fear was enough to set off a run.

Blue Owl responded by activating what the industry calls a gate — a provision, written into the fund's terms, that limits withdrawals to five percent of total assets per quarter. Investors who wanted their money back had to wait. Similar gates had been used by other large private credit managers, including Ares, KKR, and Blackstone, during past periods of stress.

To demonstrate that its loans were actually worth what it said they were worth, Blue Owl quickly sold $1.4 billion of loans from a separate fund to CalPERS, OMERS, and BCIM — at 99.7 cents on the dollar, effectively full price. That sale served as a reality check. If the loans had been impaired, pension funds would not have paid nearly full price for them.

But the episode illustrated something important about the structure of this market.

Pension funds and smaller investors had bought into these funds expecting something closer to a liquid investment. When stress arrived — even stress based on a misreading of the facts — the gates came down. The liquidity was always conditional. In a genuine credit crisis, the 99.7 cent sale might not happen.

The Experts Are Now Worried

On May 6, the Financial Stability Board (FSB) — a body of regulators from across the world's major economies, created after the 2008 financial crisis to watch for systemic risks — published its first comprehensive report on the private credit market. The language was careful. The concern was not.

AI-related deals now account for more than a third of all private credit transactions, the report noted, up from 17 percent in recent years. The market as a whole has grown to between $1.5 trillion and $2 trillion. It has never faced a severe recession.

The report flagged two specific scenarios that could cause serious losses: a significant drop in demand for AI computing power, which would undercut the value of the buildings being financed; and electricity shortfalls, which could prevent the buildings from operating at the scale the financing assumed. Either scenario could produce "sizeable losses," the FSB said — losses that would ripple through pension funds, insurance companies, and other institutions that have lent money into this market.

The same week, Mercer — one of the world's largest pension advisory firms — published guidance for pension fund trustees in Britain warning that AI bond issuance has become a concentration risk. J.P. Morgan estimates that AI-related companies will account for 14 percent of its main investment-grade bond index in 2026, surpassing U.S. banks to become the single largest sector. Pension funds running passive bond strategies — funds that automatically buy whatever is in the index — are absorbing this exposure whether or not their trustees have consciously decided to take it on.

Mercer drew an explicit comparison to the technology debt bubble of the early 2000s. Then, as now, a narrow sector was borrowing enormous sums, flooding the investment-grade bond market, on assumptions about future demand that proved optimistic.

In Washington, Senator Elizabeth Warren and three colleagues sent a letter to Treasury Secretary Scott Bessent in January requesting an investigation into AI infrastructure debt, specifically calling out the private placement structures that keep much of this lending out of public view.

"The mechanics resemble those of pre-2008 mortgage securitization," the senators wrote — not in terms of the underlying assets, but in terms of how risk gets moved from the banks that arrange it to the institutions that ultimately bear it.

And They Are Already Doing It Again in Texas

The Louisiana deal was not a one-off. It was a template.

This week, Bloomberg News and Reuters reported that Meta is working with Morgan Stanley and JPMorgan Chase on a roughly $13 billion financing package for a new data center campus in El Paso, Texas. The deal has a name: Project Sopaipilla, following the same pastry-naming convention as Beignet. The 1,000-acre campus is targeting one gigawatt of capacity by 2028.

The trajectory of Meta's El Paso commitment is worth noting.

In October 2025, when the campus was announced, Meta's committed investment was $1.5 billion. By March 2026, it had raised that figure more than sixfold to over $10 billion. The $13 billion financing package — mostly debt, with a smaller slice of equity — converts that capital spending into borrowed money, using the same basic structure established in Louisiana.

Meta's total capital spending budget for 2026, as disclosed at its first-quarter earnings call in late April, is now between $115 billion and $145 billion. Almost all of it is going toward AI data centers. The scale of that spending makes external borrowing a necessity, not a choice. Funding it from cash flow alone would leave far less money for everything else.

Oracle has run similar deals in Texas, Wisconsin, and New Mexico, structured against more than $60 billion in loans. xAI, the artificial intelligence company founded by Elon Musk, closed a $20 billion financing round in January partly through an entity that purchased chips from NVIDIA and leased them back, with the debt secured against the hardware itself — a bet on technology that depreciates rapidly as a guarantee for long-term loans.

The banks arranging all of this lending are starting to feel the weight of it.

JPMorgan, Morgan Stanley, and others are actively trying to move AI data center debt off their own books and into the hands of institutional investors. One $38 billion Oracle-linked loan package took more than six months to fully distribute. "Out of scale to anything we've thought about, ever," one banker told the Financial Times.

When banks need to move debt off their books, the buyers are typically insurance companies, pension funds, and the investment managers that act on their behalf. The risk doesn't disappear. It relocates.

What Pension Trustees Should Be Asking

For the average worker with a pension, the practical question is simple: how much of my retirement savings is exposed to this, and what happens if it goes wrong?

For most people, that question is genuinely hard to answer.

Pension funds are not required to disclose, in plain terms, how much of their investments touch AI data center debt — directly through project loans, indirectly through bonds, or passively through index funds that automatically absorb new investment-grade issuance. The Financial Stability Board noted explicitly that data gaps make it hard even for regulators to assess exposures.

For pension trustees — the people legally responsible for managing those funds — the FSB report and the Icarus Asia analysis together point toward a specific checklist.

Does the fund know its total exposure to AI infrastructure debt, including any held indirectly through passive bond mandates or through external managers? Most trustees today cannot answer this question, because no one aggregates it.

In any deal secured against an AI data center, does the guarantee from the technology company run directly to the lenders, or does it pass through intermediate entities that could complicate a claim? The difference is not a technicality. In the Beignet structure, it represents billions of dollars of potential recovery gap.

Does the deal hold enough cash in reserve to cover interest payments for at least 18 months if the guarantee is disputed in court? [Icarus Asia estimate of dispute timeline] If not, a legal fight — even one the lenders ultimately win — could push the loans into technical default before it is resolved.

And has anyone modeled what the collateral is actually worth if the technology inside the building becomes obsolete before the loans mature? A flat estimate of how much a data center is worth gets the math badly wrong once you separate the computing hardware, which has almost no resale value after a few chip generations, from the physical structure and power infrastructure, which retain value for decades.

The Bet That Defines the Cycle

PIMCO Chief Executive Officer Emmanuel Roman was at the Milken Institute International 's annual conference in Beverly Hills this week — the gathering where the world's largest investment managers talk about where money is going.

He acknowledged the obsolescence question directly. "There are issues around planned obsolescence of data centers," he told Bloomberg. His answer was the counterparty: "You think of Meta. It has a $1 trillion market cap. So when you plug this into a capital structure model, it's incredibly safe."

He may be right. PIMCO committed $18 billion to that view. Its analysts spent months reviewing the structure before making that commitment. Meta's balance sheet is, by any measure, formidable.

But the argument — essentially, that the borrower is so creditworthy that the specific risks of the thing being financed don't matter much — is an argument that sounds most persuasive right before it stops being true. It is an argument that was made about mortgage bonds in 2006, about telecom debt in 1999, about a number of things that seemed safe until they were not.

The Beignet loans mature in 2049. By then, the AI chip inside the Hyperion campus will have been superseded roughly 20 times. Whether the buildings are still worth $27 billion — or something close to it — depends on whether Meta keeps renewing leases, whether the technology demands that filled those buildings hold up, and whether the guarantee that underpins the entire structure performs the way the rating assumes it will.

Those are not unreasonable bets. They are bets nonetheless. And the people on the other side of them, whether they know it or not, include millions of Americans and Britons saving for retirement.

The author is a former journalist and the Head of Research and Analysis at Icarus Asia, a Hong Kong-based risk and advisory business.

References

Financial Stability Board — Report on Vulnerabilities in Private Credit, May 6, 2026: https://www.fsb.org/2026/05/report-on-vulnerabilities-in-private-credit/
Financial Stability Board — FSB warns on private credit vulnerabilities (press release): https://www.fsb.org/2026/05/fsb-warns-on-private-credit-vulnerabilities/
Financial Stability Board — Full report PDF: https://www.fsb.org/uploads/P060526.pdf
Meta Platforms — Joint venture with Blue Owl Capital to develop Hyperion data center, October 2025: https://about.fb.com/news/2025/10/meta-blue-owl-capital-develop-hyperion-data-center/
PE Insights — Blue Owl and Meta close record $30bn financing for AI data centre expansion in Louisiana: https://pe-insights.com/blue-owl-and-meta-close-record-30bn-financing-for-ai-data-centre-expansion-in-louisiana/
S&P Global — Key data center financing takeaways from PPIF 2026: https://www.spglobal.com/ratings/en/regulatory/article/global-infrastructure-and-project-finance-key-data-center-financing-takeaways-from-ppif-2026-s101668747
Semafor — Blue Owl credit funds face heavy redemption requests, April 2, 2026: https://www.semafor.com/article/04/02/2026/blue-owl-credit-funds-face-heavy-redemption-requests-as-private-credit-jitters-persist
The Guardian — Blue Owl Capital limits withdrawals after investors try to redeem $5.4bn, April 2, 2026: https://www.theguardian.com/business/2026/apr/02/blue-owl-capital-private-credit-investment-limits-withdrawals
Benzinga — PIMCO purchases $400M OBDC bond: https://www.benzinga.com/pressreleases/26/04/ab40264052/pimco-purchases-400m-obdc-bond
LPSC Docket U-37882 — Entergy Louisiana application for certification of generation and transmission resources, March 26, 2026: https://lpsc.louisiana.gov/
ROIC.ai — EY Critical Audit Matter flag on Meta's off-balance-sheet treatment: https://roic.ai/quote/META
Quinn Emanuel — Private credit under stress: emerging litigation risks: https://www.quinnemanuel.com/the-firm/publications/client-alert-private-credit-under-stress-emerging-litigation-risks/
Senator Warren / Senate Banking Committee — Warren, colleagues press FSOC to investigate AI debt bubble: https://www.banking.senate.gov/newsroom/minority/warren-colleagues-press-fsoc-to-launch-probe-into-financial-stability-risks-of-ai-debt-bubble
Mercer — Top considerations for UK defined benefit pension schemes in 2026: https://www.mercer.com/en-gb/insights/pensions/defined-benefit-schemes/top-db-considerations/
From The Prism — Is your pension funding the AI bubble?, February 2026: https://fromtheprism.com/pensions-ai-exposure
Bloomberg — Meta taps Morgan Stanley, JPMorgan for El Paso data center deal (Project Sopaipilla), May 4, 2026: https://www.bloomberg.com/news/articles/2026-05-04/meta-taps-morgan-stanley-jpmorgan-for-el-paso-data-center-deal
Reuters / Business Standard — Meta picks Morgan Stanley, JPMorgan for El Paso data center financing, May 5, 2026: https://www.business-standard.com/amp/world-news/meta-picks-morgan-stanley-jpmorgan-for-el-paso-data-centre-financing-126050500159_1.html
Advisor Perspectives — Meta taps Morgan Stanley, JPMorgan for new data center deal, May 5, 2026: https://www.advisorperspectives.com/articles/2026/05/05/meta-taps-morgan-stanley-jpmorgan-new-deal
Cryptopolitan — Meta $13B Texas data center financing and Q1 2026 capex guidance: https://www.cryptopolitan.com/meta-13b-texas-data-center-financing/
The Next Web — Meta El Paso $13 billion, one of the largest single-site digital infrastructure financings: https://thenextweb.com/news/meta-el-paso-13-billion-data-center-financing
Insurance Journal — The $3 trillion AI data center build-out becomes all-consuming for debt markets, February 2026: https://www.insurancejournal.com/news/international/2026/02/03/856623.htm
Insurance Journal — FSB watchdog flags risks in banks' growing private credit ties, May 2026: https://www.insurancejournal.com/news/international/2026/05/06/868689.htm
Goldman Sachs — Tracking trillions: the assumptions shaping the scale of the AI build-out: https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out
The Decoder — Building AI data centers is becoming a stress test for banks, May 2026: https://the-decoder.com/building-ai-data-centers-is-becoming-a-stress-test-for-banks/
Pensions Age — Institutional investors' private markets allocations hit record high of 12.5%: https://www.pensionsage.com/pa/Private-markets-allocations-hit-record-high-of-12-5.php
BloombergNEF — AI data center build advances at full speed: five things to know: https://about.bnef.com/insights/commodities/ai-data-center-build-advances-at-full-speed-five-things-to-know/
Bank for International Settlements Bulletin No. 120 — Financing the AI boom: from cash flows to debt: https://www.bis.org/publ/bisbull120.pdf
Cleary Gottlieb — Investment-grade private credit: how blue-chip corporates finance multi-billion-dollar projects off balance sheet: https://content.clearygottlieb.com/corporate/capital-solutions-insights/investment-grade-private-credit-how-blue-chip-corporates-finance-multi-billion-dollar-projects-off-balance-sheet/
Institutional Investor — How institutional investors are managing the private credit crisis: https://www.institutionalinvestor.com/article/how-institutional-investors-are-managing-private-credit-crisis
Icarus Asia Private Credit Research — If you build it, will they come?, April 2026 (primary analytical source)

Disclaimer: Items labeled Icarus Asia analysis, Icarus Asia estimate, or Inference represent analytical judgments based on public information and comparable transaction precedents, not primary deal documentation. The exact RVG cap for Project Beignet, the DSRA funding level, and the retrofit capital expenditure allocation mechanism are not confirmed from the Offering Memorandum. Institutional investors should verify these items independently before making investment decisions. This article is for institutional research and educational purposes only and does not constitute investment advice. No material non-public information was used.