Is Your Pension Funding the AI Bubble?

In December 2025, the Financial Times published an analysis showing that tech companies have moved more than $120 billion of data centre spending off their balance sheets using special purpose vehicles. Meta, xAI, Oracle and CoreWeave have all done it. The money comes from PIMCO, BlackRock, Apollo, Blue Owl Capital and JPMorgan. Those names matter, because those are the firms that manage money on behalf of pension funds, insurance companies and retirement savers.

How the Structure Works

I want to explain how this works, because the structure is important.

Meta needed $30 billion to build its Hyperion data centre in Louisiana. Borrowing that directly would wreck its balance sheet and damage its credit rating. So instead, Meta and Blue Owl Capital created a special purpose vehicle called "Beignet Investor". PIMCO put in $18 billion in loans. BlackRock contributed $3 billion. Apollo and others filled the rest. Meta keeps a 20% equity stake and leases the data centre back from the SPV under a long-term triple-net lease. The debt doesn't appear on Meta's books. Two weeks later, Meta sold $30 billion in corporate bonds as if the first $30 billion didn't exist.

Paul Kedrosky described it plainly: the debt belongs to "those guys over there, that SPV. Not us." He called it accounting trickery, which it is. Meta retains shared control, uses the data centre, and nothing happens without their say-so, but for accounting purposes the debt is someone else's problem. That someone else, increasingly, is your pension fund.

How the Money Trail Reaches You

Here is how the money trail reaches you.

Apollo's insurance arm, Athene, anchors much of this financing. The explicit plan, as described in deal analyses, is for Apollo to then syndicate portions of the debt to other institutional investors, including pension funds and insurance companies. This is not an accident or a side effect. It is the business model. The SPV structure is specifically designed to create paper that looks like safe, long-duration, investment-grade debt, exactly the kind of thing pension fund managers are supposed to buy.

It gets an A+ credit rating because Meta's creditworthiness props up the structure. Meta has provided a residual value guarantee, promising to cover investors if the data centre's value drops below a threshold at the end of the lease. S&P called this guarantee "the linchpin" that earned the debt its rating. So the bond looks rock solid. But what it actually represents is a bet that AI demand will grow without interruption for the next 15 to 20 years, that the hardware inside the building won't become obsolete (it will, roughly every three years), and that Meta will honour a guarantee that could cost billions if things go wrong.

And Meta's deal is just one example. Oracle has structured over $60 billion across multiple SPV deals for data centres in Texas, Wisconsin and New Mexico. xAI closed a $20 billion round in January 2026, structured partly through an SPV that purchased Nvidia GPUs and leased them back to xAI, with the debt secured by the hardware rather than xAI's corporate assets. CoreWeave, listed by the FT among the $120 billion total, has created SPVs tied to billions in compute contracts, though exact figures remain private. The $120 billion the FT documented is likely an undercount, since many of these deals are private and structured specifically to avoid public disclosure.

The UK Exposure

Now here is the UK-specific bit, and this is where I got uncomfortable.

The Bank of England's December 2025 Financial Stability Report flagged this directly. It warned that deeper links between AI firms and credit markets, and increasing interconnections between those firms, mean that a correction could create financial stability risks. It cited estimates that AI infrastructure spending could reach $5 trillion over the coming years (a figure drawn from McKinsey), with a significant portion financed through debt. It flagged compressed credit spreads, weak underwriting standards, opacity and complex structures as specific concerns. This is the Bank of England, not a blog. They are worried.

Mercer, one of the UK's largest pension advisors, published guidance for defined benefit schemes warning that AI-related bond issuance has become a concentration risk in pension portfolios. Their analysis cited JP Morgan's estimate that AI-linked companies will account for 14 per cent of its investment grade index in 2026. That is the index that passive bond portfolios and new pension mandates automatically buy into. Mercer explicitly compared the situation to the Technology Media Telecom boom of the early 2000s, where companies laden with debt participated in a bidding war for mobile licences and then underperformed expectations.

Some UK funds have already started reacting. Standard Life has been reducing its US equity allocation over concerns about AI concentration and market froth. NEST, the state-backed pension scheme managing tens of billions of pounds, is looking at ways to reduce how much younger savers depend on a single high-growth story. This matters because UK defined contribution savers who are 30 years from retirement typically have 70 to 80 per cent of their pension in global equities, heavily weighted towards US big tech. Some have 100 per cent.

The Local Government Pension Scheme, which manages roughly £392 billion on behalf of 6.7 million members across England and Wales, is in the middle of being consolidated into mega-funds. Private sector DB schemes already have nearly 69 per cent of their assets in bonds. When the investment grade bond index is 14 per cent AI-linked, that exposure is not trivial.

The Opacity Problem

But here is the really opaque part: nobody can tell you exactly how much UK pension money is sitting in AI infrastructure SPVs. The whole point of the SPV structure is that the debt doesn't appear on the tech company's balance sheet. Once it gets syndicated to institutional investors, sliced into tranches, and rated investment grade, it becomes very difficult to trace. Henry Tapper at AgeWage has been asking this exact question since Christmas and hasn't got a satisfactory answer. He noted that US private capital specialists are increasingly buying into and buying out UK pension funds in exchange for bonds that look "more and more esoteric." He compared the situation directly to the financial crisis.

Aviva Investors noted in December 2025 that as cash flows and creditworthiness become more interconnected across legally separate SPV entities, investors face indirect exposures that are hard to evaluate. They warned that lower-rated issuers may have little choice but to rely heavily on debt, and that if free cash flow alone cannot cover required investment, the risk concentrates rapidly.

The Telecom Parallel

I keep coming back to the telecom parallel because the numbers are eerily similar. In the late 1990s, telecom companies laid 80 million miles of fibre optic cable. Four years after the bubble burst, by industry post-mortem estimates, 85 to 95 per cent of it was unused, earning the nickname "dark fibre." The debt remained. The cable didn't generate revenue. Investors lost billions. The current AI data centre buildout is already spending more than the telecom boom did when adjusted for scale, driven by the same hyperbolic projections of future usage. As one analysis put it: "Every cycle offers a reason why the future will grow fast enough to justify today's debt. Leverage is not innovation; it is acceleration."

And now the securitisation has started. The FT reported in December 2025 that Wall Street has begun securitising AI debt, pooling loans from multiple SPVs and selling them as asset-backed securities to a broader range of investors, including asset managers and pension funds. The deals are currently in the single-digit billions. But they said that about mortgage-backed securities once too.

What This Means

I want to be careful here. I am not saying this is 2008 again. The asset class is different. The credit quality of the tenants (Meta, Google, Microsoft) is genuinely strong in a way that subprime mortgage borrowers never were. The Bank of England is watching. Some pension managers are already adjusting their exposure. These are meaningful differences.

But the structural parallels are hard to ignore. Off-balance-sheet vehicles hiding leverage. Debt being securitised and dispersed so widely that nobody knows who holds the risk. Credit ratings propped up by guarantees that only work if the underlying assumptions hold. An entire financing ecosystem that depends on a single thesis (AI demand will keep growing) being correct. And ordinary people's retirement savings sitting somewhere in the chain, exposed to risks they almost certainly don't know about.

The Bank of England said it plainly in December: credit spreads remain compressed, underwriting standards are weak, structures are complex, and opacity is a problem. They flagged two recent high-profile US corporate defaults as evidence that the vulnerabilities they'd been warning about were starting to materialise.

Nobody I've spoken to or read can tell me exactly how much of the average UK pension is exposed to AI infrastructure debt. Not Mercer. Not the Bank of England. Not the pension funds themselves. And that, more than any single number, is the thing that should worry you. When you can't measure the exposure, you can't manage it. When you can't manage it, you find out the hard way.

If you're a UK pension saver, you might want to ask your provider a simple question: what is my fund's exposure to AI-linked corporate bonds and private credit? If they can't give you a straight answer, that tells you something too.