AI debt migrates from public bond tables into private credit

Two true things cannot hold at once much longer. The biggest corporate borrowing surge in a generation is being priced every day on public screens, where Oracle's wider spreads and Meta's seven-percent data center prints are news. And the fastest-growing slice of that same borrowing is moving into private placements and direct-lending deals where no screen exists, no mark is published, and nobody outside the room knows the terms until something breaks. The migration is deliberate. Everyone involved has a reason to prefer it that way.

Start with the scale. Morgan Stanley expects global AI-related debt issuance to reach nearly $570 billion in 2026, with roughly $236 billion raised by May 31, about four times the pace of a year earlier (Reuters, July 2026). Private bond issuance alone hit about $81 billion through May, the strongest start since records began a decade ago, with Bloomberg reporting that life insurers hunting long-dated assets to match annuity liabilities are the buyers (Bloomberg citing Private Placement Monitor, July 2026). The Bank for International Settlements counted more than $40 billion of private credit loans to AI-related companies in 2025, up from roughly $3 billion fifteen years ago (BIS Bulletin No. 120, June 2026).

Now the actors. The hyperscalers, Alphabet, Amazon, Microsoft and Meta, want to spend more than their cash flow allows; they plan around $700 billion of capital spending this year while keeping investment-grade ratings intact (Reuters, July 2026), so project-level borrowing parked inside special-purpose vehicles lets them buy the machines without wearing all the debt on the front page of their balance sheets. The private credit managers want fee income and long-duration assets to sell to pensions and insurers, and a private deal negotiated one-on-one pays better than a public bond auction. The life insurers want thirty-year cash flows against thirty-year promises. A leased data center looks like a utility if you squint. Moody's warned in July that six major hyperscalers including Alphabet, Amazon and Microsoft are accumulating enough infrastructure debt to threaten their credit quality (Forbes, Jul 23).

The trigger this month was IREN. A former bitcoin miner turned AI cloud builder closed a $3.65 billion investment-grade GPU financing facility backing its Microsoft contract, including a $2.10 billion United States private placement priced at SOFR plus 2.13 percent, arranged by Goldman Sachs and J.P. Morgan with insurers and asset managers taking the paper (IREN company announcement, Jun 1). Read that structure again: a claim secured not on Amazon or Microsoft but on a stack of graphics processors whose resale value decays faster than the loan matures. PIMCO went further, reportedly negotiating with Bank of America to provide roughly $14 billion of debt for an Oracle data center in Michigan (Bloomberg, August 2026). The largest names in asset management are now the lenders of record for buildings they would never underwrite as real estate.

Underneath the trigger sits the slow pressure: the arithmetic of depreciation versus duration. A data center shell lasts decades; an accelerator is obsolete in a few years; the debt is written for twenty. That mismatch was survivable when the tenant was a triple-A cloud giant signing a fifteen-year lease. It becomes dangerous as financing migrates down to second-tier developers and neocloud providers whose repayment depends entirely on one lease from one customer. The BIS calls the resulting web "circular financing": chipmakers and hyperscalers invest in the labs and clouds that then commit to buying chips and computing from the same firms that funded them, with terms poorly disclosed enough that the same asset may be pledged twice (BIS Bulletin No. 120, June 2026).

History offers one clean analogue. In the late 1990s, telecom carriers borrowed hundreds of billions against fiber networks whose capacity outran demand for years; WorldCom hid the gap between booked revenue and reality off the income statement until the whole sector's debt repriced at once, and the visible bond market took the losses because the private structures were too small then to absorb them. What differs this time is that the private market has grown large enough to be the shock absorber itself. What rhymes is the confidence that demand will arrive before the debt comes due.

The counter-example argues the other way, and honest readers should weigh it. The American private placement market has financed railroads, mines and utilities for a century with default rates historically comparable to or better than public high-grade bonds, because negotiated deals carry tighter covenants than anything sold publicly. If the tenants keep paying, this is simply utilities with better yields. That is the whole bet, embedded in every deal: AI revenue grows into the obligations before refinancing windows test them.

Walk the consequences forward. First order: spreads on visible AI paper, Oracle's and Meta's, widen as supply floods the public market, which pushes marginal issuers toward the private room where pricing is opaque. Second order: insurers and pension funds end up holding claims on depreciating chips and single-tenant buildings, marked quarterly by the very managers who originated them rather than daily by a market. Third order cuts deepest. If AI monetization disappoints even modestly, the losses surface first inside insurance general accounts and retirement money, exactly the places where marks move slowly and selling is impossible. Michael Burry put the catalyst plainly: higher rates for longer could force the reckoning for private credit firms that rushed to finance the buildout (Business Insider, July 2026).

The regulators have noticed before the markets have. The European Central Bank now classifies the opacity of AI financing through private credit as a financial stability risk (ECB, August 2026), and AXA has begun reining in its AI data-center exposures and tightening its stance on private credit generally (Insurance Business Magazine, August 2026). When the sellers step back ahead of the buyers' remorse, the pricing burden shifts to whoever remains. Who remains is mostly American annuity writers reaching for yield.

The observable sequence is short. Confirm the read if private placement issuance keeps setting records into the second half while public AI bond deals start failing or pricing wide, and if any insurer discloses a data-center markdown. Break the read if hyperscaler capex guidance for 2027 gets cut sharply, forcing the giants back onto their own balance sheets, or if AI revenues grow fast enough that the leases service the debt without refinancing, in which case the opacity never mattered.

Who pays, in the end? Not the founders, whose vehicles are bankruptcy-remote by design. Not the rating agencies, who grade only what they can see. The payer is the annuitant in Ohio whose insurer needed thirty-year assets and bought a building in Michigan full of chips that stop mattering in five, on terms nobody published. The market did not lose sight of this debt. It sold the sight deliberately.