The paper has to go somewhere

When a syndicate of loans against a New Mexico data centre traded at ninety cents this week, the interesting question was not the price. It was who was holding the other end. The AI build-out is financed by a relay, and each leg hands the asset to a different kind of investor with a different tolerance for being unable to sell. The last leg is the one nobody writes about.

The Structured Finance Association set the sequence out plainly in July, in a note by its head of research, Elen Callahan. Development begins with construction loans, project finance, private credit or corporate debt, whose investors carry construction, procurement and power availability risk. Once a facility is complete, leased and generating predictable cash flow, it may refinance into asset-backed or commercial-mortgage-backed securities, stay in the corporate bond market, take further private financing, or return to the banks. The risk does not disappear at each handover. It changes owner, and the new owner is further from the site and less able to do anything about it.

The scale of the last leg has changed very fast. Citing Barclays research, the note puts outstanding data-centre securitisation at $4bn in 2020 and $61bn year-to-date in 2026. Data-centre ABS now accounts for roughly 12 per cent of the esoteric ABS market, up from 3 per cent, and data-centre CMBS for about 6 per cent of the single-asset, single-borrower market, up from nothing. Barclays projects $180bn outstanding by end-2028. Morgan Stanley's framing of the gap: about $2.9tn of global data-centre capital expenditure through 2028, of which roughly $1.4tn comes from hyperscaler cash flow and $1.5tn must be financed externally, with securitisation able to take perhaps $150bn.

Average issuance, though, is about $600m for data-centre ABS and $1.2bn for CMBS, and the largest campuses cost many multiples of that. Securitisation is not where the big projects get funded; it is where they get refinanced once somebody else has taken the build risk. The front of the chain stays private by arithmetic, not by preference.

The two structures behave differently at maturity, and that difference is the first mismatch. ABS deals are typically master trusts with anticipated repayment dates — soft deadlines that penalise the issuer for not refinancing, and so create pressure to deleverage. CMBS deals more commonly rely on balloon maturities with little scheduled amortisation, leaving investors exposed to refinancing conditions and property values on a single date years away. A balloon is a bet that a market will exist when the note comes due. In a sector whose assets are seven years old and whose tenants are three, that is a substantial bet.

The second mismatch is in the collateral itself. KBRA looked at data-centre leases in March and concluded that they are not commercial real estate leases in any familiar sense: "power, not space, drives economics", and absolute triple-net structures push operating and capital cost responsibility onto the tenant. Its warning concerned the other side of that bargain: tenants have secured expanded rights of termination, contraction and assignment, which the firm said may reduce cash flow predictability, particularly where a single tenant dominates an asset. The note is long. The lease can be shortened. The tenant can hand it to someone else.

Between the construction loan and the bond sits the part that has grown fastest and discloses least. The Federal Reserve's May financial stability report puts private credit at about $1.4tn, roughly 10 per cent of the total debt of American non-financial corporations and about a third of below-investment-grade debt excluding bank loans. Banks did not withdraw from this market; they moved up a level. Bank credit commitments to other financial entities reached $2.6tn in the fourth quarter of 2025, with private equity, business development companies and private credit vehicles the largest single category and the fastest growth in commitments to collateralised loan obligations and asset-backed vehicles. A bank that will not write a construction loan against a single-tenant campus will happily lend against a portfolio of funds that did.

The third mismatch began to bite this year, and the Fed devoted a box to it. Private credit was historically funded by closed-end institutional vehicles locked up for seven to ten years with no redemption right at all. Increasingly it is funded by retail money in semi-liquid structures: perpetual-life BDCs with $306bn of gross assets and $161bn net, which generally cap redemptions at 5 per cent of net asset value a quarter, and interval funds with $119bn and $80bn, which must accept at least 5 per cent. Together they hold $425bn gross and $241bn net, about a fifth of the sector's net assets under management.

Redemption requests rose in the fourth quarter of 2025 and accelerated in the first quarter of 2026, when a number of funds received requests well above 5 per cent of net asset value; most managers capped them. Accepted redemptions modestly exceeded new inflows in the first quarter — the first time that has happened since these vehicles were created. The Fed's explanation of the sentiment shift is worth quoting for its irony: high-profile corporate defaults, and "concerns about the potential for advances in AI to disrupt industries, particularly software". Software had become the largest sector in private credit portfolios. The same technology that is generating the new collateral is impairing the old.

Which brings the chain to its end. Somebody has to own long-dated, illiquid, often privately rated paper without needing to sell it, and the structural answer in American finance is the life insurer. Life insurance companies held $11.17tn of assets at the end of 2025, growing 8.2 per cent on the year. Leverage at the largest remained "well into the upper quartile of its historical distribution". Illiquid assets — other ABS, commercial real estate loans, alternative investments and illiquid corporate debt — ran at about 37 per cent of life insurer assets in 2024, against 14 per cent for property and casualty. The Board's own sentence is unambiguous: life insurers "have also steadily increased their investments in risky and illiquid assets over the past decade, and their activity has contributed to the expansion of private credit".

No regulator publishes an insurer-level exposure to data-centre paper. What can be said is structural: the buyer whose liabilities run for thirty years, whose accounting does not mark to market and whose capital treatment favours a rating over a price is the natural home for a twenty-year note secured on one campus leased to one tenant. That is the function insurers perform, but it moves the risk from an institution reporting quarterly to one reporting annually, on assets valued by models.

Stack the mismatches and there are four. A balloon maturity assumes a refinancing market years out. The lease is shorter and more cancellable than the note, and the equipment inside depreciates faster than either. The funds holding the middle of the chain owe quarterly liquidity against ten-year loans. And the power infrastructure built for a single campus — dedicated substations, transmission upgrades, sometimes generation — has no alternative user if the campus is not needed. The Structured Finance Association notes that stranded-cost securitisation may end up financing the recovery of utility investment. That is a remarkable sentence to find in a primer: it contemplates securitising the failure.

None of this is a crisis, and it would be careless to write it as one. Spreads are wide enough to be doing work, the deals are backed by stabilised assets with real cash flow, and the Fed judges redemption pressure manageable. But a chain lengthens in good conditions, and the test of one is never the average link. It is what happens where somebody has to sell — which, on the present structure, is a semi-liquid fund holding loans it cannot move, to meet redemptions it has already capped once.