Data center CMBS is pricing the wrong risk

The bond market has put a price on obsolescence: 72 basis points, a wide enough gap to say the market knows the collateral is different and a narrow enough one to suggest it has not worked out how different. Triple-A data center CMBS now clears at an average spread of 1.65 points over its floating-rate benchmark, against 0.93 for office, 1.05 for retail and 1.25 for industrial, according to Barclays data cited in coverage of the sector's rise.

The premium rides a market that has grown quickly: roughly $17 billion of data center CMBS issued since the start of last year, more than three times the volume of the preceding two years combined, and now about 8 percent of new commercial property bond deals, according to the same reporting, which draws on Bloomberg. Citigroup expects issuance to climb roughly 50 percent next year, to as much as $20 billion.

What a bond buyer gets for those 165 points is a building whose value turns on what commercial mortgage underwriting spent decades ignoring: power availability and transmission capacity, cooling systems and computing density determine whether a facility stays competitive, and any of them can change after the loan is written. A mortgage on an apartment building is a claim on a box in a location that can be re-leased to almost anyone; a mortgage on a data hall is closer to a claim on one tenant's electrical and thermal specification.

Tenant concentration therefore matters in a way it does not in most property debt, because data centers often depend on a small number of hyperscalers whose identities and lease terms can be closely guarded; the provisions doing the real work—who bears power costs, what minimum capacity the tenant commits to, what happens during downtime—feed straight into cash flow and therefore debt service. A lender without visibility into those terms is underwriting the sector's reputation instead.

Replacement is the harder problem, and it is where conventional CRE instincts fail. A facility engineered around one customer's power and cooling requirements does not hand over cleanly; accommodating a different occupant can require substantial capital, a cost the lender wears through longer downtime and weaker recovery value if the original tenant leaves or the building falls behind the market. Where traditional property debt can tell a re-tenanting story at maturity, this asset class must tell a re-engineering story, and it prices worse.

Then there is the clock, because rapid advances in AI chips raise power and cooling requirements that can make a facility outdated far faster than conventional real estate, and the advances do not wait for the loan to amortize. A mortgage with years left to run against a technology cycle that turns in a fraction of that time is a duration mismatch no spread fully compensates, and it turns refinancing into a question about whether the collateral is still the asset the loan was written against.

The 1.65-point spread is functioning as a technology-risk premium when the risks it actually prices are concentration and residual value, which are not the same trade. It compensates a bondholder for not knowing who the tenant is or what happens when that tenant's needs move; it does not compensate for the possibility that the box itself has no second act, because terminal value is what equities price, not mortgages.

The securitization wave is the debt-market version of a transfer this publication has already flagged on the equity side: data center capital is pushing construction and residual-value risk into public markets, and the CMBS pipeline is simply the second venue where that risk lands. If private portfolios are vulnerable to a markdown when development gets priced like the construction loan it is, the same logic applies to a bond tranche priced like a stabilized mortgage on an asset that is not stabilized in any usual sense. A buyer at 165 points is taking an equity-shaped risk at a debt-shaped return.

The market will likely sort itself rather than reprice wholesale, with contracted power and secured interconnection, leases carrying long committed terms and clean pass-throughs on energy, and cooling designs with headroom holding tighter spreads than speculative shells built to a generic specification — a tiering invisible in an average spread. The average is the least useful number in the sector, because it blends cash flows contractually insulated from the chip cadence with cash flows that are a bet on it.

The refinancing machinery is already being rearranged around exactly this kind of problem. As this publication has reported, CREFC has handed its top job to a new-issue CMBS desk head, with the council's center of gravity shifted from legislative fights to extension and modification work — and the data center loans being written today will mature into that apparatus if demand cools or a tenant's requirements drift. Premium collateral can still clear a wall: the $382.4 million refinancing at 120 Park Avenue landed 14 percent above the loan it replaced, but that was a trophy office tower in a market with a functioning clearing price. There is no comparable clearing market for a purpose-built data hall whose tenant has moved on, and no extension-and-modify playbook that fixes a cooling plant sized for last generation's chips.

The first deal to watch is one with a visible seam: a lease term shorter than the loan term, a power contract without a secured path to interconnection, or a single hyperscaler carrying the entire rent roll. Citigroup's $20 billion forecast assumes investor appetite holds through the pipeline, and if AI demand cools before that supply clears, the tranches tested first will be the ones whose collateral cannot be re-leased to anybody but the tenant it was built for.