The Refinancing Clock Inside the AI Build
Data-centre debt with legal maturities in the 2050s is being planned around repayment within five years. What happens at that earlier date is decided by the market for replacement capital, not by the building.
By the team at Banyantree Investment Group
This essay is adapted from analysis first shared in our Monthly Investment Letter of 2 March 2026.
Two dates appear in the documents of representative recent data-centre securitisations, and they disagree by a generation.
The legal maturity, the date the debt must finally be repaid, sits twenty-five or thirty years out. The anticipated repayment date, the date the transaction is actually planned around, arrives in year five. Recent deals from Aligned and Compass carry exactly this shape: no scheduled annual amortisation, a five-year anticipated repayment date, and a legal maturity a quarter of a century away.¹ The building is engineered to earn for decades. Its financing expects to be back in the market within a handful of years, on terms sufficient to repay the existing notes.
The anticipated repayment date is a soft trigger, which is why it's easy to wave past. Missing it isn't a default. The notes stay outstanding, the servers keep running, and the lawyers can point to a maturity in the 2050s. But the economics change on that day. Depending on the transaction, additional interest may accrue, excess cash is trapped and swept towards principal, and distributions to the sponsor stop. The owner's capital is still in the building; it has just stopped coming out. An asset can avoid legal default and still fail the financial plan written around it.
That's the distinction we keep coming back to. Legal maturity tells creditors when the debt must be repaid. The anticipated repayment date tells you when the transaction expects the market to cooperate.
So the same data centre carries two investment cases at once. The operating case is about the building: its tenant, its power, and the demand for computing. The financing case is about the terms on which somebody will lend against those things in five years' time. A property can stay occupied and strategically important while falling short of the return, the valuation, or the debt capacity the original transaction assumed. The label AI infrastructure obscures this by placing different claims, on different balance sheets, under one name. A hyperscaler bond is a claim on the whole company. A project loan leans on one site, its construction budget, and its lease-up. A securitisation is commonly a claim on a bankruptcy-remote vehicle holding a handful of properties and their leases, with a sponsor's equity underneath and, often, one of the strongest companies in the world paying rent above. Those parties are connected without being interchangeable, and the difference decides who wears a shortfall.
S&P Global projects the five largest US hyperscalers will spend about US$600bn on capital works in 2026, 38% more than in 2025, a year in which spending had already risen 68%. Technology's share of global non-financial corporate bond issuance jumped from 11.6% to 16.7% over 2025, and data-centre securitisation passed US$30bn for the year, close to three times the 2024 volume.² Those figures describe different layers of one financing system. The US$600bn is the technology companies' own programme, chips and servers and power and buildings, funded against corporate balance sheets that can carry it. The US$30bn is one channel used by the property owners and sponsors building alongside them. The next phase of the AI build will be settled through spreads, covenants, valuations, and refinancing terms as much as through demand for computing, and it will be settled layer by layer.
Take a completed facility with a household-name tenant. Rent is being paid, power is contracted, and the first stage is operating. On those facts the investment looks largely de-risked, and every party that underwrote it was being sensible: the board saw scarce powered land and visible demand, the lender saw a first-rate tenant on a long lease, the rating analysis saw reserves, cash traps, and decades until legal maturity.
At the anticipated repayment date, a new lender asks a different set of questions. How many years remain on the lease? How much of the site is occupied? How concentrated is the rent? What is the property worth at the return buyers now require? What rate does the replacement loan carry, and how much debt will that value support?
The refinancing market can move against that building without the building doing anything wrong. Property buyers can require a higher return on the same stream of rent, and lenders can charge a wider spread over it, which lowers the value assigned to an identical cash flow and lifts the loan-to-value ratio before the tenant has missed a dollar.
Put numbers on it. Suppose the leased space produces US$60m of annual net cash flow. At a 6% required return, that supports the US$1bn value the original financing was written against: US$650m of debt over US$350m of sponsor equity. At 6.7%, the same US$60m is worth about US$900m. A new lender advancing 60% offers US$540m against US$650m of old principal. The replacement loan comes back US$110m short, before fees and reserves, with no change in the building, the tenant, or a single month's rent. Slower lease-up elsewhere on the site would deepen the gap. The gap forms without it. The numbers are illustrative and real transactions will differ in both directions. The mechanism holds either way.
That US$110m is now a capital-allocation question, and it competes. The sponsor can write the cheque, sell the asset, bring in a partner, or accept tighter terms and trapped cash. Whichever it chooses, the same money can't also fund the next development, meet an equity call elsewhere in the portfolio, or be returned to the sponsor's own investors. A repricing of capital has met a near-dated refinancing requirement and turned into an equity call, without an operating failure anywhere in sight.
The overlooked risk sits one level up. Each of these securities is underwritten one property, one tenant, and one sponsor at a time. The refinancing market moves across all of them at once. When required returns rise broadly, debt capacity falls across vehicles owned by sponsors with no corporate relationship to one another; they share no balance sheet, only the conditions producing the shortfalls. For a single owner with several vehicles, one gap is manageable and a portfolio of gaps arriving together is a different kind of decision. Across the market, unrelated owners can face that decision in the same quarter.
And the conditions get set a long way from any data centre. They were already being revised close to home: in February the Reserve Bank lifted the cash rate 25 basis points to 3.85%, with Governor Bullock warning that inflation is likely to remain above target for some time.³ The rate every financing model treats as given is a decision somebody else makes, and revises. Late February reset it from much further away. We had been writing since late last year that a strike on Iran was coming, most likely over a weekend. It came on a Saturday, killed Supreme Leader Ali Khamenei, and drew Iranian missiles onto Israel and countries hosting US forces.⁴ Being right about the event mattered less than it sounds. The strikes were read first as a war story and an oil story, and the refinancing channel runs further back. A sustained rise in oil feeds inflation expectations. Inflation expectations narrow central banks' room to cut. Long yields and credit spreads back up, property buyers demand higher returns, the same rent is worth less, and the replacement loan shrinks. Computing demand doesn't need to weaken anywhere along that chain. A geopolitical event can change the rate the building is valued at, the spread the new lender charges, and the equity cheque the sponsor is asked to write, without touching a single workload.
The strongest objection is that this joins balance sheets that are legally and economically separate, and it deserves its full force. The hyperscaler tenant, the property sponsor, and the bankruptcy-remote borrower are usually three different entities. The borrower owes the debt. The sponsor decides whether to defend its equity. The tenant pays rent under a lease and, unless it has signed a guarantee, owes nothing more. On that reading the US$600bn tells you the scale of the build and nothing about who must fill a US$110m hole in a property vehicle, and the fund or developer actually facing the call may have resources that look nothing like Microsoft's or Amazon's.
The objection is right about the structure, and the argument doesn't need it to be wrong. If the old debt exceeds what a new lender will advance, the vehicle has a gap, and somebody supplies capital on new terms, whether that's the sponsor, an incoming partner, or a buyer. That mechanism runs whether the tenant helps or simply keeps paying rent while the owner is restructured around it. A strong tenant improves the operating case considerably; nobody underneath the lease should mistake it for a guarantee of their own return.
Scale is the harder objection. US$30bn of securitisation against a US$600bn programme looks like a rounding error, and exactly the layer a willing hyperscaler could absorb, since the sites are first-rate and a company can buy the vehicle it anchors. But the comparison runs on the wrong denominator. The investor in a property vehicle is not underwriting the AI build; their outcome turns on whether US$650m can be replaced against that one vehicle's value. The channel can stay small beside the build and still matter greatly to the capital inside it. And that capital is a widening field: by mid-2025 KBRA had counted US$48.7bn of US data-centre securitisation across 88 transactions since the structure first appeared in 2018,⁵ and public estimates point to US$30bn to US$40bn of gross supply in each of the next two years,⁶ each new cohort of five-year structures winding another clock. Gross issuance is not a maturity wall, since some of it refinances older series and terms differ deal by deal. What accumulates is successive cohorts reaching their anticipated repayment dates in successive years, a growing set of claims exposed to the conditions prevailing when each arrives.
And the absorption argument should be taken on its own terms, so suppose the tenant does choose to step in. Buying a well-performing site while its paper rolls cheaply costs a hyperscaler nothing it would notice. The intervention that matters gets priced on the day a site has wobbled and spreads have widened across the board, which is the same day the tenant is funding its own largest-ever programme into that dearer market. A cheque that size is easy to write in calm conditions. Written on that day, it competes with the tenant's own build for capital that has just repriced for everyone. And whether it is written at all remains, as the structure says, a choice.
We've watched this shape before, in a different material. Global Crossing filed for Chapter 11 in January 2002 with US$22.4bn of listed assets, about US$12.4bn of liabilities, and a fibre network linking more than 200 cities in 27 countries.⁷ Internet traffic kept growing straight through the bankruptcy and long after it. The fibre stayed useful. The company's accounting attracted investigations, which keeps it from being a clean parable, but what the episode separated was the infrastructure forecast from the capital built around it. The forecast proved broadly right over twenty years. Global Crossing's capital structure failed long before the demand matured. The analogy is narrow. Demand for AI computing looks far better supported than bandwidth demand did in 1999, and today's tenants are the strongest credits in the world. What carries over is the timetable: an infrastructure forecast can prove broadly right while the first capital structure built on it runs out of time.
None of this is a forecast. The evidence will arrive in transactions rather than in arguments about artificial intelligence, and the distinctions will show up deal by deal. An anticipated repayment date with eight years left on a strong lease is a different animal from one arriving beside a renewal decision, and a diversified vehicle has more room than one leaning on a single building. The first five-year clocks, wound in 2020 and 2021, are striking about now. The earliest are rolling: DataBank's 2021 notes, for one, are expected to be repaid in full as new notes are issued from the same trust.⁸ The cohorts wound since are far larger, and they strike into whatever market exists when they land.
The readings we're watching: securitisations extending past their anticipated repayment dates rather than refinancing; sponsors writing larger equity cheques to clear deals; refinancings that close only after amortisation is added, covenants tighten, or distributions are shut off; and, read separately, tenants stepping in with lease amendments, guarantees, or outright purchases, which would mark the facilities they genuinely can't do without. One difficult deal proves little. A pattern of extensions and top-ups would carry much more weight. The same record could just as easily weaken the concern: paper refinancing on clean terms, values holding as new supply lands, and sponsors supporting older vehicles without delaying newer ones would all say the market is deeper than the worry assumes. Either way, the transactions will speak before the demand debate resolves, and we would rather read them than argue the headline.
Permanent capital can wait out a slow leasing year. A vehicle due back in the debt market next year cannot. Global Crossing's network was still physically useful the day the petition was filed; the financial claim attached to it had run out of time. A data centre built today may still be operating in 2045. Its first capital structure has to make it through 2030.
General information only. Not personal advice. Past performance is not indicative of future performance. Examples are illustrative. This material is intended for wholesale and professional investors.
Notes
1. S&P Global's presale for Aligned Data Centers Issuer LLC Series 2023-2 describes zero scheduled annual amortisation, a 5-year anticipated repayment date, and a 25-year legal maturity. Its presale for Compass Datacenters Issuer III LLC Series 2026-1 describes zero scheduled annual amortisation, a 5-year anticipated repayment date, and a 30-year legal maturity. S&P's February 2026 Stack Infrastructure presale states that its rating did not address post-anticipated-repayment-date additional interest.
2. S&P Global Ratings, 17 February 2026: the top 5 US hyperscalers were projected to spend approximately US$600bn in 2026, 38% more than in 2025, after a 68% rise that year; technology represented 16.7% of 2025 global non-financial corporate bond issuance, up from 11.6% a year earlier; and global data-centre securitisation exceeded US$30bn in 2025, nearly 3 times the 2024 volume.
3. Reserve Bank of Australia monetary policy decision, 3 February 2026: the cash rate was raised 25 basis points to 3.85%. Governor Michele Bullock's remarks per the accompanying statement and press conference of the same date.
4. Reuters reported that US and Israeli strikes beginning 28 February 2026 killed Ali Khamenei and other senior figures, and were followed by Iranian missile attacks on Israel and countries hosting US forces.
5. KBRA, Data Centers: A Comparison of ABS and CMBS Structures, May 2025: US$48.69bn of US data-centre securitisation across 88 transactions since 2018, with ABS structures accounting for 70.8% of that issuance.
6. CRE Finance Council, Data Center E-Primer, February 2026, citing dealer research: projected gross US data-centre securitisation supply of approximately US$30bn to US$40bn in each of 2026 and 2027, with separate estimates of roughly US$130bn of net issuance over 2026 to 2028. Gross supply includes refinancing of existing series and is distinct from net new debt.
7. Global Crossing filed for bankruptcy protection on 28 January 2002. Contemporaneous reporting placed its assets at US$22.4bn and its liabilities at approximately US$12.4bn, with a network connecting more than 200 cities in 27 countries. Subsequent investigations examined its accounting and disclosures.
8. KBRA preliminary ratings, DataBank Issuer LLC and DataBank Co-Issuer LLC Series 2026-1: in conjunction with the new issuance, the Series 2021-1 Class A-2, B, and C notes are expected to be repaid, with all series sharing the same collateral pool.
