I have mapped the supply of AI debt and the structures that keep it outside hyperscaler balance sheets. The next question is when it needs to be refinanced. Most off-balance-sheet data center debt was not raised as 10-year corporate bonds. It was raised as 5-7 year private placements and project-level loans held by private credit funds and insurers. The 2023-25 vintage, the first large wave of AI data center SPVs and JVs, starts to mature in 2027. The 2025-26 vintage follows in 2030-32, but the pipeline is already building. That creates a wall that is invisible in hyperscaler filings but real at the vehicle level.
In 2024, that debt priced on the assumption of on-time construction, cheap power, and a hyperscaler lease that would be renewed without negotiation. In 2027, it will reprice on what actually happened. I do not expect a Microsoft or Amazon default to trigger this. I expect a refinancing conversation at a vehicle that needs an extension, a higher coupon, or a guarantee draw, and a private credit fund that has to explain a mark to its investors.
The BIS paper in March flagged the stock. Industry data fills in the timing. The 2023 vintage financed early hyperscale JVs where land and power were secured before the AI surge. It used 5-year private placements, first maturity in 2028, priced around 5-6% with IG-like lease coverage. The 2024 vintage financed the first large AI-optimized campuses, GPU-dense and liquid-cooled. It used 5-7 year debt, maturing 2029-31, priced around 6-7% on compressed spreads for the hyperscaler name. The 2025-26 wave is different. It combines SPVs plus project finance plus ABS, with 5-7 years plus a 3-year construction loan, maturing 2028-33 with extensions, priced at 6.5-8% with completion risk still open.
Two things matter here. A data center is a 15-year asset, lease plus renewal. The debt funding it is half that. The structure assumes refinancing succeeds. I have seen this mismatch before in 2007-08 conduits: short debt funding long assets works until it needs to be refinanced. Then terms change. The wall is also concentrated. Unlike corporate bonds that are staggered across issuers, data center private placements cluster around the same construction cycle. When power delivery slips for one campus, it often slips for the neighboring one. Same grid, same contractor, same equipment lead times.
Investors who lent against a Microsoft lease in 2024 were not really lending to Microsoft. They were lending to a single-asset vehicle whose ability to repay depends on that lease still being wanted at the same rent in 2029. That is not Microsoft risk. It is single-building risk. A 10-year Microsoft bond, MSFT 4.5% 2033 for example, is a claim on Microsoft global cash flow. If one data center underperforms, the other businesses cover it. A data center SPV private placement is a claim on one building cash flow, plus whatever guarantee was negotiated. Construction risk lives at the vehicle. A 12-18 month delay does not change Microsoft rating. It changes whether the vehicle generates cash in time to service debt before maturity. The BIS notes this explicitly: risk is transferred from corporate balance sheet to vehicle, then to non-bank holder.
Power is the bottleneck, not concrete. Grid interconnection queues in Virginia, Arizona, and parts of Europe run 2-4 years. A completed shell without power is not an operating asset. I am tracking several 2024-vintage campuses where shell completion is on time but substation delivery is 9-15 months late. The lease often does not commence until power is live. Rent does not start until power is live. Coverage is lease-dependent. Underwriting assumed 1.4-1.5x DSCR on a 15-year lease. If power costs rise 15-20% or utilization ramps slower than 70% in year 2, coverage falls below 1.2x. That is the typical threshold for refinancing without additional equity.
For 2027-28 I am watching refinancing at the vehicle first. The private credit fund that bought a 5-year note in 2023 expected to be refinanced by a similar fund in 2028, or by an ABS takeout once the asset was stabilized. That takeout requires a stabilized asset and a receptive market. If either is missing, the choice is extension at a higher spread, additional equity from the sponsor, or a guarantee draw. My base case for weak 2023 vintages is extension plus 100-200bp. I will watch for extensions described as amend and extend and for new-issue spreads versus 2024 levels.
Then guarantee activation, where off-balance-sheet becomes on-balance-sheet. Many JVs include completion guarantees, cost-overrun guarantees, or a put on capacity. These sit in footnotes, not debt tables. Hyperscalers disclose them as commitments and contingencies, often without a notional. If a 2024 campus needs EUR 200-400m of overrun funding to reach power-on, the hyperscaler can either inject equity and keep the structure off-balance-sheet or let the vehicle draw the guarantee and bring the obligation home. The first preserves optics. The second preserves cash. Management teams with leverage targets will hesitate until they have to. I will watch for changes in unconsolidated JV commitments in 10-K and 10-Q, and any shift from operating lease language to variable interest entity consolidation.
Then marks, not defaults. Private placements do not trade daily. But insurers and private credit funds must mark them for solvency and investor reporting. A vehicle that was par in 2024 at 6% with 1.5x coverage may be 92-95 in 2027 if coverage is now 1.15x and refinancing will be at 8%. That is not a default. It is a loss that has to be recognized, and it reduces capacity for the next wave. The first sign will not be a missed payment. It will be a fund reporting a selective markdown in infrastructure debt and a slower deployment pace. The second sign will be a hyperscaler quietly consolidating a JV it previously called unconsolidated.
My probabilities are straightforward. I assign 50% to an orderly roll. Assets power on within 6 months of plan, leases commence, DSCR stays above 1.3x. Refinancings in 2028 clear 50-100bp wider than 2024. The ABS market absorbs stabilized deals. Guarantees stay untouched. I assign 35% to extended and repriced. That is my central risk. 30-40% of 2023-24 vintages need a 12-18 month extension, plus 150-250bp and partial sponsor equity. Private credit marks fall 5 to 8 points. Hyperscalers inject EUR 1-3bn in aggregate completion equity. No rating action, but footnote obligations rise. I assign 15% to structure fails. AI capex slows more broadly, power constraints leave campuses underutilized, several vehicles cannot refinance without material guarantee activation. Hyperscalers consolidate EUR 5-10bn of previously off-balance-sheet debt.
I keep orderly roll at 50% because hyperscalers have the cash to bridge. Microsoft, Amazon, and Alphabet each generated USD 60-100bn plus operating cash flow in 2024-25. The question is willingness, not ability. The 35% case does not require AI to fail, only to be late.
Into September 2027 I will track a short list. At the vehicle level: completion versus plan, power-on date versus shell, any amend and extend. At the hyperscaler level: JV footnotes, guarantee language, lease reclassification. A capex cut also matters. It means fewer new vehicles will be needed. At the market level: private credit marks and deployment pace, insurance allocation to infrastructure debt, hyperscaler 10-year spreads versus Treasuries. They widened 5-10bp in 2025. A further 10-20bp would signal crowding out from my September 8 thesis. I expect the first evidence to appear in refinancing terms and private-market marks, not in public ratings.
The hyperscaler is unlikely to default. The vehicle may need additional equity, a guarantee, or an extension. That is where the squeeze becomes visible. The AI infrastructure trade has been priced as if debt is permanent and off-balance-sheet means risk-free. Neither is true. Debt that looked elegant in 2024 has to be refinanced in a world where power is scarce, construction is late, and private credit is more selective. In 2027, I expect to find out whether 2024 underwriting was right about utilization. The hyperscaler will not default. The vehicle might need to be rescued. The holder will have to mark it. That is the refinancing squeeze. It will show up first in a footnote, then in a spread, then in a fund NAV, long before it shows up in a rating action.
What do you think, does off-balance-sheet still mean off-risk when the debt comes due? Join the discussion in the comments.
Alessandro Rovelli
Rovelli Research
Independent credit investment research
