At least US$1 trillion of data-centre debt matures within the next five years, according to BloombergNEF. Data-centre developers raised approximately US$1.8 trillion between 2020 and the first half of 2026 across debt, equity and securitised instruments. Nearly half of that capital was raised between the third quarter of 2025 and the first half of 2026.
The AI infrastructure debate has moved beyond capex forecasts and regulatory concern. BloombergNEF has now quantified the financing architecture, and the result is a refinancing wall large enough to change how the entire buildout must be valued. That is the material escalation. The market no longer needs to infer that AI infrastructure has become dependent on leverage. It can now see the scale, concentration and timing of the refinancing requirement.
What BloombergNEF found about data-centre debt
BloombergNEF’s review covers more than 2,000 transactions involving the 125 largest data-centre developers. The headline figures:
| Measure | Figure |
|---|---|
| Capital raised, 2020 to first half of 2026 | Approximately US$1.8 trillion |
| Capital raised, third quarter of 2025 to first half of 2026 | US$794 billion |
| Debt maturing within the next five years | At least US$1 trillion |
| Private markets’ share of colocation financing, first half of 2026 | 56% |
| Nvidia’s direct support (guarantees, prepayments and related arrangements) | At least US$198 billion |
Colocation providers, companies that develop and operate facilities for multiple customers, have now raised more capital than the four largest hyperscalers combined.
The maturity profile of data-centre debt is uneven. Neoclouds and colocation providers face the sharpest increase in refinancing needs. BloombergNEF calculates that neocloud debt maturing in 2030 is 18 times the volume due in 2029. Smaller developers with significant construction debt face greater uncertainty because repayment depends on delivery, customer ramp-up and the continuing availability of capital.
The report also quantifies the credit support beneath the buildout. Beyond Nvidia’s support, other structures use GPU-backed loans, residual-value guarantees and support from investment-grade tenants or sponsors.
This is not conventional corporate borrowing. It is an interdependent financing network in which equipment value, tenant credit, construction completion, power availability and future refinancing can all support the same transaction.
How data-centre debt moves from construction to permanent funding
Data centres move through distinct financing stages:
- Land and power rights are typically funded with sponsor equity and development loans.
- Construction introduces cost, delivery and energisation risk.
- Lease commencement depends on tenant acceptance and contractual payment triggers.
- Stabilisation allows a facility to refinance through asset-backed securities, commercial mortgage-backed securities, project bonds or permanent loans.
The transition between stages is the vulnerability. Construction debt is often short-dated because lenders expect a completed project to access cheaper permanent capital. If completion is delayed, power is unavailable or leasing assumptions weaken, the refinancing event can require additional equity, higher spreads or a reduced valuation.
The CRE Finance Council (CREFC) counted US$116 billion of secured data-centre project bonds outstanding by 8 October. These bonds commonly finance facilities still 12 to 18 months from delivery. High-yield issues generally mature in five years and typically amortise only about 20% of principal before maturity, although terms vary materially.
For data-centre debt, refinancing is therefore not a technical footnote. It is embedded in the repayment plan.
Why the refinancing wall matters now
The US$1 trillion maturity wall arrives as sovereign yields remain high, AI equities are de-rating and bank supervisors are increasing scrutiny.
That combination changes economics before any project defaults. A higher refinancing coupon reduces equity value. A lower advance rate requires additional sponsor capital. A delayed securitisation can extend expensive construction financing. A weaker tenant or equipment residual value can reduce recoveries.
This matters for semi-liquid funds and NAV facilities. If project marks fall, fund net asset value (NAV) can decline while borrowing-base headroom tightens. If investor redemptions rise at the same time, managers may preserve cash, draw facilities or sell more liquid assets rather than the loans causing the pressure.
Where capital may move
Capital is unlikely to abandon digital infrastructure. Demand for compute, power and connectivity remains substantial. The distinction will be contractual.
Completed and powered facilities with enforceable leases, credible tenants, conservative leverage and meaningful amortisation should remain financeable. Construction-stage projects reliant on future leasing, narrow guarantees or a single refinancing event will require more equity and stronger protection.
Pricing for data-centre debt already reflects tenant quality, according to CREFC:
| Construction loans | Spread | Loan-to-cost |
|---|---|---|
| Highly rated hyperscaler tenants | Low 200 basis points | As high as 85% |
| Weaker or unrated tenants | 200 to 300 basis points wider | 70% to 80% |
Four risks in data-centre debt
- Correlation disguised as diversification. This is the largest risk. A technology company may be a tenant, customer, equipment supplier, guarantor and strategic investor across several projects. Multiple loans can therefore depend on the same demand and capital source.
- Residual value. GPUs and related equipment can support borrowing, but rapid technological change may reduce collateral value faster than loan amortisation.
- Power. A project can have land, customers and financing but still fail to secure timely grid connection or generation capacity. Delays increase interest carry and can invalidate return assumptions.
- Refinancing concentration. If many projects mature after the same buildout wave, lenders can demand higher spreads or lower leverage even when individual assets perform.
The evidence does not establish an imminent crisis. It shows that future returns increasingly depend on refinancing markets.
What Perennial is watching
The critical signals are:
- construction-to-permanent refinancing, maturity extensions and sponsor equity contributions;
- GPU residual-value assumptions, tenant guarantees, prepayments and completion support;
- loan-to-cost ratios, amortisation and debt-service reserves; and
- project-bond spreads, private marks, NAV-facility headroom and whether asset sales clear near carrying value.
The thesis would weaken if projects refinance on schedule, spreads compress and contracted utilisation converts into durable cash flow. It would strengthen if facilities are repeatedly extended, sponsors supply rescue equity, guarantees are activated or private marks lag public debt.
Frequently asked questions
How much data-centre debt matures in the next five years?
BloombergNEF estimates that at least US$1 trillion of data-centre debt matures within the next five years. Developers raised approximately US$1.8 trillion between 2020 and the first half of 2026, nearly half of it between the third quarter of 2025 and the first half of 2026.
Why do data-centre projects depend on refinancing?
Construction debt is often short-dated because lenders expect a completed project to access cheaper permanent capital. High-yield project bonds generally mature in five years and typically amortise only about 20% of principal before maturity, so refinancing is embedded in the repayment plan.
Which developers face the greatest refinancing pressure?
Neoclouds and colocation providers face the sharpest increase in refinancing needs. BloombergNEF calculates that neocloud debt maturing in 2030 is 18 times the volume due in 2029. Smaller developers with significant construction debt face greater uncertainty.
How does Nvidia support data-centre debt?
BloombergNEF found that Nvidia has extended at least US$198 billion of direct support through guarantees, prepayments and related arrangements. Other structures use GPU-backed loans, residual-value guarantees and support from investment-grade tenants or sponsors.
Does the refinancing wall mean a crisis is coming?
The evidence does not establish an imminent crisis. It shows that future returns from data-centre debt increasingly depend on refinancing markets.
The Perennial view
AI infrastructure is not facing a demand problem alone. It is facing a duration problem.
The industry has financed long-lived and construction-sensitive assets with a capital stack that often expects refinancing before underlying economics are fully proven. BloombergNEF’s US$1 trillion estimate makes that mismatch measurable.
The forward signal is not simply which projects can raise capital. It is which structures can survive a higher refinancing coupon, delayed completion and lower collateral value without depending on another round of optimistic financing.
Access only matters when the structure is worth accessing.
We are watching the conversion of construction debt into durable permanent funding. Sign up free for Perennial Market Intelligence to receive follow-up analysis as the refinancing signal develops. [Free signup — link pending]
Explore more analysis on AI infrastructure financing at Perennial Markets.
Where Insight Meets Opportunity.
Sources
- Financing Compute: The US$1.8 Trillion Bet on AI Data Centers — BloombergNEF, 19 November 2026
- Data Center Primer, Second Edition — CRE Finance Council, November 2026
- AI and private-credit scrutiny — ECB Banking Supervision, 19 November 2026
Financial-information disclaimer: This material is general information and commentary only. It is not personal financial, legal or tax advice, nor an offer or recommendation to buy or sell any security. Infrastructure and private-market investments can be illiquid and involve construction, refinancing, valuation and counterparty risk. Verify current information and consider your circumstances and eligibility before acting.





