The AI investment boom is moving beyond the balance sheets of large technology companies. As the cost of data centers, chips, power connections and networking equipment rises, more financing is being arranged through private credit, asset-backed loans, joint ventures and special-purpose vehicles.
That does not mean all AI financing is private credit. Large hyperscalers still have substantial cash flow and access to public debt markets. But the industry’s capital needs are becoming too large to be funded only from operating cash flow.
For retail investors and small business owners, the issue is not whether AI demand is real. It is how the growing chain of AI infrastructure debt behaves if revenue growth, power delivery or refinancing conditions fall short of expectations.
A large funding gap is opening
The International Monetary Fund’s April 2026 Global Financial Stability Report illustrates the scale involved. It cites a July 2025-based estimate of $2.9 trillion in data-center capital expenditure through 2028. That estimate assumes $1.4 trillion of hyperscaler cash flow and $800 billion of private bilateral credit.
This is not a forecast of actual lending or a final bill for the AI buildout. The IMF says the estimate includes asset-backed financing, excludes associated power investment and may understate needs after later increases in hyperscaler spending guidance.
Still, the funding gap is large, and AI infrastructure debt is likely to play a bigger role in filling it. The Bank for International Settlements has likewise argued that expected AI investment needs will push financing away from cash flow and toward debt. Private credit can be customized for a specific project, asset or customer contract.
How data-center financing is changing
A typical structure can look simple at first. A joint venture or special-purpose vehicle owns or develops a facility. Sponsors contribute equity, private lenders provide debt, and a hyperscaler or large customer signs a multiyear lease or capacity offtake agreement.
In some cases, the customer also provides a guarantee or other support. That can make debt easier to raise because lenders see contracted revenue behind the project.
But the arrangement can also move borrowing outside the customer’s consolidated balance sheet. Debt may sit in the vehicle rather than directly on the hyperscaler’s books, while the hyperscaler still has a long-term payment obligation that is economically debt-like.
This is not automatically a problem. Project financing has long been used in real estate, energy and infrastructure. The question is whether the contracts, customers and assets can support the debt through a full business cycle.
The underwriting questions that matter
The strongest AI private-credit deals are not necessarily those tied to the loudest AI headlines. They are the ones with durable contracts, financially sound customers, realistic construction budgets and dependable access to power.
Key questions include:
- Is the customer committed through a long-term take-or-pay contract, or can it reduce usage if demand changes?
- How concentrated is revenue in one customer or a small group of customers?
- Can the developer secure enough electricity on time and at a predictable price?
- Are construction, equipment delivery and permitting assumptions realistic?
- How quickly could expensive computing equipment lose value as chip technology changes?
- Will the project generate cash before its debt needs to be refinanced?
Power is especially important. A data center without sufficient electricity is not a productive infrastructure asset, even if AI demand is strong. Delays in grid connections, higher power costs or local restrictions can impair expected returns.
CoreWeave shows both opportunity and concentration risk
CoreWeave offers a useful company-level example, though it is not a proxy for the entire market. At December 31, 2025, the company reported $21.6 billion of indebtedness and $8.2 billion of operating-lease liabilities.
Its model relies heavily on asset-level debt supported by take-or-pay customer contracts. Contracted revenue and customer prepayments can give lenders more visibility into future cash flow and reduce demand uncertainty.
However, contracts do not remove every risk. CoreWeave reported that Microsoft represented about 67% of its 2025 revenue. A large customer can provide stability, but concentration increases exposure if that customer changes spending plans, renegotiates, delays deployment or shifts workloads elsewhere.
The company’s disclosures also highlight power constraints and costs, limited supplier options, large capital needs and fast-moving technology. Those are central issues for much of the data-center financing market.
Where stress could spread
The BIS has identified several ways pressure in a data-center project could move through the financial system. A special-purpose vehicle may need to refinance in a weaker market. Private-credit funds may become less willing to lend when asset values or demand assumptions are questioned. Guarantees may become payable if a vehicle cannot meet its obligations.
Banks can be involved even when they are not the main lender. They may provide revolving facilities, warehouse lines or other funding to private-credit funds. Insurers may also be exposed through investments or credit support.
This creates a chain: hyperscaler commitments support project vehicles; vehicles borrow from private lenders; lenders may rely on fund investors, insurers or banks. The chain can work when contracts perform and capital remains available. It becomes more fragile when several assumptions fail at once.
For investors, private-credit risks should not be judged only by default rates. Valuation, liquidity and refinancing conditions matter too.
Software credit is a separate pressure point
AI’s financing impact is not limited to physical infrastructure. It is also reaching software borrowers, particularly companies with recurring-revenue business models.
European supervisory authorities reported that several major US semiliquid business development companies received redemption requests above their limits in the first quarter of 2026. The funds said withdrawals were associated with retail concern about AI’s potential effect on software-as-a-service businesses, an area where many BDCs have exposure.
This is not proof of a broad software-credit default cycle, nor does it show that all BDCs face the same problem. Private credit is highly varied, and many funds use long-term or locked-up capital.
But it reveals a potential weak point. If AI changes software pricing, customer retention or competitive advantages, lenders may need to reassess the cash flow supporting existing loans. That can be difficult when private loans trade infrequently and valuations rely heavily on models and manager judgments.
Semiliquid funds add another complication. They offer periodic redemptions while investing in assets that may take time to sell. If many investors request withdrawals at once, managers may limit redemptions, sell more liquid holdings first or reduce new lending.
What to watch from here
The Financial Stability Oversight Council has warned about valuation difficulty, limited transparency, growing semiliquid private-credit structures and increasing links with banks and insurers. It reported approximately $445 billion in bank loan commitments to private-credit funds, while noting that the exact amount is not known.
That does not signal an immediate crisis. It does mean the AI private-credit boom deserves closer attention as debt commitments grow.
Watch contract durability, customer concentration, power availability, project completion, refinancing schedules and fund redemption activity. For data centers, AI demand alone is not enough. For software lenders, recurring revenue must remain durable even as AI changes the market.
The central risk is not that AI investment suddenly disappears. It is that long-duration obligations were priced for a future that arrives more slowly, costs more or produces less cash flow than expected.