Morgan Stanley maps $1.5 trillion AI data-center financing gap
Morgan Stanley says AI data-center spending needs about $1.5 trillion of outside capital. The underwriting lesson is lease cash flow, power access and refinancing evidence, not AI demand alone.
Morgan Stanley put a hard number on the AI data-center funding problem: roughly $2.9 trillion of global data-center spending through 2028, with about $1.5 trillion that has to come from outside capital. The bank's July research and an Aug. 6 podcast said private credit and asset-based finance could supply about $800 billion of that gap, with another $150 billion coming through data-center ABS and CMBS. For Plutus readers, the data-center financing boom is less about AI demand than the credit structure forming underneath it. Morgan Stanley estimates the buildout could require $1.5 trillion of outside financing by 2028. That pushes the story from growth projections into collateral, lease terms, power contracts, refinancing windows and the cash flows lenders can actually underwrite. Porter Stansberry flagged the financing question in an Aug. 11 thread. The harder evidence is in Morgan Stanley Research , Morgan Stanley's published podcast transcript , Moody's data-center credit-risk page and L&G's securitized-credit analysis . Morgan Stanley's split is useful because it names the capital stack. The firm expects about $1.4 trillion of hyperscaler capex to be funded with cash flow, then about $200 billion of investment-grade tech debt, $800 billion of private-credit or asset-based financing, $150 billion of securitized credit and $350 billion from other capital sources. In the podcast, Vishwas Patkar said the $1.5 trillion gap is larger than both the high-yield market and the leveraged-loan market. It means the financing is moving from one clean balance-sheet story into several underwriting channels at once. Corporate bonds are backed by a giant issuer. Private credit and ABF are backed by project cash flows, leases, collateral, construction progress and bespoke lender protections. Securitization usually arrives later, after the asset is built, leased and producing income. L&G made that sequencing explicit in a February note on data-center securitization. It said most securitized issuance refinances income-generating assets rather than funding initial construction, and that data-center ABS and CMBS issuance surpassed $25 billion in 2025. The same note flagged a later refinancing problem: many deals carry a five-year anticipated repayment date, so 2028 to 2030 becomes a real test if new takeout markets are less friendly. Moody's frames the same market through credit risk. Its data-center page says developers and landlords will need equity, bank loans, corporate bonds, securitized bonds and project-finance vehicles, and that leverage will likely rise for developers focused on hyperscale buildouts due in 2026 through 2028. It also highlights power, water and development approvals as live constraints. For an MCA, RBF or equipment-finance operator, the direct parallel is not that a restaurant file looks like a data center. It is that fast demand can hide bad repayment logic. A funder can see a merchant with real revenue and still miss that the observed cash flow depends on one processor, one seasonal buyer, one expiring contract, one second-position lender or one landlord dispute. A data-center lender can see AI demand and still be wrong about power delivery, tenant renewal, lease step-ups, obsolescence, construction timing or the takeout market. The underwriting discipline is the same shape at a different scale. Name the revenue source. Name the constraint. Name the debt-like obligation that does not sit where a casual reviewer expects it. Name the exit. If repayment depends on refinancing, say who the next buyer is supposed to be and what evidence would make that buyer show up. There is also a crowding question. If private credit is asked to absorb hundreds of billions of data-center finance, those lenders do not become less demanding elsewhere. The same insurance, pension, private-credit and ABF buyers that fund infrastructure can also sit behind warehouse lines, forward-flow programs and asset-backed facilities. When their books fill with one giant theme, smaller specialty-finance originators should expect more questions about concentration, liquidity, documentation and performance history. None of this makes AI infrastructure a bad credit. The stronger version is almost the opposite: if the leases are durable, the power is real, the tenant mix is defensible and the refinance path exists, the file may be attractive. But the phrase AI demand is not collateral. The collateral is the contract, the operating asset, the power access, the tenant payment stream and the next market that can refinance the structure. Opinion The collateral test is not only the building AI infrastructure credit will be judged on leases, power access, tenant concentration and takeout markets, not only on the fact that demand for compute is rising. The same lesson applies to smaller finance desks: a hot sector does not make a weak file better. It only makes more lenders willing to ask what is really behind the cash flow. Plutus tracks how capital-market pressure reaches the underwriting file at plutusai.net .