1. A $2.9 trillion build-out with a funding hole
Morgan Stanley estimates that global data center capital expenditure will reach about $2.9 trillion between 2025 and 2028, and that the largest technology companies can pay for only around half of it from their own cash flow. The remainder, roughly $1.5 trillion, is the data center financing gap: money that has to come from lenders and investors outside the biggest spenders. The research, reported on October 24, 2026, names private credit as a leading candidate to supply a large share of it.
The forecast arrives as artificial intelligence pushes demand for computing capacity higher. Training and running large models requires enormous numbers of specialized chips, servers and the buildings and power systems that house them. Morgan Stanley's figures put a dollar value on that appetite and, just as importantly, on how much of it the established cloud providers cannot cover alone.
A capital expenditure forecast of this size matters beyond the technology sector. Spending on this scale competes for loans, bond buyers and construction capacity, and the way it is financed determines who carries the risk if demand for AI services grows more slowly than builders expect.
2. Where does the $2.9 trillion go?
The bank splits the projected outlay into two broad buckets. About $1.6 trillion is earmarked for information technology hardware, which covers the chips and servers that perform the computing work. The remaining $1.3 trillion goes to physical infrastructure and construction, meaning the buildings, cooling, electrical systems and related works that turn a plot of land into a working facility.
The split shows that hardware is the larger line, but not by an overwhelming margin. Nearly as much money is expected to flow into concrete, steel and electrical plant as into the silicon itself. That matters for financing because the two categories behave differently as collateral: servers lose value quickly as newer generations arrive, while buildings and power connections tend to hold their worth for longer.
3. How is the data center financing gap calculated?
The gap comes from simple subtraction. Amazon, Microsoft, Alphabet, Meta and Oracle, the hyperscalers that carry the heaviest share of the build-out, are projected to fund about $1.4 trillion of the spending from their own operating cash flow. Take that figure away from the $2.9 trillion total and about $1.5 trillion is left over.
In other words, even companies with some of the strongest balance sheets in the world are not expected to pay for the whole programme from internal resources. The shortfall has to be met through borrowing, leasing arrangements, or investment from third parties. Morgan Stanley's analysis treats that external requirement as the central question for credit markets over the next several years.
4. How could private credit help cover the shortfall?
Morgan Stanley's research says private credit could provide roughly $800 billion of the missing money. Private credit refers to loans made by non-bank lenders such as asset managers and specialist funds, rather than by banks or through public bond markets. The preferred route, according to the research, is asset-specific lending.
Asset-specific lending ties a loan to a particular facility or set of equipment rather than to the borrower's entire balance sheet. Lenders look at the cash the asset is expected to produce and at what they could recover if it failed. For data centers, that structure suits projects where a tenant has signed a long lease, because the rent provides a predictable stream to service the debt.
The $800 billion estimate covers just over half of the $1.5 trillion gap. The research as relayed does not assign the remaining amount to a specific source, so the rest is left to other forms of financing that have yet to be identified.
Hyperscaler lease commitments exceed $1.2 trillion in undiscounted future payments, according to the research. That figure adds up the rent these companies have promised to pay over the life of their agreements, without adjusting for the time value of money.
Those commitments are what make much of the construction lending possible. A developer that has a hyperscaler signed up as a long-term tenant can show lenders a contracted income stream, and lenders extend credit against it. The arrangement also concentrates exposure: if a major tenant were to scale back its demand, developers and the creditors who financed them would feel the effect first.
5. What role could Nvidia play?
Nvidia has taken part in discussions about backstop guarantees of up to $250 billion for data center debt. A backstop guarantee is a commitment to step in and cover obligations if a borrower falls short, which can make lenders more comfortable extending credit. The research does not indicate that these talks have been finalized, and the discussions remain unconfirmed.
Supplier-backed financing of this kind can help keep orders flowing, since customers who can borrow more cheaply are able to buy more equipment. The trade-off is that it ties a chipmaker's financial health more closely to the debt of the customers it sells to. If those customers struggled to repay, the supplier would be exposed to losses on top of any decline in its sales.
6. Is power a limit that money cannot solve?
Funding is only one of the constraints on the build-out. Revised forecasts put the shortfall in United States data center power supply at 32 to 33 gigawatts by 2028. A gigawatt is a billion watts, and facilities of this kind can draw large amounts of electricity continuously, so a gap of that size represents a very large quantity of capacity that does not yet exist.
Capital can buy chips and pay for construction crews, but it cannot instantly produce new generation, transmission lines or grid connections. That is why the power figure sits alongside the financing numbers: a project that is fully funded can still be delayed if the electricity to run it is not available when the building is ready.
7. Could Washington end up as the backstop?
Beyond the bank's own findings, one commentator has argued that private credit is financing the AI build-out against collateral with a short useful life, and that Washington will eventually step in to backstop that lending. The argument draws on reasoning put forward in September by Alex Karp, the chief executive of Palantir, which the commentator says much of the market has not yet absorbed and which extends further down the technology stack.
This is the commentator's opinion and a forecast, not a finding of Morgan Stanley's research. It is worth noting only because it points to a concern that runs through the debate: when so much borrowing rests on assets that may age quickly, the question of who absorbs losses in a downturn becomes a policy question as well as a financial one.
8. What this means for investors and lenders
Morgan Stanley's numbers suggest that the largest buyers of AI infrastructure have moved beyond what their cash flow can finance. For credit investors, that creates opportunity: the research points to private credit chances that may offer higher yields than conventional corporate debt.
The same numbers carry risks. The lease commitments of more than $1.2 trillion support a great deal of construction lending, so a pullback in hyperscaler demand would reach developers and their creditors. Supplier-backed arrangements, if they materialize, would bind a chipmaker's fortunes to its customers' debts. Investors weighing these opportunities will want to judge how durable the underlying leases are and how quickly the equipment securing the loans loses value.
For now, the analysis remains a forecast covering 2025 through 2028. Its headline figures, $2.9 trillion in spending and a $1.5 trillion gap, give lenders, developers and policymakers a common reference point for how much outside capital the AI build-out is expected to need, and for how many of the key arrangements are still unsettled.
9. Why lenders look at data centers differently from other property
Data centers sit between real estate and technology. A conventional office building is financed against its location and its tenants, while a data center also depends on specialized equipment and on the continuing demand for the computing it hosts. Lenders therefore examine who the tenant is, how long the lease runs and how essential the facility is to the tenant's business.
This is part of the reason asset-specific structures are attractive in the Morgan Stanley framework. By isolating a single facility and its contracted rent, a lender can size a loan to the strength of that arrangement without taking a view on the borrower's entire business. It also lets investors who want exposure to AI infrastructure gain it through credit rather than through equity in the companies doing the building.
Higher yields are the usual compensation for the extra complexity. Private lenders typically negotiate tighter terms and more bespoke structures than public bond buyers, and they accept less liquidity in return for the additional income. Whether that trade-off proves adequate depends on how the underlying leases and equipment perform over several years.
10. What to watch next
The most immediate unknown is whether the reported talks over Nvidia backstop guarantees, worth up to $250 billion, are completed or fade, since the research does not treat them as settled. Any formal agreement would show how far suppliers are prepared to go to keep their customers able to borrow.
Another marker is the pace of power supply. With the forecast shortfall at 32 to 33 gigawatts by 2028, announcements about new generation and grid connections will indicate how much of the $2.9 trillion can actually be spent on schedule. Finally, the size of private credit's contribution, estimated at roughly $800 billion, will show whether the rest of the gap is filled by other lenders or remains open as the decade advances.