Goldman, JPMorgan, Morgan Stanley Agree: The AI Spending Cycle Is $5T and Nearly Half Is Debt
Goldman Sachs is projecting $5.3 trillion in total AI and data center spending through 2030. JPMorgan’s independent estimate arrives at $5.5 trillion, with a specific breakdown putting $4.1 trillion of that in debt financing. Morgan Stanley, running a shorter horizon, forecasts $2.9 trillion in data center construction alone through 2028.
Three of Wall Street’s largest research shops, working separately, have produced essentially the same number. What unites all three estimates is a structural observation starting to attract serious attention from economists: unlike the dot-com era, AI infrastructure is being financed heavily on credit.
The Financing Stack
Morgan Stanley’s breakdown of the $2.9 trillion through 2028:
| Source | Amount |
|---|---|
| Hyperscaler cash flows | $1.4T |
| Private credit / asset-based / JV debt | $800B |
| Other capital | $350B |
| Corporate debt (public bonds) | $200B |
| Securitized credit | $150B |
Nearly 48% of the build is credit-funded rather than equity. JPMorgan’s estimate implies the debt share grows as the cycle progresses, reaching $4.1 trillion against $5.5 trillion total by 2030.
Goldman adds a secondary warning: AI capex estimates are rising faster than actual construction. The constraint is shifting from model demand to financing capacity, power access, and project execution velocity. Earlier Sightline research found that only 40% of announced data center projects are likely to be completed.
The Damodaran Warning
NYU Stern professor Aswath Damodaran, known in finance as the “Dean of Valuation,” has drawn an explicit contrast between this cycle and 2000:
The dot-com boom required minimal capital expenditure. Companies raised public equity and spent it on software and headcount. When the bust came, shareholders absorbed the losses. The losses were painful but contained within equity markets.
AI infrastructure requires the largest coordinated CapEx buildup in US corporate history. When financed with debt, the correction dynamics change. Defaults do not stay with shareholders. They propagate through private credit funds, structured products, and eventually into the broader financial system.
“There’s a very real chance that if there’s a correction and companies start having problems, that problem is going to show up as distress and default,” Damodaran said. “That really doesn’t stay restricted. It spills over into the rest of society.”
He is not calling a bust. He is noting that if one comes, it will behave like 2008 more than like 2000 — and that the scale of physical infrastructure makes recovery slower than a software correction.
Three Structural Differences From 2000
Capital intensity: AI data centers require land, power transmission, cooling systems, buildings, and bespoke silicon. The dot-com era involved almost none of this. Stranded AI assets are physical and expensive to repurpose.
Debt maturity concentration: Much of the AI infrastructure debt is structured as private credit with 3-7 year terms. A demand correction across multiple hyperscaler tenants would trigger covenant tests across hundreds of billions in credit facilities simultaneously.
Financing counterparties: The dot-com era used public equity markets as the primary loss-absorption vehicle. AI infrastructure is backed by private credit funds, infrastructure investors, and asset-based lenders. These entities have different liquidity profiles and different contagion paths than public equity shareholders.
Goldman’s observation that estimates are rising faster than construction is a distinct, near-term risk. Projects financed on projected utilization could face stress before longer-cycle credit risk materializes, if model demand softens before capacity is deployed.
The banks are not forecasting a collapse. They are producing, for the first time with precision, a map of how one would propagate — and it does not look like the last time.