AI’s $1.5 trillion funding wall is a credit-market test

01 · The information delta

The relevant financing number is larger than the cash-flow gap

A capex-minus-cash-flow calculation produces a useful estimate of how much cash the largest technology companies may lack internally. Our narrower operating-cash shortfall is roughly $200–250B. That is not the same question as how much gross financing the entire AI buildout may send through capital markets.

Morgan Stanley Research estimates that capital expenditure by the five largest U.S. technology companies rises from nearly $800B in 2026 to approximately $1.2T in 2027 and $1.4T in 2028. The underlying estimates used in our worksheet are $1.230T and $1.396T for 2027 and 2028 respectively.

Direct hyperscaler borrowing is only the first layer. Goldman Sachs expects roughly $300B of project-finance and data-center transactions in 2027, above and beyond direct hyperscaler issuance. Those structures can involve investment-grade bonds, high yield, leveraged loans, private credit and later ABS or CMBS takeouts.

Combining direct corporate debt with that project layer produces an estimated $700–730B of gross external financing in 2027. For 2028, holding project finance flat or scaling it with projected capacity additions gives a scenario range of $790–865B. The two-year total is approximately $1.5–1.6T.

“The market risk is not simply whether hyperscalers can pay. It is whether credit can absorb the entire financing stack.”

Fiat Elpis research note, 21 August 2026

02 · Signals to track

Four channels determine whether the wall becomes a market event

03 · What may be mispriced

Markets may be watching solvency while the pressure builds in supply

The aggregate estimates do not imply an imminent hyperscaler default cycle. Large technology companies can issue ahead of need, refinance maturities or preserve liquidity by choice. The more plausible pressure point is the marginal project, where credit terms determine whether construction remains economic and on schedule.

The sensitivity is material. A 29% historical debt-to-capex ratio produces roughly $1.44T of two-year financing. The 35% base case produces about $1.60T. A more debt-intensive 40% outcome reaches approximately $1.73T. The precise figure matters less than recognizing the scale and the number of financing channels involved.

1. Separate the $200–250B operating-cash shortfall from the much larger gross supply that markets must absorb.
2. Do not count the same compute stack at both the vendor and customer level, or add a hyperscaler prepayment again as neocloud financing.
3. Treat the $300–375B 2028 project layer as a Fiat Elpis scenario; no published 2028 project-finance forecast is being claimed.
4. Watch credit concessions and project delays, not only headline capex guidance or issuer-level leverage.

04 · What would change my mind

What would shrink or defer the funding wall

This is a financing-supply thesis. It weakens if the buildout needs less external capital or markets absorb the supply without repricing:

• Hyperscalers cut 2027–2028 capex materially rather than merely shifting it between quarters.
• AI revenue, tax shields and operating cash conversion improve fast enough to reduce direct issuance.
• Customer prepayments fund a substantially larger share of neocloud and project costs.
• Power, permitting or construction delays defer enough capacity to flatten the annual financing peak.
• Investment-grade, structured and private-credit markets digest the issuance without wider spreads, larger concessions or weaker secondary performance.