Goldman Sachs Raises AI Project Financing Forecast to $545 Billion; Fed Warns Off-Balance-Sheet Liabilities Exceed $3 Trillion
nashnova research
Goldman Sachs lifted its 2027 global AI financing forecast from $300 billion to $545 billion, while the Fed warned that seven hyperscalers' off-balance-sheet liabilities now exceed $3 trillion — a funding surge and systemic-risk alarm arriving in the same window.
Where does $545 billion go — how is "shell" vs. "compute" split?
Goldman breaks the forecast into two buckets: $225 billion for "shells" — data-center buildings that may not yet be powered — and $320 billion for "compute" — chip procurement that may not yet be deployed.
This means → nearly sixty percent of the money is locked in chip purchases, still a long way from becoming usable compute capacity.
A growing share of these projects is financed through SPVs — bankruptcy-remote special-purpose vehicles that wall off debt from the parent company. In plain terms = borrowers pay a higher premium to buy certainty that the funding will close and the money will flow.
Can the money actually be borrowed — why has Nvidia's $50 billion platform seen zero deals?
Nvidia launched a $50 billion AI financing platform in August. To date it has closed no transactions, because lenders demanded larger credit backstops.
JPMorgan, Citi, and Goldman have all publicly tightened underwriting discipline, capping exposure to junk-rated issuers. This reflects big banks starting to pump the brakes on AI exuberance.
Apollo president Jim Zelter drew a hard line: loans must fully amortize within five years, carry investment-grade guarantees, and bear zero residual-value risk. This means → the private-credit market is no longer willing to underwrite a revolving, borrow-to-refinance model for AI infrastructure.
How much market do export curbs cut — what is the hit to Nvidia?
The U.S. tightened AI chip export controls this week, adding Singapore, Indonesia, Thailand, Malaysia, and Brazil — aiming to close the loophole of Chinese firms leasing compute through Southeast Asian data centers.
Jefferies estimates the new rules put $44 billion to $62 billion of Nvidia's 2027 chip sales at risk — roughly 11% to 16% of its data-center revenue.
Separately, about 600,000 Nvidia chips worth roughly $18 billion are stranded in Hong Kong with no import license. In plain terms = these chips can neither enter China nor easily be resold, leaving them as inventory suspended in limbo.
Why is the Fed drawing a telecom-bubble parallel?
Fed Governor Michael Barr cited the latest Financial Stability Report: capital spending and debt issuance by the seven largest AI firms since 2020 are now comparable to the 1995 telecom broadband investment cycle — and the recent acceleration may soon surpass it.
The key number: off-balance-sheet obligations — long-term contracts, lease commitments, residual-value guarantees — are growing "far faster" than on-balance-sheet debt. The FSR's notional ceiling estimate tops $3 trillion, versus on-balance-sheet debt projected to exceed $800 billion by end-2026.
This means → regulators are less worried about the debt you can see on a balance sheet than about the hidden load buried in contracts and leases that never shows up directly in earnings filings.
What does the 33-gigawatt gap reveal?
An telling contradiction has surfaced inside Goldman itself: the equity-research team's chip-sales forecast and the credit-strategy team's data-center capacity forecast diverge by roughly 33 gigawatts through end-2028 — equivalent to the entire baseload electricity demand of the United Kingdom.
In plain terms = two desks at the same bank disagree: one says the chips will sell; the other says the data centers won't be built in time. Buying a chip is not the same as bringing compute online.
This reflects the market's deepest divide over this AI capex wave: money is being spent, but will it actually produce working capacity? Whether real hyperscaler demand can absorb the SPV premium is the pivotal test of this expansion's sustainability.
市场有风险,内容仅供研究参考,不构成投资建议。
