Institution: The Financing Situation for AI Computing Power Construction Next Year Becomes Increasingly Abnormal
Cognitive dissonance— the discomfort experienced when holding two contradictory beliefs simultaneously— seems to have no effect on investment banks. For example, take a look at this chart:
The chart comes from a Goldman Sachs report on the financing structure of AI data center projects. The light blue line represents chip sales, sourced from Goldman Sachs' equity research team; the dark blue line represents the new installed capacity of data centers, sourced from Goldman Sachs' credit strategy team. The gap between the two lines is projected to reach 33 gigawatts by the end of 2028, roughly equivalent to the basic electricity demand of the UK. {#quote_GS}
This gap could potentially be larger. Goldman Sachs credit analysts, led by Shamsher Ali, have raised their forecast for global AI project financing in 2027 to $545 billion, more than double the previous estimate. {#quote_BABA}
The timing seems quite delicate. Colette Kress, the interim CEO of "Nvidia Bank," will make her public debut after today’s market close, with multiple reports indicating that her attitude towards supplier financing is much more cautious compared to her predecessor.
Even before Jensen Huang stepped down, Nvidia's "balance sheet as a service" model had already shown signs of nearing its expansion peak. Media reports last week indicated that a $500 billion AI financing platform launched in August has yet to close a single deal, as lenders demand stronger guarantees. Several banks, including JPMorgan, Citigroup, and Goldman Sachs, are emphasizing underwriting discipline and controlling risk exposure to junk-rated issuers. {#quote_JPM}
The bank's credit team previously predicted a global AI project financing supply of $300 billion. They now break down the estimate: $225 billion is directed towards "shell data centers" (i.e., construction of facilities that may not yet be powered on), while another $320 billion is aimed at "computing equipment" (i.e., chips, which also may not yet be powered on).
An increasing proportion of AI capital expenditures will be financed through bankruptcy-remote special purpose vehicles (SPVs), even if issuers need to pay a higher risk premium. As the scale of computing projects continues to expand, the ability of the private equity market to secure financing and ensure certainty of execution becomes particularly valuable— even if the cost of capital is higher.
GPU financing is now accounted for separately, with one core variable being the lifespan of the chips. Nvidia's product iteration cycle turns hardware upgrades into an unending loop...
To maintain this model, Nvidia and cloud giants set a floor price to cover lenders' default risks while also presuming that hardware retains residual value after lease expiration.
This interconnected off-balance-sheet financing raises concerns in the market, and signs of cooling demand have already emerged. Jim Zelter, President of Apollo Global Management, stated during the third-quarter earnings call that Apollo hopes to fully amortize loans within five years, backed by investment-grade counterparties, without bearing residual value risk.
Federal Reserve Governor Michael Barr mentioned in a speech yesterday:
By 2026, the on-balance-sheet debt of the seven major AI giants is expected to surge, likely surpassing $800 billion by year-end. The expansion of corporate debt in this sector evokes memories of the investment cycle in the telecommunications industry at the turn of the millennium. The Federal Reserve's latest "Financial Stability Report" compares the cumulative capital expenditures and bond issuance of cloud vendors since 2020 with all telecommunications companies during the broadband investment cycle of 1995. Although cloud vendors have not yet surpassed the telecom companies of that era, the recent acceleration in capital expenditures suggests they may soon catch up. {#quote_LNVGY}
The scale of these companies' on-balance-sheet debt has significantly increased, while other obligations such as long-term contracts, lease commitments, and residual value guarantees have grown even faster. The "Financial Stability Report" estimates that these obligations (i.e., off-balance-sheet liabilities) have a potential nominal ceiling exceeding $30 trillion. A large portion stems from complex transaction arrangements: high-rated cloud giants and large enterprises indirectly provide financing support for third-party projects with weaker credit profiles and utility investments. These off-balance-sheet liabilities include energy supply contracts, long-term data center and computing leases, and various off-balance-sheet financing structures. As I mentioned in my previous speech, ordinary investors find it challenging to see through these financial structures; once market sentiment shifts, the lack of transparency can amplify selling pressure, as the true risk points are difficult to locate. Therefore, these large off-balance-sheet structures warrant close monitoring. Equally concerning is that, aside from leading cloud vendors, companies with a high reliance on external financing and weaker credit quality have seen a significant rise in capital costs in recent months, turning to more complex asset-backed financing.
Despite these concerns, Goldman Sachs' core logic is that the scale of AI project financing must continue to expand; otherwise, the entire computing power expansion will be unsustainable.
The Goldman Sachs team stated: "Securitized loans backed by GPUs primarily come from emerging cloud vendors, low-rated cloud service providers, or the chip suppliers themselves, while the actual end users often exist in a gray area. However, considering the enormous capital requirements, even high-rated companies will rely on project financing markets for some financing plans, especially for business segments that provide computing services externally, albeit to a lesser extent than their low-rated counterparts."
The reality is that the market continues to flood into short-term data center debt, artificially suppressing the yield curve while continuously raising the issuer's guarantee obligations. Structured project financing typically does not fall within mainstream corporate bond indices— the structures are highly customized and overly complex, making them unsuitable for conservative retail funds. Thus, the pricing logic for data center debt is entirely different from that of ordinary credit bonds.
So where do the capital buyers come from? The Goldman Sachs team did not provide a clear answer, only mentioning that "overall, investors find it difficult to significantly overweight." Financing parties need to pay a higher risk premium, which also means that evidence of commercial viability is becoming increasingly important for credit investors.
In our view, the market is caught in a contradiction. Projects must be feasible, and the capitalization rate (annual return rate of data centers) must reach a value-added level. When the coupon rate rises, if borrowers want to maintain their existing debt financing ratio, rents must be adjusted upwards. If tenants refuse to bear the rent increase, the adjustment will shift towards reducing leverage, such as increasing equity input, lowering loan-to-value ratios, and tightening transaction terms. We expect that under high interest rates or other friction factors, investors will demand more conservative transaction structures, and the aforementioned methods will appear simultaneously.
To add a note: Goldman Sachs credit analysts have already doubled their forecast for AI project financing in 2027, but this figure is still nearly 30% lower than the implied capital expenditure estimate from the bank's chip sales group. Who will untangle this logic remains a challenging question.
-- Price
This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.
You may also like

FinCEN Withdraws Crypto Travel Rule Proposal: What Still Applies

TD Cowen Raises Coca-Cola Target Price to $112

This Magnificent Growth Stock Just Entered the Artificial Intelligence (AI) Cloud Race, and It Looks Dirt Cheap

CoinGecko Acquired: Two People Started with $200, Once Valued at $500 Million

4 Chip Stocks Down as Much as 36% From Their Highs to Buy and Hold for the Next Decade

$LIBRA: One of the defendants seeks to remove Judge Ledesma and halt the review that could reactivate the case

Solmate plans AI GPU service with Solana payments: What do we know?

Is Apple Losing Its Memory?

Tech Giants Should Diversify Data Center Investments

Blue Wave Sweeps the U.S., Halting Digital Asset Legislation

Bitcoin: Michael Saylor Hints at New Strategy Purchase

Wall Street Tests the AI Boom: What the Industry's Balances Reveal After the Democratic Victory

Perpetuals on Stocks: SEC Comes to CFTC's Aid Against Cboe and Citadel

The Privacy Paradox of Bitcoin: Satoshi Acknowledged It Long Ago, Have the Successors Solved It?

Prediction: Archer Aviation Will Fly Paying Passengers Before the 2028 Olympics. Here’s What $5,000 Invested Today Could Be Worth by Then.

Liftoff Mobile Stock Rises 30% After Strong Earnings Report

BCRA's Demand for Pesos Fell in October, but There Were Positive Signs: Is Remonetization Starting?

In Anthropic's Doomsday Narrative, There Is No Room for Jensen Huang

China Uncovers AI Projects Worth $3.2 Billion Involving NVIDIA Chips

What Investment Signals Are Released as the 'Big Three' of AI New Cloud Submit Their Reports?

Strive CEO States No Common Stock Issued, mNAV Metric Removed

In-Depth Interview | Fu Peng: I Have No Faith in BTC

Why AMD Shares Are Surging and What a $10,000 Investment in the Stock Could Be Worth in 3 Years

Apple Becomes a Safe Haven Amid AI Sell-off, Bernstein: 37x P/E Still Not Expensive

9-Minute Attack Window, 7.1 Million BTC Exposed: How Bitcoin Can Save Itself from Quantum Threats

Wall Street's Next Blockbuster Stock Split Could Come From This Chip Equipment Giant

Kraft Heinz's Revival Plan: Cheese Ramen and High-Protein Jell-O

Bernstein: Why You Absolutely Shouldn't Buy the Dip on Nvidia Right Now?

Progress in Arbitration of Financial Loan Contract Dispute Involving ST Sanmu








