The narrative around AI investment is shifting. It began as a technology story, evolved into a capital expenditure story, and is now becoming a financing story.
Recently, Bridgewater's Greg Jensen summarized the current situation in one sentence: "We are entering a critical moment for capital."
Morgan Stanley predicts that total spending on AI infrastructure construction will reach $3.2 trillion by 2028, with approximately $1.75 trillion needing to be raised through the credit markets. Funding sources have expanded from traditional investment-grade bonds to leveraged loans, private credit, and securitized products. Money is available, but it won't be cheap.
The bond market has already started to react. Apollo data shows that AI-related bond issuance now accounts for 40% of long-duration supply. Credit spreads for hyperscale cloud providers (Hyperscalers) have widened significantly this year, while the overall investment-grade market has remained virtually unchanged.
The "new bond king" Gundlach was even more direct, stating that issuing long-term bonds backed by GPUs is akin to "making a 30-year ABS out of bananas."
Cash Flow is Moving in the Wrong Direction
Capital expenditure forecasts are repeatedly revised upwards, while free cash flow forecasts continue to slide.
Morgan Stanley has significantly lowered its free cash flow forecasts for major hyperscale cloud providers for 2027. Oracle's situation is the most prominent, with its 2027 free cash flow forecast now approaching -$40 billion.
The scale of spending commitments is equally staggering. Purchase commitments from hyperscalers have exploded to $982 billion. Morgan Stanley points out that a significant portion of the real capital expenditure now exists off-balance-sheet. This means that looking only at the balance sheet severely underestimates the actual funding pressure.
Credit Markets Have Begun Pricing Risk
Investors are demanding higher compensation for risk.
Morgan Stanley data shows that credit spreads for hyperscale cloud providers have "widened significantly this year, far exceeding the overall investment-grade market"—spreads for high-quality hyperscalers (HQ Hyperscalers) widened by about 25 basis points, regular hyperscalers widened by about 22 basis points, while the overall investment-grade market spread changed by 0.
Oracle is the most watched case among them. While other hyperscalers' credit default swaps (CDS) generally trade in the 30 to 80 basis points range, Oracle's CDS surged from around 40 basis points in mid-2025 to a peak near 190 basis points in April 2026, and currently hovers around 180 basis points. Meta's CDS has risen more modestly to about 75 basis points. The credit market has clearly singled out the company it is most concerned about.
Balance Sheets are Strong, But the Problem Lies Ahead
Morgan Stanley's Q1 2026 data shows that hyperscalers' total leverage ratio is only 1.3x, net leverage is 0.5x, the cash-to-debt ratio is as high as 128%, and the median credit rating is AA-. In contrast, the overall non-financial investment-grade universe has a total leverage ratio of 2.4x and a rating of BBB.
But the problem lies ahead. Morgan Stanley has sharply raised its forecast for cloud computing capital expenditure growth in 2027 from 14% to 29%, while hyperscalers are "reaffirming confidence in investment returns." As spending forecasts double, the financing gap widens accordingly.
"New Bond King" Gundlach: Using GPUs as Collateral is Like Using Bananas for ABS
Regarding financing schemes emerging in the market that use AI assets as collateral, renowned bond market investor Jeff Gundlach offered sharp criticism.
He warned about a planned $50 billion fund consortium, stating the plan is "likely not to stand the test of time." He asked on social media: "Using assets with unknown life spans as collateral for long-term debt? Why not do a 30-year ABS trade backed by bananas in a warehouse?" He added, these are "brand new engineered bananas with unknown life spans."
Gundlach's analogy gets to the core issue: GPUs depreciate extremely quickly, with technology refresh cycles far shorter than the tenure of the debt. When using such assets as collateral for long-term financing, the actual value of the collateral is highly uncertain.
