Core Scientific’s $3.3B Junk Bond Exposes AI Infrastructure’s Financing Fracture
A 7.933% yield on data center debt reveals the unstable economics of GPU-backed financing as Big Tech's $650B capex spree migrates from equity to bond markets.
Core Scientific raised $3.3 billion in high-yield debt on 22 April 2026 at a 7.933% yield—one of the highest borrowing costs for data-center-linked junk bonds this year—exposing a bifurcated financing model where smaller AI infrastructure operators refinance at double-digit spreads while hyperscalers access investment-grade capital for identical hardware risk.
The deal, structured around a 12-year lease to CoreWeave worth approximately $10 billion, marks the CoinDesk-reported pivot of a company that emerged from Chapter 11 bankruptcy in January 2024 from crypto mining to GPU hosting. The transaction sits within a broader explosion of AI Infrastructure debt: $17.9 billion raised in junk bonds for Data Centers year-to-date in 2026, more than double the $7 billion issued across all of 2025.
What makes Core Scientific’s yield notable is timing. While Amazon, Alphabet, Meta, Microsoft, and Oracle have collectively guided toward $650-720 billion in combined 2026 capital expenditure—a 65-75% increase from 2025’s $381 billion, per Bloomberg—these hyperscalers issued $121 billion in bonds last year at investment-grade rates. Core Scientific, hosting the same Nvidia H100 and H200 GPUs for the same workloads, pays nearly 8% for five-year money.
The GPU Depreciation Problem
The 7.933% yield reflects a harsh reality in GPU-backed financing: collateral that depreciates 30-60% in year one with no standardised secondary market. Asset-backed lenders impose 40-50% haircuts and charge 12-15% interest rates on GPU financing, according to Introl research, despite H100 rental prices collapsing from $8 per hour in early 2024 to $2-3 per hour by late 2025—a 60-70% decline documented by Dave Friedman.
No forward curve exists for compute pricing. GPU residual value carries no standardised hedging instruments. When OpenAI signalled its exit from Oracle’s Stargate facility in Abilene, Texas in March 2026—preferring newer GPU generations—it illustrated the contract-termination risk during obsolescence cycles that makes five-year debt instruments perilous against three-year hardware refresh cycles.
“There is a speculative edge to this form of lending, with no history to refer back to. Data centre lending represents a small percentage of the US high yield index currently, but we expect more issuance in coming years.”
— Vontobel Manager, quoted in Portfolio Adviser
Moral Hazard in Two-Tier Markets
The financing bifurcation creates systemic fragility. While Core Scientific borrows at 7.933%, Alphabet-backed Fluidstack raised a record $5.7 billion junk bond in April 2026 for identical data center infrastructure, per Bloomberg. Combined with CoreWeave’s concurrent offering, Google-linked entities alone raised $6.7 billion in high-yield markets within weeks.
Meanwhile, the hyperscalers funding these facilities access bond markets at spreads 500 basis points tighter. Oracle—despite carrying a 500% debt-to-equity ratio and $100 billion in balance-sheet debt—maintains investment-grade ratings even as its credit default swap spreads widened to 125 basis points, three times wider than Microsoft’s, according to analysis by Implicator.ai. Barclays projected Oracle would exhaust cash by November 2026 in a December 2025 research note.
Big Tech’s capital intensity reached 45-57% of revenue in 2026 guidance, versus historical software industry norms of 5-15%. This represents a fundamental shift from equity-funded R&D to debt-funded infrastructure speculation. CNBC reported that AI-related bonds now constitute roughly 30% of investment-grade corporate issuance.
Cash Flow Compression
The debt binge coincides with free cash flow deterioration across hyperscalers. Meta is modeling negative free cash flow for 2027 and 2028, while Amazon faces $17-28 billion negative FCF in 2026, representing up to 90% declines from prior years. “We are now modeling negative FCF for ’27 and ’28, which is somewhat shocking to us but likely what we eventually see for all companies in the AI infrastructure arms race,” Barclays equity research told CNBC in February 2026.
Microsoft’s AI revenue target for fiscal 2026 sits at $25 billion against estimated capex of $97.7-150 billion—implying multi-year payback periods on infrastructure that may be technically obsolete before generating positive returns. JPMorgan projects $1.5 trillion in AI-related bond issuance will be needed over the next five years, with Morgan Stanley estimating $900 billion in 2026 alone from global technology companies, per Mellon Investments analysis.
- GPU collateral depreciates 30-60% annually with no liquid secondary market or hedging instruments
- Five-year debt maturities mismatched against three-year hardware refresh cycles
- Hyperscaler free cash flow projected to decline 90% in 2026 despite investment-grade ratings
- AI revenue monetization remains unproven while $1.5T bond issuance pipeline builds
The Unspoken Contract Breaks
“For years, we’ve been told this AI spend would be funded by generated cash flow—that it is equity risk, it is speculative, and not to worry about it from a credit point of view. There now seems to be a change in the unspoken contract,” Al Cattermole, fixed income portfolio manager at Mirabaud Asset Management, told CNBC in February 2026.
The shift from equity-funded to debt-funded AI infrastructure represents a transfer of speculative risk from shareholders to bondholders. Columbia Threadneedle Investments notes that “the economics of AI infrastructure at scale remain unproven. The ratio of incremental revenue to capex has declined, and the timing of revenue realisation is uncertain.”
Data center securitization issuance reached $17.4 billion year-to-date in 2025, up 50% from 2024’s full-year total. But these structures assume stable cash flows from long-term leases—the same assumption tested when OpenAI walked from Oracle’s facility despite contractual commitments.
What to Watch
The financing model faces three near-term stress tests. First, whether Core Scientific and similar operators can refinance 2027-2028 maturities if GPU rental rates remain compressed—H100s that commanded $8 per hour 24 months ago now struggle to hold $3. Second, whether rating agencies maintain investment-grade ratings for hyperscalers projecting negative free cash flow through 2028 while carrying record debt loads. Oracle’s November 2026 cash exhaustion timeline, if accurate, would force the first high-profile refinancing at distressed spreads.
Third, whether the $230-240 billion in projected 2026 public debt issuance for AI infrastructure finds buyers if early revenue disappointments emerge. Microsoft, Meta, and Alphabet must begin demonstrating AI monetization at scale by late 2026 to justify capex-to-revenue ratios that have inverted historical software economics. As BofA’s head of global credit Matt McQueen noted: “The numbers are like nothing any of us who have been in this business for 25 years have seen. You have to turn over all avenues to make this work.”
The 7.933% yield Core Scientific paid on 22 April 2026 may mark the canary: the rate at which GPU-backed infrastructure debt must clear when stripped of hyperscaler credit ratings. If AI returns disappoint or capital costs rise, that 8% spread becomes the true cost of compute—and the $1.5 trillion bond pipeline reprices accordingly.