Milken's Junk Bond Insight Echoes in Today's AI Chip Financing Structures

Michael Milken's foundational work on high-yield 'junk' bonds in the 1970s and 1980s stemmed from his encounter with W. Braddock Hickman's 1958 NBER study, which showed that investors often overestimated risks in lower-rated bonds, leading to superior long-term returns for diversified portfolios. Milken applied this at Drexel Burnham Lambert, growing the public junk bond market from roughly $8.5 billion in 1977 to over $150-200 billion by the late 1980s, financing companies like MCI and Turner Broadcasting.

Contemporary AI infrastructure deals show parallels in creative financing. Amazon is exploring an SPV to shift about $8 billion of installed Nvidia Grace Blackwell chips off its balance sheet, issuing debt and leasing them back while potentially offering investors up to 10% equity. Broadcom has committed to lend Anthropic up to $42 billion via convertible notes to support TPU leases, part of broader patterns where chipmakers backstop customer purchases. Nvidia is involved in platforms potentially backstopping up to $125 billion (with total commitments and guarantees exceeding $500 billion across partners like Apollo, BlackRock, and others), alongside SPVs for data centers leased to entities like OpenAI.

Analyses note AI-related high-yield and leveraged debt issuance reaching $88 billion in 2026, with project finance, asset-backed structures, and off-balance-sheet vehicles proliferating. Milken himself later emphasized that debt suitability depends on stable revenues and warned against it for high-risk tech sectors reliant on equity for growth. Critics highlight risks from rapid GPU obsolescence and concentrated exposures, echoing historical questions about who ultimately holds the risk once paper is sold.

These structures enable massive capex (hyperscalers issuing hundreds of billions in bonds) but raise questions about collateral valuation and systemic leverage, much as Milken's market tested rating agencies' assessments.