Citadel Securities projected on August 3 that more than $500 billion of additional public- and private-market debt will be raised by 2028 to finance chips inside AI campuses. The number landed with force, yet it describes an acceleration of something already visible in first-half 2026 filings: Amazon, Alphabet, Meta and their peers have collectively shifted from financing AI infrastructure out of operating cash flow to financing it through a proliferating mix of long-term bonds, finance leases, equipment loans, SPVs and private-credit facilities. The critical change is structural, and its consequences fall on credit-market investors who are still pricing these instruments as though balance-sheet strength alone determines return.
The Cash Flow Inflection That Balance Sheets Are Concealing
The Q2 numbers are unambiguous. Alphabet spent $44.9 billion on capex, reported negative $5.9 billion of free cash flow and raised long-term debt to $98.2 billion, issuing approximately $51.8 billion of new notes in the process. Amazon spent $54.2 billion on property and equipment, ran negative $7.6 billion of trailing-twelve-month free cash flow and nearly doubled long-term debt from $65.6 billion at year-end to $128.9 billion. Meta generated $784 million of quarterly free cash flow after roughly $31.1 billion of capex—a 91% collapse from prior levels—and disclosed approximately $279 billion in future data-center and network lease obligations, plus another $68 billion signed in July carrying 18–20-year terms. Microsoft remains the exception: it expects to stay free-cash-flow positive through fiscal 2027 and represents the strongest internal case for the bull thesis.
Morgan Stanley estimated AI-related global debt issuance at nearly $236 billion through May 31, 2026, projecting roughly $570 billion for the full year. Investment-grade technology paper now represents approximately 15% of all US corporate issuance year-to-date, a 1,300% year-over-year jump. Subscription coverage for hyperscaler deals has fallen from roughly 5x in February to below 2x in July. Ten-year AI-related bonds were trading around 121 basis points over Treasurys versus roughly 80 basis points for the broader investment-grade market—an approximately 41-basis-point premium—and CDS on Alphabet, Amazon and Meta sat near 49 basis points, the widest since 2018.
A Prisoner's Dilemma Wearing a Return-on-Investment Costume
The standard framing—that strong cloud revenue validates the spend—conflates technological monetization with capital-provider economics. AWS grew 37% in Q2, its fastest rate in 18 quarters. Google Cloud grew 82% with backlog reaching $514 billion. Azure grew approximately 43%. Capacity is being consumed. The question credit investors must answer is whether dollar revenue outpaces dollar cost at the asset level, not whether aggregate cloud markets are growing.
The spending is partly prisoner's dilemma, not pure demand response. Each firm is investing because underinvestment risks losing developers, enterprise workloads and model distribution to whoever builds first. That asymmetric payoff structure makes individually rational overinvestment collectively probable regardless of whether aggregate returns justify the capital deployed. Telecommunications operators in 1999 and 2000 understood this dynamic intimately; the Internet continued growing for decades while overleveraged carriers restructured.
The accounting obscures the exposure further. Off-balance-sheet structures—SPVs, take-or-pay capacity contracts, finance leases—reduce reported leverage without eliminating the obligation. Meta's $347 billion in disclosed lease commitments, for instance, sits largely outside the standard net-debt calculation most investors apply to the credit. The BIS has described these as "shadow borrowing," noting linkages through bank warehouse lines, subscription facilities and guarantee chains that can transmit stress across vehicles that appear separately capitalized.
Why Duration Is the Operative Risk, Not Default
Spreads do not need to signal distress for long-dated AI bonds to deliver poor risk-adjusted returns. A bond can repay at par after years of underperformance relative to where it priced—and the supply math is hostile. Combined 2026 capex guidance from Alphabet, Amazon, Microsoft and Meta runs roughly $735–760 billion, including non-AI expenditure. As that spending enters bond benchmarks, investment-grade portfolio managers face growing sector concentration irrespective of their individual security preferences.
The core mismatch—and the analysis's sharpest observation—is temporal. Microsoft disclosed that approximately two-thirds of recent capex went into short-lived GPUs and CPUs, with the balance supporting assets expected to monetize across 15 years or longer. Amazon shortened server and networking useful lives from six to five years because AI development is accelerating. Meta previously extended certain server lives to 5.5 years. There is no universal accounting life for AI compute. The relevant variable is whether revenue per unit of compute falls faster than the combined weight of capital cost, power expense, lease obligations and financing charges amortize—and that race is not settled by revenue growth headlines alone.
The Structural Takeaway That Rewrites the Investment Thesis
Sophisticated credit investors are already voting with their structures. Apollo's framework for AI project debt demands shorter amortization, triple-net leases and unconditional payment obligations. CoreWeave's $2.6 billion loan—backed in part by Anthropic contracts—cleared only after yields approached 9.1% and lenders secured a lockbox directing contract revenue toward repayment before general creditors could access it. Nvidia has discussed guaranteeing as much as $250 billion of OpenAI obligations tied to an Ohio data-center project. Apollo and Blackstone have worked on roughly $36 billion of debt to finance Google TPUs for Anthropic, with Broadcom reportedly backstopping major payment tranches.
These structures reveal the correct credit thesis. Secured data-center construction debt—some project bonds reportedly offering 50–200 basis points above unsecured hyperscaler paper—can carry better downside architecture than the long-dated investment-grade bonds that dominate mainstream portfolio allocations. Project creditors hold collateral over identifiable assets and can enforce amortization schedules tied to construction or lease milestones. Unsecured 10–30-year hyperscaler bondholders hold a contractual claim on a company whose incremental AI investments may or may not earn their capital cost, with no direct recourse to the financed assets and full exposure to future capital-allocation decisions by management.
The dominant thesis circulating in fixed-income markets treats creditworthiness and bond attractiveness as equivalent. They are not. A company can remain highly solvent while its marginal investments destroy economic value and its bonds underperform duration-matched alternatives for years. The hyperscalers' AI infrastructure build will very likely produce enormous technological value. Whether the long-dated debt financing it produces acceptable returns for bondholders is a separate calculation—one that the 41-basis-point spread premium and collapsing deal coverage ratios suggest the market is only beginning to price.
not investment advice
Sources: https://x.com/Polymarket/status/2084297329545294085