AI Bond Buyers Now Sort Issuers by Business Model — Oracle's 18-Year CDS High Marks the Crack

The supply story is no longer news. By August 2026, the five large hyperscalers had collectively issued more than $360 billion of US-dollar debt since September 2025 — more than doubling their USD bond footprint in less than a year — and now represent roughly 5% of the US-dollar investment-grade index (PIMCO, Aug 10 2026). Vanguard estimates full-year 2026 AI-related issuance across the broader ecosystem at $300-570 billion, with hyperscaler capex projected to approach $800 billion this year and exceed $1 trillion annually from 2027 through 2030.

Through 2025, every AI-debt story was really a single trade: hyperscalers issuing investment-grade paper into a hungry market, with spreads compressing in lockstep. In 2026, the lines cracked. Hyperscaler composite spreads widened roughly 30 bps year-to-date while the broader USD IG index widened only 2 bps (J.P. Morgan Asset Management, Aug 20 2026), and the gap between hyperscalers and the index stretched from 20-25 bps at the start of 2025 to as much as 50 bps in the first half of 2026.

The clearest evidence sits in the names that have fallen out of bed. Oracle's 5-year CDS hit 215 bps in late July, an 18-year high, against a peer hyperscaler average near 80 bps (Reuters, Aug 4 2026). Broadcom's 5-year CDS jumped 28 bps in August — a larger move than Oracle's or SpaceX's in the same window — as the market began pricing in $60-100 billion of off-balance-sheet AI chip debt (Bloomberg Tax, Aug 24 2026). Yet Amazon's €25 billion jumbo priced flat, forcing only 20 bps of concession on its existing 30-year bonds (Sage Advisory, Jul 16 2026). Three different reactions to the same theme.

The split runs along three predictable lines: how diversified the cash flow is, how much of the capex sits on- versus off-balance-sheet, and how concentrated the AI customer is. Oracle was already a problem child before the AI bond wave started — Insight Investment noted in January 2026 that its bonds had been trading 100-150 bps wider than other hyperscalers since the 2022 A-to-BBB downgrade. The rest of the IG market kept treating Oracle as a generic hyperscaler until S&P moved it to one notch above junk on July 9, 2026, then the spread gap became impossible to ignore.

For Oracle, the spread split is now mechanical. Fiscal-2027 capex guidance of $90-95 billion — up 63% from the $55.7 billion spent in fiscal 2026 — points to a $42 billion free-operating-cash-flow deficit in fiscal 2027 (S&P Global Ratings, Jul 9 2026), and $25 billion of customer repayments are expected to flow back into the cash stack. Even with that offset, S&P forecasts adjusted leverage in the mid-4x area. The discrimination has therefore moved from bond traders' preferences into rating-agency math.

Broadcom is the more interesting case because most of its AI debt is invisible. The chip designer is in talks with a group of lenders to raise more than $60 billion — potentially $70 billion to $100 billion across all structures — through a special-purpose vehicle that finances custom AI chips for Anthropic, with Apollo Global Management and Blackstone acting as the lead lenders alongside global banks (Bloomberg/Reuters, Aug 20-21 2026). Broadcom would guarantee part of the senior tranche and provide 'residual value support' if the AI customer stops paying and repossessed chips sell for less than the balance owed. The earlier $35 billion AI XPV Platform transaction closed on June 9, 2026 with Apollo Global Management and Blackstone leading the syndicate (Apollo press release, Jun 9 2026).

Broadcom's 5-year CDS jumped 28 bps in August — a larger move than Oracle's or SpaceX's — and yields on its 5.15% bonds due 2031 climbed about 14 bps in the same window (Bloomberg Tax, Aug 24 2026). The bond market has moved from treating Broadcom's AI exposure as a forward-looking sales pitch to treating it as a contingent liability that has to be marked. On Broadcom's FY2025 numbers ($63.9 billion of revenue, $34.7 billion of EBITDA, $3.21 billion of interest expense), the on-balance-sheet debt service is rounding error — but the AI XPV backstop is not, and the market is now demanding 28 bps of CDS protection for it.

Above the hyperscalers sit the neoclouds — CoreWeave, Nebius, and others — which lease GPU capacity to the same frontier-model labs the hyperscalers themselves are building for. The funding stack is different: not public IG, but convertible bonds, GPU-backed term loans, and asset-based facilities. CoreWeave's long-term debt reached $27.8 billion at June 30, 2026, up 274% year-over-year (Macrotrends, Q2 2026 10-Q), against $2.58 billion of Q2 revenue and a net loss of $626 million. The $2.6 billion delayed-draw term loan closed in August at SOFR plus 5.50% — almost triple the rate Nebius pays on similar facilities (Dealroom, Sep 14 2026). Full-year 2026 capex guidance of $30-35 billion means interest expense keeps compounding against an already-thinning operating margin.

The risk bond buyers are pricing in the public IG market is already adjacent to private credit — but the private market has not yet started to differentiate. Business-development companies — the listed funds that have been the marginal lender to U.S. middle-market software for a decade — had accumulated $115 billion of software loans by mid-2026, roughly a fifth of all their lending and over 80% of their fast-growing technology portfolios (BIS Bulletin 128, Jul 14 2026). By Q1 2026, software represented 32% of BDC portfolio holdings at both cost and fair value (Octus, Jul 27 2026).

The discrimination is also visible across borders. AI hyperscalers are roughly 5% of the USD IG index but only 1.2% of the EUR IG index (PIMCO, Aug 10 2026), so the supply pressure sits squarely with dollar buyers. 2026 year-to-date net supply of index-eligible AI hyperscaler debt is $120 billion in USD against €22 billion in EUR. USD IG dispersion — the share of issuers whose spreads widen more than 30 bps over two months — has reached levels last seen during the April 2025 'Liberation Day' tariff shock; EUR IG dispersion has barely moved (PIMCO, Aug 10 2026).

Three things should move in the next few quarters. More rating actions: Oracle sits one notch above junk, and the agencies have no template for Broadcom's off-balance-sheet AI structure. More concessions on each hyperscaler jumbo: every $20 billion-plus issuance in 2026 has pressured existing bonds 15-20 bps wider before stabilizing (Sage Advisory, Jul 16 2026). And BDC NAV marks: BIS finds software BDC loans have not yet been repriced for AI exposure, but 35% of software principal matures in 2028-2029 (Octus, Jul 27 2026).

Over a 1-3 year horizon, the picture depends on whether hyperscaler free cash flow recovers on schedule. J.P. Morgan Asset Management expects hyperscaler gross debt issuance to rise into 2027 and peak around 2028, then improve as data-center build-out matures and capex growth levels off toward 2030 (JPMorgan, Aug 20 2026). If the recovery is delayed — by capex overruns, AI demand disappointment, or rates staying higher for longer — the spread split widens: Tier 1 issuers gain a relative-value trade, Tier 2 names test the fallen-angel cliff, and Tier 3 names become the speculative-grade casualties of the cycle. The roughly $300 billion of IG bond supply expected over the next 12 months sets the floor for how much dispersion the market can absorb.