What the Bond Market Knows: Credit Is Repricing the AI Buildout Before Equity Does
Spinel Weekly — Industry Note
Published: July 11, 2026 · Subject: Oracle Corporation (NYSE: ORCL) and the debt-funded AI infrastructure buildout
The divergence
Two markets are currently pricing the same company, and they disagree.
Oracle’s five-year credit default swaps have moved from roughly 108 basis points in November 2025 to a record near 198 basis points by April 2026, more than doubling in five months. A spread in that upper range implies, assuming a standard 40% recovery rate, a cumulative default probability in the low-to-mid teens over five years. This repricing began the moment Oracle signed its $300 billion compute agreement with OpenAI in September 2025, and it has not moved in only one direction. Spreads briefly compressed in early February 2026 after Oracle priced a $25 billion financing package that reassured bond investors, only to blow back out to the April record within weeks. That reversal is itself informative: the relief rally lasted exactly as long as it took the market to notice that a financing plan solves near-term liquidity timing, not the underlying counterparty concentration sitting inside the backlog. The wider structural move, not the interim relief, is what has persisted.
A third reference point sat between those two until days before publication. Moody’s still rates Oracle Baa2 with a negative outlook. S&P, which had held BBB with a negative outlook through the entire CDS repricing, cut the rating to BBB- in the week of this note’s publication, the last rung of investment grade, citing rising business risk and weaker cash flow. Alongside the downgrade, S&P now projects a fiscal 2027 free operating cash flow deficit of roughly $42 billion, nearly double its prior estimate. The sequence is the point: the CDS market repriced first, the rating agency followed nine months later, and the equity consensus has done neither. The slow end of the adjustment has started to move. Equity analysts, it turns out, are slower than either.
Equity analysts, meanwhile, maintain a Strong Buy consensus on the same issuer, with a mean price target in the $252–264 range against a share price near $140, an implied upside of roughly 75%. That figure deserves no analytical weight. The bulk of those targets were set between March and June 2026, before the July correction in AI infrastructure names, and the post-correction revisions have only started. A consensus built on a market regime that no longer exists is not a bullish signal. It is a lagging indicator wearing a bullish costume.
The interesting question is not which market is right about Oracle. It is what the credit market’s early move tells us about the entire debt-funded AI buildout, three weeks before Microsoft and Meta update their capital expenditure plans. At issuers under credit stress, the CDS market has repeatedly moved before the equity market. This note takes that pattern seriously.
The mechanics: who actually pays for the buildout
The AI infrastructure trade has quietly changed its funding model. Through 2024, hyperscaler capex was financed almost entirely from internal cash generation. That era is ending. Morgan Stanley expects AI-related debt issuance to roughly double toward $570 billion in 2026 and hyperscaler spending to cross $1 trillion by 2027.
Oracle sits at the sharp end of this shift, not because it runs the weakest business, but because it is the marginal borrower of the trade. The fiscal 2026 numbers, published June 10, describe the position plainly. Capital expenditures rose 162% to $55.7 billion. Free cash flow swung to negative $23.7 billion, from a deficit of just $394 million a year earlier. Management guided fiscal 2027 net capex to roughly $70 billion and announced a further $40 billion raise in debt and equity, on top of the $43 billion of debt and $5 billion of equity raised during fiscal 2026. Total borrowings now stand near $167 billion.
Unlike Amazon, Alphabet, or Microsoft, Oracle cannot fund this buildout from operating cash flow. Every incremental gigawatt of capacity is financed on the balance sheet, in advance, against contracts whose revenue arrives years later. That sequencing is the entire story: Oracle carries the credit risk of its own order book. Which is precisely why the bond market, whose job is pricing balance-sheet risk, moved before the equity market, whose reflex is pricing growth.
The concentration problem
Oracle ended fiscal 2026 with $638 billion in remaining performance obligations, up 363% year over year. The headline is extraordinary. The composition is the risk.
According to Bank of America, more than half of that RPO is attributable to a single customer: OpenAI. The September 2025 compute agreement alone accounts for a reported $300 billion of it; BofA’s higher figure suggests additional commitments beyond that contract. Consider what stands behind that exposure. OpenAI ended 2025 at an annualized revenue run rate of roughly $20 billion, projects revenue of $280 billion by 2030, and expects to burn through more than $650 billion in cash along the way. It has made compute commitments on a comparable scale to several other cloud providers simultaneously, and its IPO, once expected earlier, is now reported by The New York Times to be under consideration for 2027. These are reported figures and reported plans, not audited outcomes; the point is not that they are wrong, but that a receivable of this size currently rests on them.
Oracle itself has told investors as much. The fiscal 2026 annual report, filed June 10, contains unusually specific risk language for a company of this size: it warns that some customers may be highly leveraged and could fail to pay or perform, that certain OCI offerings concentrate revenue among a small number of large customers in ways that amplify this exposure, and that the company may be overestimating demand and might be unable to re-lease or repurpose AI data center capacity if a customer falls away. When an issuer documents the counterparty risk on its own backlog with this level of precision, the credit market’s near-record spread stops looking like pessimism and starts looking like reading comprehension.
The equity framing treats $638 billion of RPO as deferred revenue. The credit framing treats it as a concentrated, unsecured exposure to a pre-profit counterparty, funded in advance with borrowed money. Both descriptions are technically accurate. Only one of them prices the difference between a contract and a collection.
Management’s strongest rebuttal deserves the floor. Oracle discloses that $75 billion of the large AI contracts is already covered by customer prepayments or customer-supplied hardware, which materially reduces the capital it must raise against that portion of the backlog. The mitigant is real, and it is bounded: it covers the hardware, not the buildout. The remaining obligation still requires Oracle to finance construction years ahead of collection, and $75 billion against $638 billion leaves the proportion of the order book resting on future counterparty fundraising largely intact.
The rhyme: vendor financing, 1999–2001
The structure has a precedent. In the late-1990s telecom buildout, equipment vendors extended credit to their own customers to finance purchases of their own equipment. Backlogs hit records at the exact peak of the cycle. When the customers, many of them pre-profit carriers built on capital-market access rather than cash flow, lost that access, the backlogs did not convert. They evaporated, and the vendors were left holding both the receivables and the debt raised to serve them.
The rhyme with today is structural, not cosmetic: a supplier’s balance sheet standing behind its own demand, counterparty concentration among cash-burning customers, and a market narrative that reads backlog size as backlog quality. The differences are equally structural and should be stated plainly. End demand for AI compute is real and monetizing today, not speculative. The largest builders, Microsoft, Alphabet, Amazon, and Meta, fund from operating cash flow at a scale the 1999 carriers never approached. And Oracle’s counterparty, whatever its burn rate, has revenue growing at a pace no 1990s carrier ever achieved. The analogy is a risk framework for one funding structure at the margin of the trade, not a forecast of sector collapse. Used that way, it earns its place. Stretched further, it would be narrative outrunning numbers.
What to watch: the July test
This note publishes three weeks before the thesis gets its first live test. The signals worth tracking, in order of information value:
The hyperscaler capex prints, July 29. Microsoft reports its June quarter on July 29, with Meta scheduled for the same day, and both will update capital expenditure guidance. The number matters less than the funding language. Capex raised and funded from operating cash flow extends the buildout on the old model. Capex raised alongside new debt issuance, or softened with talk of “discipline” and “phasing,” confirms that the funding constraint has become the binding one. Watch the balance sheet commentary, not the headline figure.
The credit tape. Oracle’s five-year CDS and cash bond spreads are now the cleanest real-time gauge of how professional risk-takers price the AI buildout’s weakest funding link. Spread stability through the earnings window would suggest the July equity correction overshot. Further widening, particularly if it spreads to CoreWeave-type pure plays or shows up in the terms of new AI-linked issuance, would suggest it has not finished.
Conversion, not accumulation. From Oracle’s September quarter onward, the metric that matters is the rate at which RPO converts to billed revenue, and the size of customer prepayments against future capacity. A backlog that converts is an asset. A backlog that compounds faster than it converts is a claim on someone else’s future fundraising. Disclosure changes around customer concentration would be the earliest warning of all: companies expand risk language before they restate numbers.
The OpenAI funding calendar. Every reported capital raise, IPO timing update, or new compute commitment shifts the probability distribution under more than half of Oracle’s backlog. This is the uncomfortable core of the trade: the marginal variable behind a $638 billion order book belongs to a private company that publishes no financial statements.
Conclusion
The equity market spent the first half of 2026 debating whether AI infrastructure demand is real. That was the wrong debate; the demand is visible and monetizing. The credit market has moved on to the right one: whether the weakest-funded commitments behind that demand are bankable. Having roughly doubled since November and set a record in April, Oracle’s CDS says the bond market has already answered provisionally. The late-July earnings window will show whether the equity market catches up, or whether the two continue to price different companies wearing the same ticker.
The question for August is not whether the backlog is big. It is whether it is bankable.
This framework now has fixed appointments with reality. Spinel will revisit it after the July 29 hyperscaler prints, and again when Oracle’s September quarter shows the first clean conversion data. Whichever way the signals resolve, the follow-up will say so plainly, including if they resolve against the concerns laid out here. A framework that cannot be revised by evidence is a narrative, and this publication does not sell narratives.
This note is independent research published for informational purposes only. It is not investment advice, and nothing in it constitutes a recommendation to buy or sell any security. Figures are sourced from company filings and press reports as dated in the text.