One Market Already Priced In AI Failing

Stocks price hope. Credit prices fear of loss. Both markets are watching the same AI buildout and disagreeing, because they get paid in opposite shapes. A shareholder who is right about a data-centre bet can make several times their money. A lender who is right gets par plus a coupon, which is what the contract promised anyway. The most a lender can win is the agreed amount. The most a lender can lose is all of it. That asymmetry is why credit desks register trouble before equity investors accept there is any.

In late July 2026, insuring ten million dollars of Oracle’s debt against default for five years cost roughly $219,000 a year, the most expensive that protection has been in at least six years. This article takes that number apart: what it measures, what S&P published as the trigger for the next downgrade, and the $260 billion of data-centre leases that have not started and therefore do not appear as debt.

The nine-minute video version is above. Everything in it is written out below, so the text stands on its own.

What 219 basis points actually buys

A credit default swap is shaped like insurance. The buyer pays an annual premium, quoted in basis points of the amount covered and settled quarterly. If the reference company hits a defined credit event, the seller makes the buyer whole on the covered debt. Standardised contracts now trade with a fixed coupon plus an upfront payment, so the quoted spread is best read as the equivalent annual cost rather than a literal invoice.

The arithmetic is direct. A spread of 219 basis points on $10 million of notional is 2.19% a year, or $219,000. At 100 basis points the same protection would cost $100,000. Nothing about the contract changed between those two prices. What changed is what somebody is willing to charge to stand behind the debt.

That is why credit is a fast signal. Buying protection costs cash every quarter and needs a counterparty willing to take the other side at an agreed price. A spread is a position, not an opinion.

Converting the spread into a default probability is tempting, and the rough translation is worth knowing along with its limits. The standard approximation divides the spread by one minus the assumed recovery rate. Assume a 40% recovery, a convention rather than a measurement: 219 divided by 60 gives an annual hazard rate near 3.65%, compounding to roughly 17% cumulative over five years. Spreads also pay the seller for illiquidity, for funding the position, and for being wrong at the worst moment, so the priced number sits above the odds a statistician would quote. It is the price of the fear, not a forecast.

The rating that leaves nothing below it

S&P Global cut Oracle to BBB-/A-3 on 9 July 2026. BBB- is the last investment-grade rung on the long-term scale. The next notch down, BB+, is high yield.

The reason that boundary matters is structural rather than reputational. A large share of the buy side cannot hold sub-investment-grade paper: index funds tracking investment-grade benchmarks, and insurance and pension mandates written against ratings. If an issuer falls out of the investment-grade indices, the funds tracking them have to sell, and the selling is not a judgement about the company. It is a rule being followed. The price moves because of who is no longer permitted to own the bonds.

S&P also published what would trigger the next action: leverage above 4.5x EBITDA. That ratio is net debt divided by annual earnings before interest, tax, depreciation and amortisation, and it can be breached from either side.

Take a hypothetical issuer, chosen for round numbers rather than to describe anyone: $100 billion of net debt against $25 billion of EBITDA sits at 4.0x. Two paths cross 4.5x. Borrow another $12.5 billion with earnings flat, or hold debt flat and let EBITDA slip to $22.2 billion, a shortfall of about 12%. In the middle of a capital expenditure cycle funded by borrowing, both variables move at once, and they move in the same direction. That is why a published threshold is more useful than a published outlook. It converts a rating decision into a number you can track yourself.

The $260 billion that is not debt yet

Reuters put the lease burden built by the AI data-centre race near $1 trillion, of which roughly $260 billion covers leases that have been signed but have not commenced.

The accounting is the whole point. Under US lease rules, a lessee recognises a right-of-use asset and a matching lease liability at the commencement date, meaning when the asset is actually made available for use. A lease that is signed but has not started is disclosed in the footnotes as a commitment and is not carried on the balance sheet. Nothing is being hidden. It is simply not yet recognised.

So a leverage ratio read off reported debt is incomplete. Credit analysts add committed leases back before comparing issuers. A screen that reads the balance sheet alone does not.

The timing is the part worth sitting with. Those leases commence on a schedule, arriving as liabilities and as rent expense on dates already agreed. The revenue meant to justify them arrives on no schedule at all. The cost is contractual and dated. The income is a forecast. Sequencing risk is what credit analysts price when they cannot price the business.

The quarter the cash flow turned

Alphabet reported its first negative free cash flow quarter since 2004 in the second quarter of 2026.

Free cash flow is operating cash flow minus capital expenditure, which makes it the measure of whether a company can fund its own buildout. A large cash pile does not change the mechanism. When capex runs ahead of operating cash flow, the gap is closed by borrowing, and the borrowing has to be sold to somebody.

That is the volume side of this story. Morgan Stanley expects roughly $570 billion of AI-related debt issuance across 2026. Debt of that size does not clear by assumption. It clears at a price, and the price is set deal by deal.

The early warning almost nobody watches

Apollo tracks the cover ratio on new hyperscaler bond deals: how many dollars of investor orders arrive for every dollar of bonds on offer. A cover ratio of 3x means the deal was three times oversubscribed. On Apollo’s figures the ratio ran near 5x in February 2026 and fell below 2x by July.

Read that carefully, because it is not a failure. The deals cleared. What shrank was the queue behind them. When the order book thins, the issuer offers a larger concession to fill it, which means a higher yield, which raises the cost of capital on every deal after it. Reuters reported exactly that pattern on 29 July: a hyperscaler debt binge pushing yields up as investor demand cooled.

Cover ratios lead spreads for a simple reason. A spread is a mark on existing paper. A cover ratio is measured at the moment new paper changes hands, deal by deal, before any of it settles into an index.

The six numbers on one page

Number	Level, August 2026	What it measures	What a move would say
Oracle 5-year CDS spread	219 bps	Annual cost to insure the debt	Wider means a higher price to stand behind it
S&P rating	BBB-/A-3	Last investment-grade rung	One notch down forces index-mandated selling
Published leverage threshold	4.5x EBITDA	S&P’s stated trigger	Breached by more debt or weaker earnings
Leases signed, not commenced	~$260 bn	Obligation not yet on the balance sheet	Commencement turns footnote into liability
Expected 2026 AI debt issuance	~$570 bn	Supply that must find buyers	Volume sets the price of the marginal deal
Hyperscaler bond cover ratio	Below 2x, from near 5x	Order book per dollar issued	Thinner books raise the next deal’s yield

Where this reading goes wrong

A spread is not a probability. It is a price that includes compensation for risk, illiquidity and funding. Quoting a 17% implied default chance as an estimate of default confuses the two.

A downgrade is not a default. BBB- companies overwhelmingly pay their debts. Crossing into high yield forces selling by mandate-bound holders, which is not insolvency.

One issuer is not the sector. Oracle carries an enormous buildout against a smaller earnings base than the cash-rich hyperscalers, which is why its CDS is the loudest print on the screen. Loudest is not representative.

Equity agreement is not the test. Credit and equity can disagree for a long time and either can be wrong. The claim is not that bond investors are smarter, only that they are paid differently and notice different things first.

Single-name CDS can be thin. A large move on light volume may be positioning, not information. Check the single name against a broad index first.

Not on the balance sheet is not the same as not owed. A commitment disclosed in a footnote is still a contract.

Three things you can check without a terminal

Read the rating action, not the headline about it. Agencies publish their decisions, their outlooks and the thresholds that would move the rating again. The 4.5x figure came from a public document, not a data subscription.

Open the lease footnote and the cash flow statement. In any 10-Q or 10-K, search for leases that have not yet commenced and compare that figure against reported debt. Then take operating cash flow, subtract capital expenditure, and see whether the buildout is self-funded this quarter.

Watch a spread index alongside the single name. FRED publishes the ICE BofA BBB US corporate index option-adjusted spread daily and free, and FINRA’s TRACE shows actual corporate bond trade prices and yields. If one issuer widens while the index sits still, the story is about that issuer. If both widen, it is about credit conditions.

What would settle this the other way

If leases commence alongside the revenue that pays for them, if cover ratios recover toward early-2026 levels, if leverage stays under the published threshold and spreads compress back, then credit was pricing a risk that did not arrive. That happens constantly. Most priced probabilities never occur, which is the point of pricing them.

The narrower claim survives either outcome: one market has already put a number on this buildout going wrong, the number is public, and the inputs behind it can be checked by anyone willing to read a footnote.

Sources

Every figure above is as of August 2026 and comes from a primary source:

This channel and this blog are for education and analysis. Nothing here is investment advice. The hypothetical leverage example uses round numbers chosen for illustration and does not describe any specific company’s accounts. Figures are as of August 2026 and can change.

Music in the video: “Crypto” by Kevin MacLeod (incompetech.com), licensed under Creative Commons: By Attribution 4.0.