How a Credit Default Swap Actually Works — and What Its Price Is Telling You

Last week, the number that defined the AI-debt scare wasn't a stock price. It was the cost of protecting CoreWeave's debt against default: more than $1 million a year for every $10 million covered, a level generally seen as a sign of distress. We quoted it ourselves.

Here's what that number leaves out. CoreWeave isn't part of the contract it comes from, and never sees a cent of the money. You don't need to own any of its debt to buy the protection. Nobody actually pays that $1 million as a yearly bill. And the odds of default some reports attached to it aren't anyone's forecast.

The simple version, and what it skips

The simple version: a credit default swap, or CDS, is insurance on a loan. The lender pays a premium, and if the borrower goes bust, the insurer pays out.

That isn't wrong. But it skips the parts that make the headline number so easy to misread: who's allowed to buy, what actually gets paid, who decides whether there's a payout, and what besides fear of default can push the price up.

The Money Map

Notice where the borrower sits: outside the box.

Watch the machine work

Say a bank has lent $10 million to a company building AI data centers. It likes the loan. It's also losing sleep over it.

1. The bank buys protection. It calls a dealer and asks for five years of protection on $10 million of the company's debt. Five years is the standard length. The dealer finds an investment fund willing to sell it.

2. They agree a price, and here's the first surprise. The market quotes the company at 1,000 basis points. A basis point is a hundredth of a percentage point, so that's 10% a year, or $1 million a year on $10 million. That's the number in the headlines. It's not what the bank pays.

Since 2009, the yearly fee on North American single-company contracts has been fixed at one of two round numbers: 1%, generally for investment-grade borrowers, or 5% for riskier ones. Our company is risky, so the bank pays 5%: $500,000 a year, about $125,000 a quarter. To make up the gap between 5% and the market's 10%, it also pays the fund a lump sum on day one: roughly $1.5 million, by our estimate using the market's standard assumptions.

Why not the full $500,000 a year for five years, or $2.5 million? Because money paid today is worth more than money paid later, and the fees stop early if the company defaults. Allow for both, and five years of missing fees come to about three years' worth today.

3. Time passes, and the worry moves. If the company's spread jumps, the protection the bank owns is worth more, and the fund typically has to post collateral to cover it. If the worry fades, the collateral flows back. Many contracts never get past this step: the borrower doesn't default, or the position is closed out along the way.

4. The company misses a payment. The bank asks the region's Determinations Committee to rule on whether a credit event, the contract's definition of default, has happened. A ruling needs an 80% vote; if the committee can't reach that, a panel of outside reviewers decides.

5. An auction prices the damage. Dealers bid for the company's bonds, and the auction sets a final price: the market's verdict on what the debt is still worth. Say it's 40 cents on the dollar.

6. The fund pays the gap. The fund pays the bank the difference between face value and the auction price: $6 million on $10 million of protection. The quarterly fees stop.

The payout follows the auction price, not the bank's own loss on its loan, which may turn out bigger or smaller than the payout. And the bank has been paying for the protection all along.

The real thing runs the same way. When Lehman Brothers' contracts were settled in October 2008, the auction priced its bonds at 8.625 cents on the dollar, so sellers paid about $9.1 million for every $10 million of protection.

Now swap the bank for a buyer that owns no debt

Here's the second surprise. Run the same contract again, with one change: this time the buyer has never lent the company a dollar.

Every number is identical: the same 5%, the same lump sum. If the company defaults, the new buyer collects the same $6 million, except now there's no loss to offset. If the worry simply grows, it can close out the contract at a gain without anyone missing a payment.

It can be a side bet on someone else's loan, but not always. A buyer without the debt might be hedging a related exposure, or trading one borrower against another. Or it might simply be betting the borrower's credit will get worse. In one Banque de France sample of French financial firms' CDS on companies, direct hedging accounted for only about a fifth of the protection bought.

When the machine breaks

A CDS doesn't make risk disappear. It moves it. Our bank swapped the risk that the company won't pay for the risk that the fund won't.

The seller can't pay. Before 2008, AIG sold enormous amounts of protection on mortgage-linked securities. When those securities lost value and AIG's credit rating was cut, buyers demanded collateral. The Financial Crisis Inquiry Commission concluded AIG failed mainly because it had sold the protection without putting up initial collateral, setting aside capital or hedging its bets. The government committed $182.3 billion to stabilize it, and later got it all back with a profit. Today the most standardized contracts go through clearing houses that collect collateral from both sides; more than half of CDS, measured by the amount of debt covered, are cleared this way.

The referees may have a stake. The committees that decide whether a default counts are made up of firms that trade the contracts they rule on. A 2024 review commissioned by ISDA, the derivatives industry's trade body, flagged concerns about opacity and potential conflicts of interest. S&P Global is expected to take over administering the committees this autumn.

Reading the price

Back to the 1,000 basis points our bank was quoted. What was it pricing?

A rule of thumb: the spread is roughly the yearly chance of default multiplied by how much of the debt would be lost. Run that backwards and you can turn a spread into odds yourself, using the market's standard assumption that lenders get back 40 cents on the dollar of senior unsecured debt after a default:

• 100bp works out to roughly an 8% chance of default within five years
• 500bp to roughly 34%
• 1,000bp to roughly 57%

That arithmetic is how CoreWeave's record of more than 1,100bp last Thursday came to be described as implying about a 63% chance of default over five years, in a Reuters report citing S&P Global Market Intelligence data.

Here's where some of last week's names sat. On Monday, CoreWeave's five-year CDS stood around 1,080bp and Oracle's around 315bp, according to CNBC. Meta's was near 105bp last Thursday, according to Reuters.

Treat those odds as a price, not a forecast:

1. They include a fee for carrying the risk, so market-implied odds tend to run above real-world default probabilities, as research published by the BIS found.
2. They lean on the 40-cent assumption. Real auctions swing far wider: Atos's bonds were priced at 3 cents on the euro in 2024, Intrum's at 76 cents in early 2025. Assume a lower recovery and the same spread implies lower odds.
3. They price the contract's definition of default, as a committee applies it, which isn't always what "default" means in everyday English.
4. They come from a thin market. Dealers and their clients trade privately, and even a large company's contracts sometimes change hands fewer than 10 times a day, Reuters notes. Data providers disagree, too: last Thursday, CoreWeave was quoted anywhere from about 1,040 to about 1,195bp, depending on the provider and the time of day.

The part nobody explains

A spread measures what people will pay for protection. That depends on how badly they need it, not only on how likely a default is.

Most of the time the two move together. But the spread is the price of handing off the worry, and like any price, it can climb simply because a lot of people need the same thing at once. Remember our bank, buying cover on a loan it wanted to keep? In AI, it has company: some of the keenest buyers of protection have been the lenders themselves.

Banks lending tens of billions to build data centers for Oracle and other tech giants have been buying protection on those companies.

Last November, Morgan Stanley's credit strategists said lenders on Oracle-linked data-center loans were likely a key driver of the surge in Oracle CDS trading. In May, Bloomberg reported that banks were buying protection on the biggest tech borrowers so they could keep lending without breaching limits on their exposure to any one company. Their demand, it said, had pushed the cost of that protection to unusually high levels for the borrowers' credit ratings. By late July, a record $25.1 billion of Oracle protection was outstanding, up from $6.9 billion a year earlier, IFR reported, citing DTCC data.

So when a spread jumps, ask who's buying. CoreWeave's move last week was, by most accounts, about distress over cash flow and refinancing. But on the evidence above, part of Oracle's long climb has been lenders queuing up for cover, which says something about how much they've lent, not only about how likely Oracle is to fail.

Why it matters

For borrowers, a widely watched price they don't set can, when it widens, make the next lender warier and new borrowing dearer, which can raise the very risk the spread was pricing, a report published by the European Systemic Risk Board warns.

For lenders, CDS make it possible to keep lending to a borrower while handing off some of the risk, which helps explain why AI lending and AI protection have grown together.

For governments, the machine is the same but the rules differ. Standard contracts on Western European governments, France's included, pay out on a missed payment, a refusal or suspension of payments, or a restructuring, but not on bankruptcy, and the odds shortcut above doesn't apply mechanically. France's five-year CDS rose about 5bp to roughly 82bp on Monday as riots deepened the country's political crisis, Reuters reported, ahead of a Moody's review of its rating on Friday; Germany's was around 12bp on Monday, according to Investing.com data.

The next time a CDS number makes the news

Meta reports quarterly results after Wednesday's close, so you may not have to wait long. Ask:

1. Which contract, and when? Five years? A closing price or an intraday high? From which provider?
2. Wide compared with what? Its own history, its peers, the investment-grade index?
3. Who's buying the protection? Lenders hedging, or investors betting on trouble?
4. What would count as default, and who decides?

Two sides, a fee, a committee and an auction: that's the whole machine. The borrower isn't in it, though its own lenders may be, buying protection.

It just has to live with the price.