The S&P 500 lost about $2.9tn of value yesterday - the largest single-day loss in dollar terms on record. Chip stocks fell 12%.

Three weeks ago, with CoreWeave’s bonds yielding 15%, its co-founder Brannin McBee was asked by the FT where the next round of financing for the AI build-out would come from. His answer:

“[Insurance companies]. It’s unlimited. Like, that’s all the money in the world that sits in the insurance tranche.”¹

The risk in the AI trade passes through four markets - equity, bonds, private credit, and insurance wrappers on bank loans - each one is slower to reprice than the one before. When the fast ones get expensive, the risk moves down to the slow ones.

What's at the end of the chain? Insurance (twice).

Equity marks every second, as we saw yesterday. The tech companies that spent a decade buying back their shares are now selling them to fund the build-out.

Oracle sold $10bn of stock this week at an 11% discount. Meta sold $37.5bn the week before. With the Nasdaq more than 10% below the high it set a month ago, selling equity is becoming more expensive.

Bonds are marked in real time but tend to reprice in spread over weeks - AI debt has been repricing wider for a month.

This week Oracle borrowed $10bn at up to 8.7%. Meta borrowed $15bn, paying 7.5% for 30-year money that cost it 5.6% a year ago. For years tech bonds traded tighter than the average investment-grade bond. This week they traded wider by the largest margin since Bloomberg's data began in 2003. S&P put Meta's AA- rating on negative outlook last Friday. Oracle is already BBB-, one notch above junk.

The insurance tranche McBee was referring to is the ‘float’: insurance premiums being invested until they are owed back. Issue 2 and Issue 3 covered how that money has been increasingly moving into private credit. The AI build-out is the newest thing its buying.

CoreWeave borrows in two ways:

Life insurers are the natural buyers because of how they’re regulated.

A life insurer has a leveraged balance sheet, like a bank. It buys assets with policyholder premiums and counts its liabilities as what it owes policyholders. The thin gap between the two is its capital, and the regulator says how big that gap has to be, asset by asset, according to the rating: about $1m for every $100m of A-rated debt, about $9.5m for every $100m of a B-rated bond like CoreWeave’s.³

Once the loan is on the books the insurer carries it at cost in its regulatory accounts until the rating changes. Spread widening doesn’t affect the book - only a downgrade does. (That’s by design - these balance sheets are built to hold to maturity).

How much AI exposure is in there?

A bank that insures a data-centre loan with an A-rated insurer can hold capital against the insurer's rating instead of the borrower's. The insurer takes the risk unfunded.

Barclays is looking to buy default protection from insurers on more than $10bn of AI-infrastructure debt, and on some facilities wants them to carry more than half the credit risk. HSBC is in early talks on project loans including a data-centre portfolio of more than $1bn. Deutsche Bank has been working on a similar transfer of about $2bn.⁶

The product is credit insurance. The bank pays a premium and the insurer covers a share of the losses if the borrower defaults. Banks have bought it for years, mostly on trade finance. Around 60 insurers write it, all rated A- or better, and about 98% of bank claims by value were paid in full between 2007 and 2022.

Nothing is posted, the book is marked only at a loss or when the policy expires, and a claim arrives only after a default and a waiting period. As a result, of all the ways to hold AI credit risk, this one prices it least often.

Insurers are struggling to digest the demand for the product.

Some credit insurers told WTW this year they were "approaching sector limits or are already full on data center exposure".⁶ Reinsurers have flagged data-centre concentrations for closer scrutiny at the 1 January renewals.

Citadel's co-CIO described the banks' problem as a credit market "getting jammed right now because it hasn't figured out a good way" to offload these risks. In other words, insurers are being asked to take this risk off banks’ balance sheets, and the underwriters are telling the banks they're full.

Yep. A credit insurer invests its premium, and its capital, in bonds, and bond portfolios are filling up with the same names. The five hyperscalers were 2.7% of the US investment-grade index in July 2025 and 4.8% a year later.

Unless a credit underwriter is deliberately screening them out, it already owns Meta and Oracle on one side of its balance sheet while it guarantees data-centre loans on the other. If the data centres stop paying, it loses on the loans it guaranteed and on the bonds it owns, in the same year.

Pretty much. People remember AIG for the $379bn of protection it sold to banks for capital relief, but that book lost 0.1%. What broke AIG was the $78bn of protection on a single factor (mortgage CDOs), while its life insurers held mortgage bonds bought with securities-lending cash and lost on the same factor at the same time.⁸

Guaranteeing bank loans didn't kill AIG. Guaranteeing one thing and owning the same thing did.

Apollo’s chief economist, Torsten Slok, wrote in August that “a portfolio can appear diversified across sectors and asset classes and still be exposed to the same underlying factor: AI.”⁷ Apollo owns Athene, one of the biggest annuity writers in America. He certainly knows the balance sheets he’s describing…

Yup. The credit rating sets the price. CoreWeave's A3 loan costs it about 6%. Its junk-rated bonds yield 15%.

The ratings are moving. Meta’s AA- is now on negative outlook. Oracle, the biggest hyperscaler name in life insurers’ portfolios, is one notch from junk.

We don’t have good visibility. More than $500bn of insurers’ bonds carry private letter ratings the market never sees and private rated bonds held by life insurers are "twice as likely to impair yet are downgraded less often" than public bonds with the same rating.⁹

When ratings go wrong, balance sheets that carry assets at cost find out late and all at once. We all know about mortgage bonds in 2008.

As Bloomberg’s Matt Levine put it last month, “there’s a kajillion dollars of systematic data-center credit risk, and someone has to own it.”¹⁰ We know who owns it - the question is how much we trust the ratings.

CoreWeave reports on Tuesday. I'll be watching whether the Barclays and HSBC deals get done and at what price, and how credit reinsurers treat data-centre concentrations at the Jan 1 renewals.

The standing invitation holds. If you've been sitting on a question - about a product, a structure, an idea - send it over.

Until then.

Charlie Evans, CAIA
Chief Investment Officer, Zephyros Investments

———

¹ Financial Times, 14 October 2026. The FT original is paywalled; quotes via Seeking Alpha and Dealroom. https://dealroom.co/news/162445-coreweave-wants-investment-grade-ratings-to-tap-unlimited-insurance-mone/

² CoreWeave, 30 March 2026 - $8.5bn facility rated Moody's A3 and DBRS A (low), SOFR plus 2.25%, anchored by Blackstone Credit & Insurance with insurance investors participating. https://www.businesswire.com/news/home/20260330529766/en/CoreWeave-Closes-Landmark-$8.5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing

³ NAIC life risk-based capital factors (pre-tax): 1.016% for A3, 9.535% for B. https://content.naic.org/sites/default/files/inline-files/2021-11-L%20RBC%20Proposal.pdf

⁴ Moody's, reported by Bloomberg, 22 October 2026 - $85.2bn, 1.4% of cash and invested assets; Oracle about $10bn. https://www.royalgazette.com/reinsurance/business/article/20261025/life-insurers-investment-exposure-to-ai-spikes-24/

⁵ NAIC Capital Markets Bureau, "Private Credit Analysis of the U.S. Insurance Industry", September 2026 - about $1.2tn of private credit and about $512bn of privately rated bonds at year-end 2025; "Quantifying U.S. insurers' exposure to private credit is challenging." https://content.naic.org/sites/default/files/capital-markets-special-reports-private-credit-analysis.pdf

⁶ Bloomberg, 29 and 30 October 2026, on Barclays and HSBC, including the WTW survey and Pablo Salame's comments. https://www.insurancejournal.com/news/international/2026/10/29/890719.htm · https://www.insurancejournal.com/news/international/2026/10/30/890898.htm

⁷ Torsten Slok, Apollo, "How Diversified Is Your Portfolio, Really?", 18 August 2026 - hyperscalers 2.7% of the US investment-grade index in July 2025, 4.8% a year later. https://www.apollo.com/wealth/insights-news/insights/daily-spark/how-diversified-is-your-portfolio-really

⁸ Robert McDonald and Anna Paulson, "AIG in Hindsight", NBER Working Paper 21108 - the $379bn regulatory-capital book and its $379m loss; $78bn of multi-sector CDO protection losing $11.5bn in 2007 and $28.6bn in 2008; about $21bn of securities-lending losses. https://www.nber.org/papers/w21108.pdf

⁹ Institutional Investor, 6 July 2026, on the Columbia Business School study by Li, Oh and Ricciardi; NAIC Investment Designation Analysis (E) Working Group, Summer 2026 National Meeting summary, on the SVO's discretion over filing-exempt ratings becoming operational on 13 August 2026. https://www.institutionalinvestor.com/article/study-says-life-insurers-may-need-45b-more-capital-private-credit-holdings · https://content.naic.org/sites/default/files/national_meeting/2026sumnm-summary-idawgv3.pdf

¹⁰ Matt Levine, "Someone Has to Take the AI Risk", Money Stuff, Bloomberg, 20 October 2026. https://www.bloomberg.com/opinion/newsletters/2026-10-20/someone-has-to-take-the-ai-risk