The market now charges more to insure Big Tech than it does banks

Roughly 60 basis points. That is the gap, measured by Apollo's economists, between what the market now charges to insure against a hyperscaler — one of the handful of tech giants building out artificial intelligence — defaulting on its debt, and what it charges to insure a bank against the same. A basis point is a hundredth of a percentage point, the unit in which this insurance is quoted. The instrument is a credit default swap: a contract that pays out if a borrower stops paying. Banks are the institutions the public last had to rescue, in 2008, and the bank side is now the cheaper side of this insurance. ^[1] On August 11 the same insurance was as cheap as it has ever been. Andromeda Capital described the terms on offer to AI credit investors as a "heads I win, tails you lose" setup, with spreads — the premium over a safe benchmark — at record lows. ^[2] The repricing then moved in short, dated steps. Broadcom's default-insurance spreads roughly tripled after it raised $60 billion in debt on August 21 ^[3]. By September 27 the widening had reached Nvidia, Oracle and Meta at once ^[4]. The clearest ledger belongs to Oracle, which S&P cut to BBB- — the bottom rung of investment grade, the rating tier most funds still treat as safe and one notch above junk — back in July, before the market itself moved. The arithmetic the agency published as the repricing ran: revenue projected to grow 239% between fiscal 2022 and 2028, debt growing 412%, and 2027 capital-spending guidance lifted from $60 billion to $95 billion. Its label for the company was a "show me" story. ^[5] Already, by mid-September, the old hierarchy had begun to invert — TSMC's bonds trading cheaper than Meta's was one marker ^[6] — and the repricing was about to get a name. One explanation is the closest thing to a mechanism the run has produced.

What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets. — Torsten Slok

The naming came from institutions. On September 20 the IMF's managing director called AI's financing "front-loaded financial stability risks," pointing at circular financing — AI companies borrowing from the same lenders and investors who are also their customers ^[7]. Five days later, Kansas City Fed president Jeffrey Schmid asked the question out loud, borrowing the phrase once reserved for the banks rescued after 2008 — too big to fail, the label for an institution judged too central to let collapse without a public rescue. ^[8]

Where we have to start to really synthesize what's happening in the AI and the data center build-out is are we moving to a too-big-to-fail AI ecosystem. — Jeffrey Schmid

The same day, Japan's Financial Services Agency said it was stepping up scrutiny of the AI data-center loans its banks and life insurers had made, most of them to US projects ^[9]. The prices had moved first; the official language caught up. By mid-October the numbers had reached records. Oracle's default-insurance spread hit 288 basis points, SoftBank's 424, and CoreWeave — a loss-making data-center operator carrying nearly $50 billion in debt — passed 1,000 ^[10][11]. The cost of insuring the buildout rose while the plans for building it got bigger — the market charging more for the risk at the very moment the builders committed to more of it. Hyperscalers kept lifting their 2026 and 2027 capital-spending guidance through September ^[12], and the US government's own projection put the buildout at $10.3 trillion through 2032 ^[13]. The repricing landed in the same weeks the governments that would have to catch a failure were borrowing at their most expensive rates in a generation — coincidence of timing, not one story causing the other. The US 10-year Treasury yield rose above 5.3%, its highest since 2002 ^[14], while the government needs to sell roughly $2 trillion in bonds a year into a market where foreign buyers are stepping back, with China and Japan's central banks cutting Treasury holdings by $120 billion ^[15]. The IMF warned of a feedback loop in which heavy debt and rising yields reinforce one another ^[16]. The selloff's proximate causes were the oil shock from the Iran war and fiscal worry, not AI ^[17][18]. Andrew Bailey, the Bank of England governor, put the brittleness plainly, warning that leveraged positions can unwind into forced selling when fiscal credibility is lacking. ^[19] The counter-movement is real, and it reads best as a diagnosis of the present calm rather than a rival thesis. Corporate credit stayed stable through the bond rout for a dated, specific reason: the market's single biggest borrower has temporarily stopped borrowing. Goldman Sachs credit trader Jeffrey Papai says the calm has roughly two months left before the AI capital-expenditure debt cycle resumes into a repriced market. ^[20] The placement machine still works, and the sell-side is blunt about why.

We don’t doubt that we will be able to place it, it’s all a question of price. — Chase Bank

These businesses are still very solid and are raising money cheaply relative to their credit rating. — Hugh Briscoe

So the reclassification is on the books, but it has not yet been tested against fresh supply. The test has a date — one the market set itself. The biggest borrower paused. Goldman's own trader gives the pause about two months ^[20]. When the AI debt machine restarts, it will do so into a market that charges more to insure Big Tech than it charges to insure banks, and into a sovereign backstop priced at its tightest in more than twenty years. Whether the repricing means anything will be answered not by another warning but by the first bond sale.